Transistors, the building blocks of modern electronics, are typically made of silicon. Because it’s a semiconductor, this material can control the flow of electricity in a circuit. But silicon has fundamental physical limits that restrict how compact and energy-efficient a transistor can be.

MIT researchers have now replaced silicon with a magnetic semiconductor, creating a magnetic transistor that could enable smaller, faster, and more energy-efficient circuits. The material’s magnetism strongly influences its electronic behavior, leading to more efficient control of the flow of electricity. 

The team used a novel magnetic material and an optimization process that reduces the material’s defects, which boosts the transistor’s performance.

The material’s unique magnetic properties also allow for transistors with built-in memory, which would simplify circuit design and unlock new applications for high-performance electronics.

“People have known about magnets for thousands of years, but there are very limited ways to incorporate magnetism into electronics. We have shown a new way to efficiently utilize magnetism that opens up a lot of possibilities for future applications and research,” says Chung-Tao Chou, an MIT graduate student in the departments of Electrical Engineering and Computer Science (EECS) and Physics, and co-lead author of a paper on this advance.

Chou is joined on the paper by co-lead author Eugene Park, a graduate student in the Department of Materials Science and Engineering (DMSE); Julian Klein, a DMSE research scientist; Josep Ingla-Aynes, a postdoc in the MIT Plasma Science and Fusion Center; Jagadeesh S. Moodera, a senior research scientist in the Department of Physics; and senior authors Frances Ross, TDK Professor in DMSE; and Luqiao Liu, an associate professor in EECS, and a member of the Research Laboratory of Electronics; as well as others at the University of Chemistry and Technology in Prague. The paper appears today in Physical Review Letters.

Overcoming the limits

In an electronic device, silicon semiconductor transistors act like tiny light switches that turn a circuit on and off, or amplify weak signals in a communication system. They do this using a small input voltage.

But a fundamental physical limit of silicon semiconductors prevents a transistor from operating below a certain voltage, which hinders its energy efficiency.

To make more efficient electronics, researchers have spent decades working toward magnetic transistors that utilize electron spin to control the flow of electricity. Electron spin is a fundamental property that enables electrons to behave like tiny magnets.

So far, scientists have mostly been limited to using certain magnetic materials. These lack the favorable electronic properties of semiconductors, constraining device performance.

“In this work, we combine magnetism and semiconductor physics to realize useful spintronic devices,” Liu says.

The researchers replace the silicon in the surface layer of a transistor with chromium sulfur bromide, a two-dimensional material that acts as a magnetic semiconductor.

Due to the material’s structure, researchers can switch between two magnetic states very cleanly. This makes it ideal for use in a transistor that smoothly switches between “on” and “off.”

“One of the biggest challenges we faced was finding the right material. We tried many other materials that didn’t work,” Chou says.

They discovered that changing these magnetic states modifies the material’s electronic properties, enabling low-energy operation. And unlike many other 2D materials, chromium sulfur bromide remains stable in air.

To make a transistor, the researchers pattern electrodes onto a silicon substrate, then carefully align and transfer the 2D material on top. They use tape to pick up a tiny piece of material, only a few tens of nanometers thick, and place it onto the substrate.

“A lot of researchers will use solvents or glue to do the transfer, but transistors require a very clean surface. We eliminate all those risks by simplifying this step,” Chou says.

Leveraging magnetism

This lack of contamination enables their device to outperform existing magnetic transistors. Most others can only create a weak magnetic effect, changing the flow of current by a few percent or less. Their new transistor can switch or amplify the electric current by a factor of 10.

They use an external magnetic field to change the magnetic state of the material, switching the transistor using significantly less energy than would usually be required.

The material also allows them to control the magnetic states with electric current. This is important because engineers cannot apply magnetic fields to individual transistors in an electronic device. They need to control each one electrically.

The material’s magnetic properties could also enable transistors with built-in memory, simplifying the design of logic or memory circuits.

A typical memory device has a magnetic cell to store information and a transistor to read it out. Their method can combine both into one magnetic transistor.

“Now, not only are transistors turning on and off, they are also remembering information. And because we can switch the transistor with greater magnitude, the signal is much stronger so we can read out the information faster, and in a much more reliable way,” Liu says.

Building on this demonstration, the researchers plan to further study the use of electrical current to control the device. They are also working to make their method scalable so they can fabricate arrays of transistors.

This research was supported, in part, by the Semiconductor Research Corporation, the U.S. Defense Advanced Research Projects Agency (DARPA), the U.S. National Science Foundation (NSF), the U.S. Department of Energy, the U.S. Army Research Office, and the Czech Ministry of Education, Youth, and Sports. The work was partially carried out at the MIT.nano facilities.

MIT.nano has added a new X-ray diffraction (XRD) instrument to its characterization toolset, enhancing facility users’ ability to analyze materials at the nanoscale. While many XRD systems exist across MIT’s campus, this new instrument, the Bruker D8 Discover Plus, is unique in that it features a high-brilliance micro-focus copper X-ray source — ideal for measuring small areas of thin film samples using a large area detector.

The new system is positioned within Characterization.nano’s X-ray diffraction and imaging shared experimental facility (SEF), where advanced instrumentation allows researchers to “see inside” materials at very small scales. Here, scientists and engineers can examine surfaces, layers, and internal structures without damaging the material, and create detailed 3D images to map composition and organization. The information gathered is supporting materials research for applications ranging from electronics and energy storage to health care and nanotechnology.

“The Bruker instrument is an important addition to MIT.nano that will help researchers efficiently gain insights into their materials’ structure and properties,” says Charlie Settens, research specialist and operations manager in the Characterization.nano X-ray diffraction and imaging SEF. “It brings high-performance diffraction capabilities to our lab, supporting everything from routine phase identification to complex thin film microstructural analysis and high-temperature studies.”

What is X-ray diffraction?

When people think of X-rays, they often picture medical imaging, where dense structures like bones appear in contrast to soft tissue. X-ray diffraction takes that concept further, revealing the crystalline structure of materials by measuring the interference patterns that form when X-rays interact with atomic planes. These diffraction patterns provide detailed information about a material’s crystalline phase, grain size, grain orientation, defects, and other structural properties.

XRD is essential across many fields. Civil engineers use it to analyze the components of concrete mixtures and monitor material changes over time. Materials scientists engineer new microstructures and track how atomic arrangements shift with different element combinations. Electrical engineers study crystalline thin film deposition on substrates — critical for semiconductor manufacturing. MIT.nano’s new X-ray diffractometer will support all of these applications, and more.

“The addition of another high-resolution XRD will make it a lot easier to get time on these very popular tools,” says Fred Tutt, PhD student in the MIT Department of Materials Science and Engineering. “The wide variety of options on the new Bruker will also make it easier for myself and my group members to take some of the more atypical measurements that aren't readily accessible with the current XRD tools.”

A closer, clearer look

Replacing two older systems, the Bruker D8 Discover Plus introduces the latest in X-ray diffraction technology to MIT.nano, along with several major upgrades for the Characterization.nano facility. One key feature is the high-brilliance microfocus copper X-ray source, capable of producing intense X-rays from a small spot size — ranging from 2mm down to 200 microns.

“It’s invaluable to have the flexibility to measure distinct regions of a sample with high flux and fine spatial resolution,” says Jordan Cox, MIT.nano research specialist in the MIT.nano X-ray diffraction and imaging facility.

Another highlight is in-plane XRD, a technique that enables surface diffraction studies of thin films with non-uniform grain orientations.

“In-plane XRD pairs well with many thin film projects that start in the fab,” says Settens. After researchers deposit thin film coatings in MIT.nano’s cleanroom, they can selectively measure the top 100 nanometers of the surface, he explains.

But it’s not just about collecting diffraction patterns. The new system includes a powerful software suite for advanced data analysis. Cox and Settens are now training users how to operate the diffractometer, as well as how to analyze and interpret the valuable structural data it provides.

Visit Characterization.nano for more information about this and other tools.

With respiratory illness season in full swing, a bad night’s sleep, sore throat, and desire to cancel dinner plans could all be considered hallmark symptoms of the flu, Covid-19 or other illnesses. Although everyone has, at some point, experienced illness and these stereotypical symptoms, the mechanisms that generate them are not well understood.

Zuri Sullivan, a new assistant professor in the MIT Department of Biology and core member of the Whitehead Institute for Biomedical Research, works at the interface of neuroscience, microbiology, physiology, and immunology to study the biological workings underlying illness. In this interview, she describes her work on immunity thus far as well as research avenues — and professional collaborations — she’s excited to explore at MIT.

Q: What is immunity, and why do we get sick in the first place? 

A: We can think of immunity in two ways: the antimicrobial programs that defend against a pathogen directly, and sickness, the altered organismal state that happens when we get an infection. 

Sickness itself arises from brain-immune system interaction. The immune system is talking to the brain, and then the brain has a system-wide impact on host defense via its ability to have top-down control of physiologic systems and behavior. People might assume that sickness is an unintended consequence of infection, that it happens because your immune system is active, but we hypothesize that it’s likely an adaptive process that contributes to host defense. 

If we consider sickness as immunity at the organismal scale, I think of my work as bridging the dynamic immunological processes that occur at the cellular scale, the tissue scale, and the organismal scale. I’m interested in the molecular and cellular mechanisms by which the immune system communicates with the brain to generate changes in behavior and physiology, such as fever, loss of appetite, and changes in social interaction. 

Q: What sickness behaviors fascinate you? 

A: During my thesis work at Yale University, I studied how the gut processes different nutrients and the role of the immune system in regulating gut homeostasis in response to different kinds of food. I’m especially interested in the interaction between food, the immune system, and the brain. One of the things I’m most excited about is the reduction in appetite, or changes in food choice, because we have what I would consider pretty strong evidence that these may be adaptive. 

Sleep is another area we’re interested in exploring. From their own subjective experience, everyone knows that sleep is often altered during infection. 

I also don’t just want to examine snapshots in time. I want to characterize changes over the course of an infection. There’s probably going to be individual variability, which I think may be in part because pathogens are also changing over the course of an illness — we’re studying two different biological systems interacting with each other. 

Q: What sorts of expertise are you hoping to recruit to your lab, and what collaborations are you excited about pursuing?

A: I really want to bring together different areas of biology to think about organism-wide questions. The thing that’s most important to me is people who are creative — I’d rather trainees come in with an interesting idea than a perfectly formed question within the bounds of what we already believe to be true. I’m also interested in people who would complement my expertise; I’m fascinated by microbiology, but I don’t have any formal training.

The Whitehead Institute is really invested in interdisciplinary work, and there’s a natural synergy between my work and the other labs in this small community at the Whitehead Institute.

I’ve been collaborating with Sebastian Lourido’s lab for a few years, looking at how Toxoplasma gondii influences social behavior, and I’m excited to invest more time in that project. I’m also interested in molecular neuroscience, which is a focus of Siniša Hrvatin’s lab. That lab is interested in the hypothalamus, and trying to understand the mechanisms that generate torpor. My work also focuses on the hypothalamus because it regulates homeostatic behaviors that change during sickness, such as appetite, sleep, social behavior, and body temperature. 

By studying different sickness states generated by different kinds of pathogens — parasites, viruses, bacteria — we can ask really interesting questions about how and why we get sick. 

Numerous potential treatments for neurological conditions, including autism spectrum disorders, have worked well in mice but then disappointed in humans. What would help is a non-invasive, objective readout of treatment efficacy that is shared in both species. 

In a new open-access study in Nature Communications, a team of MIT researchers, backed by collaborators across the United States and in the United Kingdom, identifies such a biomarker in fragile X syndrome, the most common inherited form of autism.

Led by postdoc Sara Kornfeld-Sylla and Picower Professor Mark Bear, the team measured the brain waves of human boys and men, with or without fragile X syndrome, and comparably aged male mice, with or without the genetic alteration that models the disorder. The novel approach Kornfeld-Sylla used for analysis enabled her to uncover specific and robust patterns of differences in low-frequency brain waves between typical and fragile X brains shared between species at each age range. In further experiments, the researchers related the brain waves to specific inhibitory neural activity in the mice and showed that the biomarker was able to indicate the effects of even single doses of a candidate treatment for fragile X called arbaclofen, which enhances inhibition in the brain.

Both Kornfeld-Sylla and Bear praised and thanked colleagues at Boston Children’s Hospital, the Phelan-McDermid Syndrome Foundation, Cincinnati Children’s Hospital, the University of Oklahoma, and King’s College London for gathering and sharing data for the study.

“This research weaves together these different datasets and finds the connection between the brain wave activity that’s happening in fragile X humans that is different from typically developed humans, and in the fragile X mouse model that is different than the ‘wild-type’ mice,” says Kornfeld-Sylla, who earned her PhD in Bear’s lab in 2024 and continued the research as a FRAXA postdoc. “The cross-species connection and the collaboration really makes this paper exciting.”

Bear, a faculty member in The Picower Institute for Learning and Memory and the Department of Brain and Cognitive Sciences at MIT, says having a way to directly compare brain waves can advance treatment studies.

“Because that is something we can measure in mice and humans minimally invasively, you can pose the question: If drug treatment X affects this signature in the mouse, at what dose does that same drug treatment change that same signature in a human?” Bear says. “Then you have a mapping of physiological effects onto measures of behavior. And the mapping can go both ways.”

Peaks and powers

In the study, the researchers measured EEG over the occipital lobe of humans and on the surface of the visual cortex of the mice. They measured power across the frequency spectrum, replicating previous reports of altered low-frequency brain waves in adult humans with fragile X and showing for the first time how these disruptions differ in children with fragile X.

To enable comparisons with mice, Kornfeld-Sylla subtracted out background activity to specifically isolate only “periodic” fluctuations in power (i.e., the brain waves) at each frequency. She also disregarded the typical way brain waves are grouped by frequency (into distinct bands with Greek letter designations delta, theta, alpha, beta, and gamma) so that she could simply juxtapose the periodic power spectra of the humans and mice without trying to match them band by band (e.g., trying to compare the mouse “alpha” band to the human one). This turned out to be crucial because the significant, similar patterns exhibited by the mice actually occurred in a different low-frequency band than in the humans (theta vs. alpha). Both species also had alterations in higher-frequency bands in fragile X, but Kornfeld-Sylla noted that the differences in the low-frequency brainwaves are easier to measure and more reliable in humans, making them a more promising biomarker.

So what patterns constitute the biomarker? In adult men and mice alike, a peak in the power of low-frequency waves is shifted to a significantly slower frequency in fragile X cases compared to in neurotypical cases. Meanwhile, in fragile X boys and juvenile mice, while the peak is somewhat shifted to a slower frequency, what is really significant is a reduced power in that same peak.

The researchers were also able to discern that the peak in question is actually made of two distinct subpeaks, and that the lower-frequency subpeak is the one that varies specifically with fragile X syndrome.

Curious about the neural activity underlying the measurements, the researchers engaged in experiments in which they turned off activity of two different kinds of inhibitory neurons that are known to help produce and shape brain wave patterns: somatostatin-expressing and parvalbumin-expressing interneurons. Manipulating the somatostatin neurons specifically affected the lower-frequency subpeak that contained the newly discovered biomarker in fragile X model mice.

Drug testing

Somatostatin interneurons exert their effects on the neurons they connect to via the neurotransmitter chemical GABA, and evidence from prior studies suggest that GABA receptivity is reduced in fragile X syndrome. A therapeutic approach pioneered by Bear and others has been to give the drug arbaclofen, which enhances GABA activity. In the new study, the researchers treated both control and fragile X model mice with arbaclofen to see how it affected the low-frequency biomarker.

Even the lowest administered single dose made a significant difference in the neurotypical mice, which is consistent with those mice having normal GABA responsiveness. Fragile X mice needed a higher dose, but after one was administered, there was a notable increase in the power of the key subpeak, reducing the deficit exhibited by juvenile mice.

The arbaclofen experiments therefore demonstrated that the biomarker provides a significant readout of an underlying pathophysiology of fragile X: the reduced GABA responsiveness. Bear also noted that it helped to identify a dose at which arbaclofen exerted a corrective effect, even though the drug was only administered acutely, rather than chronically. An arbaclofen therapy would, of course, be given over a long time frame, not just once.

“This is a proof of concept that a drug treatment could move this phenotype acutely in a direction that makes it closer to wild-type,” Bear says. “This effort reveals that we have readouts that can be sensitive to drug treatments.”

Meanwhile, Kornfeld-Sylla notes, there is a broad spectrum of brain disorders in which human patients exhibit significant differences in low-frequency (alpha) brain waves compared to neurotypical peers.

“Disruptions akin to the biomarker we found in this fragile X study might prove to be evident in mouse models of those other disorders, too,” she says. “Identifying this biomarker could broadly impact future translational neuroscience research.”

The paper’s other authors are Cigdem Gelegen, Jordan Norris, Francesca Chaloner, Maia Lee, Michael Khela, Maxwell Heinrich, Peter Finnie, Lauren Ethridge, Craig Erickson, Lauren Schmitt, Sam Cooke, and Carol Wilkinson.

The National Institutes of Health, the National Science Foundation, the FRAXA Foundation, the Pierce Family Fragile X Foundation, the Autism Science Foundation, the Thrasher Research Fund, Harvard University, the Simons Foundation, Wellcome, the Biotechnology and Biological Sciences Research Council, and the Freedom Together Foundation provided support for the research.

Just like each person has unique fingerprints, every CMOS chip has a distinctive “fingerprint” caused by tiny, random manufacturing variations. Engineers can leverage this unforgeable ID for authentication, to safeguard a device from attackers trying to steal private data.

But these cryptographic schemes typically require secret information about a chip’s fingerprint to be stored on a third-party server. This creates security vulnerabilities and requires additional memory and computation.

To overcome this limitation, MIT engineers developed a manufacturing method that enables secure, fingerprint-based authentication, without the need to store secret information outside the chip.

They split a specially designed chip during fabrication in such a way that each half has an identical, shared fingerprint that is unique to these two chips. Each chip can be used to directly authenticate the other. This low-cost fingerprint fabrication method is compatible with standard CMOS foundry processes and requires no special materials.

The technique could be useful in power-constrained electronic systems with non-interchangeable device pairs, like an ingestible sensor pill and its paired wearable patch that monitor gastrointestinal health conditions. Using a shared fingerprint, the pill and patch can authenticate each other without a device in between to mediate.

“The biggest advantage of this security method is that we don’t need to store any information. All the secrets will always remain safe inside the silicon. This can give a higher level of security. As long as you have this digital key, you can always unlock the door,” says Eunseok Lee, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on this security method.

Lee is joined on the paper by EECS graduate students Jaehong Jung and Maitreyi Ashok; as well as co-senior authors Anantha Chandrakasan, MIT provost and the Vannevar Bush Professor of Electrical Engineering and Computer Science, and Ruonan Han, a professor of EECS and a member of the MIT Research Laboratory of Electronics. The research was recently presented at the IEEE International Solid-States Circuits Conference.

“Creation of shared encryption keys in trusted semiconductor foundries could help break the tradeoffs between being more secure and more convenient to use for protection of data transmission,” Han says. “This work, which is digital-based, is still a preliminary trial in this direction; we are exploring how more complex, analog-based secrecy can be duplicated — and only duplicated once.”

Leveraging variations

Even though they are intended to be identical, each CMOS chip is slightly different due to unavoidable microscopic variations during fabrication. These randomizations give each chip a unique identifier, known as a physical unclonable function (PUF), that is nearly impossible to replicate.

A chip’s PUF can be used to provide security just like the human fingerprint identification system on a laptop or door panel.

For authentication, a server sends a request to the device, which responds with a secret key based on its unique physical structure. If the key matches an expected value, the server authenticates the device.

But the PUF authentication data must be registered and stored in a server for access later, creating a potential security vulnerability.

“If we don’t need to store information on these unique randomizations, then the PUF becomes even more secure,” Lee says.

The researchers wanted to accomplish this by developing a matched PUF pair on two chips. One could authenticate the other directly, without the need to store PUF data on third-party servers.

As an analogy, consider a sheet of paper torn in half. The torn edges are random and unique, but the pieces have a shared randomness because they fit back together perfectly along the torn edge.

While CMOS chips aren’t torn in half like paper, many are fabricated at once on a silicon wafer which is diced to separate the individual chips.

By incorporating shared randomness at the edge of two chips before they are diced to separate them, the researchers could create a twin PUF that is unique to these two chips.

“We needed to find a way to do this before the chip leaves the foundry, for added security. Once the fabricated chip enters the supply chain, we won’t know what might happen to it,” Lee explains.

Sharing randomness

To create the twin PUF, the researchers change the properties of a set of transistors fabricated along the edge of two chips, using a process called gate oxide breakdown.

Essentially, they pump high voltage into a pair of transistors by shining light with a low-cost LED until the first transistor breaks down. Because of tiny manufacturing variations, each transistor has a slightly different breakdown time. The researchers can use this unique breakdown state as the basis for a PUF.

To enable a twin PUF, the MIT researchers fabricate two pairs of transistors along the edge of two chips before they are diced to separate them. By connecting the transistors with metal layers, they create paired structures that have correlated breakdown states. In this way, they enable a unique PUF to be shared by each pair of transistors.

After shining LED light to create the PUF, they dice the chips between the transistors so there is one pair on each device, giving each separate chip a shared PUF.

“In our case, transistor breakdown has not been modeled well in many of the simulations we had, so there was a lot of uncertainty about how the process would work. Figuring out all the steps, and the order they needed to happen, to generate this shared randomness is the novelty of this work,” Lee says.

After finetuning their PUF generation process, the researchers developed a prototype pair of twin PUF chips in which the randomization was matched with more than 98 percent reliability. This would ensure the generated PUF key matches consistently, enabling secure authentication.

Because they generated this twin PUF using circuit techniques and low-cost LEDs, the process would be easier to implement at scale than other methods that are more complicated or not compatible with standard CMOS fabrication.

“In the current design, shared randomness generated by transistor breakdown is immediately converted into digital data. Future versions could preserve this shared randomness directly within the transistors, strengthening security at the most fundamental physical level of the chip,” Lee says.

“There is a rapidly increasing demand for physical-layer security for edge devices, such as between medical sensors and devices on a body, which often operate under strict energy constraints. A twin-paired PUF approach enables secure communication between nodes without the burden of heavy protocol overhead, thereby delivering both energy efficiency and strong security. This initial demonstration paves the way for innovative advancements in secure hardware design,” Chandrakasan adds.

This work is funded by Lockheed Martin, the MIT School of Engineering MathWorks Fellowship, and the Korea Foundation for Advanced Studies Fellowship.

Large language models (LLMs) have been championed as tools that could democratize access to information worldwide, offering knowledge in a user-friendly interface regardless of a person’s background or location. However, new research from MIT’s Center for Constructive Communication (CCC) suggests these artificial intelligence systems may actually perform worse for the very users who could most benefit from them.

A study conducted by researchers at CCC, which is based at the MIT Media Lab, found that state-of-the-art AI chatbots — including OpenAI’s GPT-4, Anthropic’s Claude 3 Opus, and Meta’s Llama 3 — sometimes provide less-accurate and less-truthful responses to users who have lower English proficiency, less formal education, or who originate from outside the United States. The models also refuse to answer questions at higher rates for these users, and in some cases, respond with condescending or patronizing language.

“We were motivated by the prospect of LLMs helping to address inequitable information accessibility worldwide,” says lead author Elinor Poole-Dayan SM ’25, a technical associate in the MIT Sloan School of Management who led the research as a CCC affiliate and master’s student in media arts and sciences. “But that vision cannot become a reality without ensuring that model biases and harmful tendencies are safely mitigated for all users, regardless of language, nationality, or other demographics.”

A paper describing the work, “LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users,” was presented at the AAAI Conference on Artificial Intelligence in January.

Systematic underperformance across multiple dimensions

For this research, the team tested how the three LLMs responded to questions from two datasets: TruthfulQA and SciQ. TruthfulQA is designed to measure a model’s truthfulness (by relying on common misconceptions and literal truths about the real world), while SciQ contains science exam questions testing factual accuracy. The researchers prepended short user biographies to each question, varying three traits: education level, English proficiency, and country of origin.

Across all three models and both datasets, the researchers found significant drops in accuracy when questions came from users described as having less formal education or being non-native English speakers. The effects were most pronounced for users at the intersection of these categories: those with less formal education who were also non-native English speakers saw the largest declines in response quality.

The research also examined how country of origin affected model performance. Testing users from the United States, Iran, and China with equivalent educational backgrounds, the researchers found that Claude 3 Opus in particular performed significantly worse for users from Iran on both datasets.

“We see the largest drop in accuracy for the user who is both a non-native English speaker and less educated,” says Jad Kabbara, a research scientist at CCC and a co-author on the paper. “These results show that the negative effects of model behavior with respect to these user traits compound in concerning ways, thus suggesting that such models deployed at scale risk spreading harmful behavior or misinformation downstream to those who are least able to identify it.”

Refusals and condescending language

Perhaps most striking were the differences in how often the models refused to answer questions altogether. For example, Claude 3 Opus refused to answer nearly 11 percent of questions for less educated, non-native English-speaking users — compared to just 3.6 percent for the control condition with no user biography.

When the researchers manually analyzed these refusals, they found that Claude responded with condescending, patronizing, or mocking language 43.7 percent of the time for less-educated users, compared to less than 1 percent for highly educated users. In some cases, the model mimicked broken English or adopted an exaggerated dialect.

The model also refused to provide information on certain topics specifically for less-educated users from Iran or Russia, including questions about nuclear power, anatomy, and historical events — even though it answered the same questions correctly for other users.

“This is another indicator suggesting that the alignment process might incentivize models to withhold information from certain users to avoid potentially misinforming them, although the model clearly knows the correct answer and provides it to other users,” says Kabbara.

Echoes of human bias

The findings mirror documented patterns of human sociocognitive bias. Research in the social sciences has shown that native English speakers often perceive non-native speakers as less educated, intelligent, and competent, regardless of their actual expertise. Similar biased perceptions have been documented among teachers evaluating non-native English-speaking students.

“The value of large language models is evident in their extraordinary uptake by individuals and the massive investment flowing into the technology,” says Deb Roy, professor of media arts and sciences, CCC director, and a co-author on the paper. “This study is a reminder of how important it is to continually assess systematic biases that can quietly slip into these systems, creating unfair harms for certain groups without any of us being fully aware.”

The implications are particularly concerning given that personalization features — like ChatGPT’s Memory, which tracks user information across conversations — are becoming increasingly common. Such features risk differentially treating already-marginalized groups.

“LLMs have been marketed as tools that will foster more equitable access to information and revolutionize personalized learning,” says Poole-Dayan. “But our findings suggest they may actually exacerbate existing inequities by systematically providing misinformation or refusing to answer queries to certain users. The people who may rely on these tools the most could receive subpar, false, or even harmful information.”

Eight MIT faculty and 22 additional MIT alumni are among 126 early-career researchers honored with 2026 Sloan Research Fellowships by the Alfred P. Sloan Foundation.

The fellowships honor exceptional researchers at U.S. and Canadian educational institutions, whose creativity, innovation, and research accomplishments make them stand out as the next generation of leaders. Winners receive a two-year, $75,000 fellowship that can be used flexibly to advance the fellow’s research.

"The Sloan Research Fellows are among the most promising early-career researchers in the U.S. and Canada, already driving meaningful progress in their respective disciplines," says Stacie Bloom, president and chief executive officer of the Alfred P. Sloan Foundation. "We look forward to seeing how these exceptional scholars continue to unlock new scientific advancements, redefine their fields, and foster the well-being and knowledge of all."

Including this year’s recipients, a total of 341 MIT faculty have received Sloan Research Fellowships since the program’s inception in 1955. The MIT recipients are:

Jacopo Borga is interested in probability theory and its connections to combinatorics, and in mathematical physics. He studies various random combinatorial structures — mathematical objects such as graphs or permutations — and their patterns and behavior at a large scale. This research includes random permutons, meanders, multidimensional constrained Brownian motions, Schramm-Loewner evolutions, and Liouville quantum gravity. Borga earned bachelor’s and master’s degrees in mathematics from the Università degli Studi di Padova in Italy, and a master’s degree in mathematics from Université Sorbonne Paris Cité in France, then proceeded to complete a PhD in mathematics at Unstitut für Mathematik at the Universität Zürich in Switzerland. Borga was an assistant professor at Stanford University before joining MIT as an assistant professor of mathematics in 2024.

Anna-Christina Eilers is an astrophysicist and assistant professor at MIT’s Department of Physics. Her research explores how black holes form and evolve across cosmic time, studying their origins and the role they play in shaping our universe. She leverages multi-wavelength data from telescopes all around the world and in space to study how the first galaxies, black holes, and quasars emerged during an epoch known as the Cosmic Dawn of our universe. She grew up in Germany and completed her PhD at the Max Planck Institute for Astronomy in Heidelberg. Subsequently, she was awarded a NASA Hubble Fellowship and a Pappalardo Fellowship to continue her research at MIT, where she joined the faculty in 2023. Her work has been recognized with several honors, including the PhD Prize of the International Astronomical Union, the Otto Hahn Medal of the Max Planck Society, and the Ludwig Biermann Prize of the German Astronomical Society.

Linlin Fan is the Samuel A. Goldblith Career Development Assistant Professor of Applied Biology in the Department of Brain and Cognitive Sciences and the Picower Institute for Learning and Memory at MIT. Her lab focuses on the development and application of advanced all-optical physiological techniques to understand the plasticity mechanisms underlying learning and memory. She has developed and applied high-speed, cellular-precision all-optical physiological techniques for simultaneously mapping and controlling membrane potential in specific neurons in behaving mammals. Prior to joining MIT, Fan was a Helen Hay Whitney Postdoctoral Fellow in Karl Deisseroth’s laboratory at Stanford University. She obtained her PhD in chemical biology from Harvard University in 2019 with Adam Cohen. Her work has been recognized by several awards, including the Larry Katz Memorial Lecture Award from the Cold Spring Harbor Laboratory, Helen Hay Whitney Fellowship, Career Award at the Scientific Interface from the Burroughs Wellcome Fund, Klingenstein-Simons Fellowship Award, Searle Scholar Award, and NARSAD Young Investigator Award.

Yoon Kim is an associate professor in the Department of EECS and a principal investigator in the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and the MIT-IBM Watson AI Lab, where he works on natural language processing and machine learning. Kim earned a PhD in computer science at Harvard University, an MS in data science from New York University, an MA in statistics from Columbia University, and BA in both math and economics from Cornell University. He joined EECS in 2021, after spending a year as a postdoc at MIT-IBM Watson AI Lab.

Haihao Lu PhD ’19 is the Cecil and Ida Green Career Development Assistant Professor, and an assistant professor of operations research/statistics at the MIT Sloan School of Management. Lu’s research lies at the intersection of optimization, computation, and data science, with a focus on pushing the computational and mathematical frontiers of large-scale optimization. Much of his work is inspired by real-world challenges faced by leading technology companies and optimization software companies, such as first-order methods and scalable solvers and data-driven optimization for resource allocation. His research has had real-world impact, generating substantial revenue and advancing the state of practice in large-scale optimization, and has been recognized by several research awards. Before joining MIT Sloan, he was an assistant professor at the University of Chicago Booth School of Business and a faculty researcher at Google Research’s large-scale optimization team. He obtained his PhD in mathematics and operations research at MIT in 2019.

Brett McGuire is the Class of 1943 Career Development Associate Professor of Chemistry at MIT. He completed his undergraduate studies at the University of Illinois at Urbana-Champaign before earning an MS from Emory University and a PhD from the Caltech, both in physical chemistry. After Jansky and Hubble postdoctoral fellowships at the National Radio Astronomy Observatory, he joined the MIT faculty in 2020 and was promoted to associate professor in 2025. The McGuire Group integrates physical chemistry, molecular spectroscopy, and observational astrophysics to explore how the chemical building blocks of life evolve alongside the formation of stars and planets.

Anand Natarajan PhD ’18 is an associate professor in EECS and a principal investigator in CSAIL and the MIT-IBM Watson AI Lab. His research is mainly in quantum complexity theory, with a focus on the power of interactive proofs and arguments in a quantum world. Essentially, his work attempts to assess the complexity of computational problems in a quantum setting, determining both the limits of quantum computers’ capability and the trustworthiness of their output. Natarajan earned his PhD in physics from MIT, and an MS in computer science and BS in physics from Stanford University. Prior to joining MIT in 2020, he spent time as a postdoc at the Institute for Quantum Information and Matter at Caltech.

Mengjia Yan is an associate professor in the Department of EECS and a principal investigator in CSAIL. She is a security computer architect whose research advances secure processor design by bridging computer architecture, systems security, and formal methods. Her work identifies critical blind spots in hardware threat models and improves the resilience of real-world systems against information leakage and exploitation. Several of her discoveries have influenced commercial processor designs and contributed to changes in how hardware security risks are evaluated in practice. In parallel, Yan develops architecture-driven techniques to improve the scalability of formal verification and introduces new design principles toward formally verifiable processors. She also designed the Secure Hardware Design (SHD) course, now widely adopted by universities worldwide to teach computer architecture security from both offensive and defensive perspectives.

The following MIT alumni also received fellowships:

Ashok Ajoy PhD ’16
Chibueze Amanchukwu PhD ’17
Annie M. Bauer PhD ’17
Kimberly K. Boddy ’07
danah boyd SM ’02
Yuan Cao SM ’16, PhD ’20
Aloni Cohen SM ’15, PhD ’19
Fei Dai PhD ’19
Madison M. Douglas ’16
Philip Engel ’10
Benjamin Eysenbach ’17
Tatsunori B. Hashimoto SM ’14, PhD ’16
Xin Jin ’10
Isaac Kim ’07
Christina Patterson PhD ’19
Katelin Schutz ’14
Karthik Shekhar PhD ’15
Shriya S. Srinivasan PhD ’20
Jerzy O. Szablowski ’09
Anna Wuttig PhD ’18
Zoe Yan PhD ’20
Lingfu Zhang ’18

By now, ChatGPT, Claude, and other large language models have accumulated so much human knowledge that they’re far from simple answer-generators; they can also express abstract concepts, such as certain tones, personalities, biases, and moods. However, it’s not obvious exactly how these models represent abstract concepts to begin with from the knowledge they contain.

Now a team from MIT and the University of California San Diego has developed a way to test whether a large language model (LLM) contains hidden biases, personalities, moods, or other abstract concepts. Their method can zero in on connections within a model that encode for a concept of interest. What’s more, the method can then manipulate, or “steer” these connections, to strengthen or weaken the concept in any answer a model is prompted to give.

The team proved their method could quickly root out and steer more than 500 general concepts in some of the largest LLMs used today. For instance, the researchers could home in on a model’s representations for personalities such as “social influencer” and “conspiracy theorist,” and stances such as “fear of marriage” and “fan of Boston.” They could then tune these representations to enhance or minimize the concepts in any answers that a model generates.

In the case of the “conspiracy theorist” concept, the team successfully identified a representation of this concept within one of the largest vision language models available today. When they enhanced the representation, and then prompted the model to explain the origins of the famous “Blue Marble” image of Earth taken from Apollo 17, the model generated an answer with the tone and perspective of a conspiracy theorist.

The team acknowledges there are risks to extracting certain concepts, which they also illustrate (and caution against). Overall, however, they see the new approach as a way to illuminate hidden concepts and potential vulnerabilities in LLMs, that could then be turned up or down to improve a model’s safety or enhance its performance.

“What this really says about LLMs is that they have these concepts in them, but they’re not all actively exposed,” says Adityanarayanan “Adit” Radhakrishnan, assistant professor of mathematics at MIT. “With our method, there’s ways to extract these different concepts and activate them in ways that prompting cannot give you answers to.”

The team published their findings today in a study appearing in the journal Science. The study’s co-authors include Radhakrishnan, Daniel Beaglehole and Mikhail Belkin of UC San Diego, and Enric Boix-Adserà of the University of Pennsylvania.

A fish in a black box

As use of OpenAI’s ChatGPT, Google’s Gemini, Anthropic’s Claude, and other artificial intelligence assistants has exploded, scientists are racing to understand how models represent certain abstract concepts such as “hallucination” and “deception.” In the context of an LLM, a hallucination is a response that is false or contains misleading information, which the model has “hallucinated,” or constructed erroneously as fact.

To find out whether a concept such as “hallucination” is encoded in an LLM, scientists have often taken an approach of “unsupervised learning” — a type of machine learning in which algorithms broadly trawl through unlabeled representations to find patterns that might relate to a concept such as “hallucination.” But to Radhakrishnan, such an approach can be too broad and computationally expensive.

“It’s like going fishing with a big net, trying to catch one species of fish. You’re gonna get a lot of fish that you have to look through to find the right one,” he says. “Instead, we’re going in with bait for the right species of fish.”

He and his colleagues had previously developed the beginnings of a more targeted approach with a type of predictive modeling algorithm known as a recursive feature machine (RFM). An RFM is designed to directly identify features or patterns within data by leveraging a mathematical mechanism that neural networks — a broad category of AI models that includes LLMs — implicitly use to learn features.

Since the algorithm was an effective, efficient approach for capturing features in general, the team wondered whether they could use it to root out representations of concepts, in LLMs, which are by far the most widely used type of neural network and perhaps the least well-understood.

“We wanted to apply our feature learning algorithms to LLMs to, in a targeted way, discover representations of concepts in these large and complex models,” Radhakrishnan says.

Converging on a concept

The team’s new approach identifies any concept of interest within a LLM and “steers” or guides a model’s response based on this concept. The researchers looked for 512 concepts within five classes: fears (such as of marriage, insects, and even buttons); experts (social influencer, medievalist); moods (boastful, detachedly amused); a preference for locations (Boston, Kuala Lumpur); and personas (Ada Lovelace, Neil deGrasse Tyson).

The researchers then searched for representations of each concept in several of today’s large language and vision models. They did so by training RFMs to recognize numerical patterns in an LLM that could represent a particular concept of interest.

A standard large language model is, broadly, a neural network that takes a natural language prompt, such as “Why is the sky blue?” and divides the prompt into individual words, each of which is encoded mathematically as a list, or vector, of numbers. The model takes these vectors through a series of computational layers, creating matrices of many numbers that, throughout each layer, are used to identify other words that are most likely to be used to respond to the original prompt. Eventually, the layers converge on a set of numbers that is decoded back into text, in the form of a natural language response.

The team’s approach trains RFMs to recognize numerical patterns in an LLM that could be associated with a specific concept. As an example, to see whether an LLM contains any representation of a “conspiracy theorist,” the researchers would first train the algorithm to identify patterns among LLM representations of 100 prompts that are clearly related to conspiracies, and 100 other prompts that are not. In this way, the algorithm would learn patterns associated with the conspiracy theorist concept. Then, the researchers can mathematically modulate the activity of the conspiracy theorist concept by perturbing LLM representations with these identified patterns. 

The method can be applied to search for and manipulate any general concept in an LLM. Among many examples, the researchers identified representations and manipulated an LLM to give answers in the tone and perspective of a “conspiracy theorist.” They also identified and enhanced the concept of “anti-refusal,” and showed that whereas normally, a model would be programmed to refuse certain prompts, it instead answered, for instance giving instructions on how to rob a bank.

Radhakrishnan says the approach can be used to quickly search for and minimize vulnerabilities in LLMs. It can also be used to enhance certain traits, personalities, moods, or preferences, such as emphasizing the concept of “brevity” or “reasoning” in any response an LLM generates. The team has made the method’s underlying code publicly available.

“LLMs clearly have a lot of these abstract concepts stored within them, in some representation,” Radhakrishnan says. “There are ways where, if we understand these representations well enough, we can build highly specialized LLMs that are still safe to use but really effective at certain tasks.”

This work was supported, in part, by the National Science Foundation, the Simons Foundation, the TILOS institute, and the U.S. Office of Naval Research. 

A new artificial intelligence control system enables soft robotic arms to learn a wide repertoire of motions and tasks once, then adjust to new scenarios on the fly, without needing retraining or sacrificing functionality. 

This breakthrough brings soft robotics closer to human-like adaptability for real-world applications, such as in assistive robotics, rehabilitation robots, and wearable or medical soft robots, by making them more intelligent, versatile, and safe.

The work was led by the Mens, Manus and Machina (M3S) interdisciplinary research group — a play on the Latin MIT motto “mens et manus,” or “mind and hand,” with the addition of “machina” for “machine” — within the Singapore-MIT Alliance for Research and Technology. Co-leading the project are researchers from the National University of Singapore (NUS), alongside collaborators from MIT and Nanyang Technological University in Singapore (NTU Singapore).

Unlike regular robots that move using rigid motors and joints, soft robots are made from flexible materials such as soft rubber and move using special actuators — components that act like artificial muscles to produce physical motion. While their flexibility makes them ideal for delicate or adaptive tasks, controlling soft robots has always been a challenge because their shape changes in unpredictable ways. Real-world environments are often complicated and full of unexpected disturbances, and even small changes in conditions — like a shift in weight, a gust of wind, or a minor hardware fault — can throw off their movements. 

Despite substantial progress in soft robotics, existing approaches often can only achieve one or two of the three capabilities needed for soft robots to operate intelligently in real-world environments: using what they’ve learned from one task to perform a different task, adapting quickly when the situation changes, and guaranteeing that the robot will stay stable and safe while adapting its movements. This lack of adaptability and reliability has been a major barrier to deploying soft robots in real-world applications until now.

In an open-access study titled “A general soft robotic controller inspired by neuronal structural and plastic synapses that adapts to diverse arms, tasks, and perturbations,” published Jan. 6 in Science Advances, the researchers describe how they developed a new AI control system that allows soft robots to adapt across diverse tasks and disturbances. The study takes inspiration from the way the human brain learns and adapts, and was built on extensive research in learning-based robotic control, embodied intelligence, soft robotics, and meta-learning.

The system uses two complementary sets of “synapses” — connections that adjust how the robot moves — working in tandem. The first set, known as “structural synapses”, is trained offline on a variety of foundational movements, such as bending or extending a soft arm smoothly. These form the robot’s built‑in skills and provide a strong, stable foundation. The second set, called “plastic synapses,” continually updates online as the robot operates, fine-tuning the arm’s behavior to respond to what is happening in the moment. A built-in stability measure acts like a safeguard, so even as the robot adjusts during online adaptation, its behavior remains smooth and controlled.

“Soft robots hold immense potential to take on tasks that conventional machines simply cannot, but true adoption requires control systems that are both highly capable and reliably safe. By combining structural learning with real-time adaptiveness, we’ve created a system that can handle the complexity of soft materials in unpredictable environments,” says MIT Professor Daniela Rus, co-lead principal investigator at M3S, director of the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), and co-corresponding author of the paper. “It’s a step closer to a future where versatile soft robots can operate safely and intelligently alongside people — in clinics, factories, or everyday lives.”

“This new AI control system is one of the first general soft-robot controllers that can achieve all three key aspects needed for soft robots to be used in society and various industries. It can apply what it learned offline across different tasks, adapt instantly to new conditions, and remain stable throughout — all within one control framework,” says Associate Professor Zhiqiang Tang, first author and co-corresponding author of the paper who was a postdoc at M3S and at NUS when he carried out the research and is now an associate professor at Southeast University in China (SEU China).

The system supports multiple task types, enabling soft robotic arms to execute trajectory tracking, object placement, and whole-body shape regulation within one unified approach. The method also generalizes across different soft-arm platforms, demonstrating cross-platform applicability. 

The system was tested and validated on two physical platforms — a cable-driven soft arm and a shape-memory-alloy–actuated soft arm — and delivered impressive results. It achieved a 44–55 percent reduction in tracking error under heavy disturbances; over 92 percent shape accuracy under payload changes, airflow disturbances, and actuator failures; and stable performance even when up to half of the actuators failed. 

“This work redefines what’s possible in soft robotics. We’ve shifted the paradigm from task-specific tuning and capabilities toward a truly generalizable framework with human-like intelligence. It is a breakthrough that opens the door to scalable, intelligent soft machines capable of operating in real-world environments,” says Professor Cecilia Laschi, co-corresponding author and principal investigator at M3S, Provost’s Chair Professor in the NUS Department of Mechanical Engineering at the College of Design and Engineering, and director of the NUS Advanced Robotics Centre.

This breakthrough opens doors for more robust soft robotic systems to develop manufacturing, logistics, inspection, and medical robotics without the need for constant reprogramming — reducing downtime and costs. In health care, assistive and rehabilitation devices can automatically tailor their movements to a patient’s changing strength or posture, while wearable or medical soft robots can respond more sensitively to individual needs, improving safety and patient outcomes.

The researchers plan to extend this technology to robotic systems or components that can operate at higher speeds and more complex environments, with potential applications in assistive robotics, medical devices, and industrial soft manipulators, as well as integration into real-world autonomous systems.

The research conducted at SMART was supported by the National Research Foundation Singapore under its Campus for Research Excellence and Technological Enterprise program.

It happens every day — a motorist heading across town checks a navigation app to see how long the trip will take, but they find no parking spots available when they reach their destination. By the time they finally park and walk to their destination, they’re significantly later than they expected to be.

Most popular navigation systems send drivers to a location without considering the extra time that could be needed to find parking. This causes more than just a headache for drivers. It can worsen congestion and increase emissions by causing motorists to cruise around looking for a parking spot. This underestimation could also discourage people from taking mass transit because they don’t realize it might be faster than driving and parking.

MIT researchers tackled this problem by developing a system that can be used to identify parking lots that offer the best balance of proximity to the desired location and likelihood of parking availability. Their adaptable method points users to the ideal parking area rather than their destination.

In simulated tests with real-world traffic data from Seattle, this technique achieved time savings of up to 66 percent in the most congested settings. For a motorist, this would reduce travel time by about 35 minutes, compared to waiting for a spot to open in the closest parking lot.

While they haven’t designed a system ready for the real world yet, their demonstrations show the viability of this approach and indicate how it could be implemented.

“This frustration is real and felt by a lot of people, and the bigger issue here is that systematically underestimating these drive times prevents people from making informed choices. It makes it that much harder for people to make shifts to public transit, bikes, or alternative forms of transportation,” says MIT graduate student Cameron Hickert, lead author on a paper describing the work.

Hickert is joined on the paper by Sirui Li PhD ’25; Zhengbing He, a research scientist in the Laboratory for Information and Decision Systems (LIDS); and senior author Cathy Wu, the Class of 1954 Career Development Associate Professor in Civil and Environmental Engineering (CEE) and the Institute for Data, Systems, and Society (IDSS) at MIT, and a member of LIDS. The research appears today in Transactions on Intelligent Transportation Systems.

Probable parking

To solve the parking problem, the researchers developed a probability-aware approach that considers all possible public parking lots near a destination, the distance to drive there from a point of origin, the distance to walk from each lot to the destination, and the likelihood of parking success.

The approach, based on dynamic programming, works backward from good outcomes to calculate the best route for the user.

Their method also considers the case where a user arrives at the ideal parking lot but can’t find a space. It takes into the account the distance to other parking lots and the probability of success of parking at each.

“If there are several lots nearby that have slightly lower probabilities of success, but are very close to each other, it might be a smarter play to drive there rather than going to the higher-probability lot and hoping to find an opening. Our framework can account for that,” Hickert says.

In the end, their system can identify the optimal lot that has the lowest expected time required to drive, park, and walk to the destination.

But no motorist expects to be the only one trying to park in a busy city center. So, this method also incorporates the actions of other drivers, which affect the user’s probability of parking success.

For instance, another driver may arrive at the user’s ideal lot first and take the last parking spot. Or another motorist could try parking in another lot but then park in the user’s ideal lot if unsuccessful. In addition, another motorist may park in a different lot and cause spillover effects that lower the user’s chances of success.

“With our framework, we show how you can model all those scenarios in a very clean and principled manner,” Hickert says.

Crowdsourced parking data

The data on parking availability could come from several sources. For example, some parking lots have magnetic detectors or gates that track the number of cars entering and exiting.

But such sensors aren’t widely used, so to make their system more feasible for real-world deployment, the researchers studied the effectiveness of using crowdsourced data instead.

For instance, users could indicate available parking using an app. Data could also be gathered by tracking the number of vehicles circling to find parking, or how many enter a lot and exit after being unsuccessful.

Someday, autonomous vehicles could even report on open parking spots they drive by.

“Right now, a lot of that information goes nowhere. But if we could capture it, even by having someone simply tap ‘no parking’ in an app, that could be an important source of information that allows people to make more informed decisions,” Hickert adds.

The researchers evaluated their system using real-world traffic data from the Seattle area, simulating different times of day in a congested urban setting and a suburban area. In congested settings, their approach cut total travel time by about 60 percent compared to sitting and waiting for a spot to open, and by about 20 percent compared to a strategy of continually driving to the next closet parking lot.

They also found that crowdsourced observations of parking availability would have an error rate of only about 7 percent, compared to actual parking availability. This indicates it could be an effective way to gather parking probability data.

In the future, the researchers want to conduct larger studies using real-time route information in an entire city. They also want to explore additional avenues for gathering data on parking availability, such as using satellite images, and estimate potential emissions reductions.

“Transportation systems are so large and complex that they are really hard to change. What we look for, and what we found with this approach, is small changes that can have a big impact to help people make better choices, reduce congestion, and reduce emissions,” says Wu.

This research was supported, in part, by Cintra, the MIT Energy Initiative, and the National Science Foundation.

Training for a clerical military role in France, Gustavo Barboza felt a spark he couldn’t ignore. He remembered his love of learning, which once guided him through two college semesters of mechanical engineering courses in his native Colombia, coupled with supplemental resources from MIT Open Learning’s OpenCourseWare. Now, thousands of miles away, he realized it was time to follow that spark again.

“I wasn’t ready to sit down in the classroom,” says Barboza, remembering his initial foray into higher education. “I left to try and figure out life. I realized I wanted more adventure.”

Joining the military in France in 2017 was his answer. For the first three years of service, he was very military-minded, only focused on his training and deployments. With more seniority, he took on more responsibilities, and eventually was sent to take a four-month training course on military correspondence and software. 

“I reminded myself that I like to study,” he says. “I started to go back to OpenCourseWare because I knew in the back of my mind that these very complete courses were out there.”

At that point, Barboza realized that military service was only a chapter in his life, and the next would lead him back to learning. He was still interested in engineering, and knew that MIT OpenCourseWare could help prepare him for what was next. 

He dove into OpenCourseWare’s free, online, open educational resources — which cover nearly the entire MIT curriculum — including classical mechanics, intro to electrical engineering, and single variable calculus with David Jerison, which he says was his most-visited resource. These allowed him to brush up on old skills and learn new ones, helping him tremendously in preparing for college entrance exams and his first-year courses. 

Now in his third year at Grenoble-Alpes University, Barboza studies electrical engineering, a shift from his initial interest in mechanical engineering.

“There is an OpenCourseWare lecture that explains all the specializations you can get into with electrical engineering,” he says. “They go from very natural things to things like microprocessors. What interests me is that if someone says they are an electrical engineer, there are so many different things they could be doing.” 

At this point in his academic career, Barboza is most interested in microelectronics and the study of radio frequencies and electromagnetic waves. But he admits he has more to learn and is open to where his studies may take him. 

MIT OpenCourseWare remains a valuable resource, he says. When thinking about his future, he checks out graduate course listings and considers the different paths he might take. When he is having trouble with a certain concept, he looks for a lecture on the subject, undeterred by the differences between French and U.S. conventions.  

“Of course, the science doesn't change, but the way you would write an equation or draw a circuit is different at my school in France versus what I see from MIT. So, you have to be careful,” he explains. “But it is still the first place I visit for problem sets, readings, and lecture notes. It’s amazing.”

The thoroughness and openness of MIT Open Learning’s courses and resources — like OpenCourseWare — stand out to Barboza. In the wide world of the internet, he has found resources from other universities, but he says their offerings are not as robust. And in a time of disinformation and questionable sources, he appreciates that MIT values transparency, accessibility, and knowledge. 

“Human knowledge has never been more accessible,” he says. “MIT puts coursework online and says, ‘here’s what we do.’ As long as you have an internet connection, you can learn all of it.”

“I just feel like MIT OpenCourseWare is what the internet was originally for,” Barboza continues. “A network for sharing knowledge. I’m a big fan.”

Explore lifelong learning opportunities from MIT, including courses, resources, and professional programs, on MIT Learn.

Many of the latest large language models (LLMs) are designed to remember details from past conversations or store user profiles, enabling these models to personalize responses.

But researchers from MIT and Penn State University found that, over long conversations, such personalization features often increase the likelihood an LLM will become overly agreeable or begin mirroring the individual’s point of view.

This phenomenon, known as sycophancy, can prevent a model from telling a user they are wrong, eroding the accuracy of the LLM’s responses. In addition, LLMs that mirror someone’s political beliefs or worldview can foster misinformation and distort a user’s perception of reality.

Unlike many past sycophancy studies that evaluate prompts in a lab setting without context, the MIT researchers collected two weeks of conversation data from humans who interacted with a real LLM during their daily lives. They studied two settings: agreeableness in personal advice and mirroring of user beliefs in political explanations.

Although interaction context increased agreeableness in four of the five LLMs they studied, the presence of a condensed user profile in the model’s memory had the greatest impact. On the other hand, mirroring behavior only increased if a model could accurately infer a user’s beliefs from the conversation.

The researchers hope these results inspire future research into the development of personalization methods that are more robust to LLM sycophancy.

“From a user perspective, this work highlights how important it is to understand that these models are dynamic and their behavior can change as you interact with them over time. If you are talking to a model for an extended period of time and start to outsource your thinking to it, you may find yourself in an echo chamber that you can’t escape. That is a risk users should definitely remember,” says Shomik Jain, a graduate student in the Institute for Data, Systems, and Society (IDSS) and lead author of a paper on this research.

Jain is joined on the paper by Charlotte Park, an electrical engineering and computer science (EECS) graduate student at MIT; Matt Viana, a graduate student at Penn State University; as well as co-senior authors Ashia Wilson, the Lister Brothers Career Development Professor in EECS and a principal investigator in LIDS; and Dana Calacci PhD ’23, an assistant professor at the Penn State. The research will be presented at the ACM CHI Conference on Human Factors in Computing Systems.

Extended interactions

Based on their own sycophantic experiences with LLMs, the researchers started thinking about potential benefits and consequences of a model that is overly agreeable. But when they searched the literature to expand their analysis, they found no studies that attempted to understand sycophantic behavior during long-term LLM interactions.

“We are using these models through extended interactions, and they have a lot of context and memory. But our evaluation methods are lagging behind. We wanted to evaluate LLMs in the ways people are actually using them to understand how they are behaving in the wild,” says Calacci.

To fill this gap, the researchers designed a user study to explore two types of sycophancy: agreement sycophancy and perspective sycophancy.

Agreement sycophancy is an LLM’s tendency to be overly agreeable, sometimes to the point where it gives incorrect information or refuses the tell the user they are wrong. Perspective sycophancy occurs when a model mirrors the user’s values and political views.

“There is a lot we know about the benefits of having social connections with people who have similar or different viewpoints. But we don’t yet know about the benefits or risks of extended interactions with AI models that have similar attributes,” Calacci adds.

The researchers built a user interface centered on an LLM and recruited 38 participants to talk with the chatbot over a two-week period. Each participant’s conversations occurred in the same context window to capture all interaction data.

Over the two-week period, the researchers collected an average of 90 queries from each user.

They compared the behavior of five LLMs with this user context versus the same LLMs that weren’t given any conversation data.

“We found that context really does fundamentally change how these models operate, and I would wager this phenomenon would extend well beyond sycophancy. And while sycophancy tended to go up, it didn’t always increase. It really depends on the context itself,” says Wilson.

Context clues

For instance, when an LLM distills information about the user into a specific profile, it leads to the largest gains in agreement sycophancy. This user profile feature is increasingly being baked into the newest models.

They also found that random text from synthetic conversations also increased the likelihood some models would agree, even though that text contained no user-specific data. This suggests the length of a conversation may sometimes impact sycophancy more than content, Jain adds.

But content matters greatly when it comes to perspective sycophancy. Conversation context only increased perspective sycophancy if it revealed some information about a user’s political perspective.

To obtain this insight, the researchers carefully queried models to infer a user’s beliefs then asked each individual if the model’s deductions were correct. Users said LLMs accurately understood their political views about half the time.

“It is easy to say, in hindsight, that AI companies should be doing this kind of evaluation. But it is hard and it takes a lot of time and investment. Using humans in the evaluation loop is expensive, but we’ve shown that it can reveal new insights,” Jain says.

While the aim of their research was not mitigation, the researchers developed some recommendations.

For instance, to reduce sycophancy one could design models that better identify relevant details in context and memory. In addition, models can be built to detect mirroring behaviors and flag responses with excessive agreement. Model developers could also give users the ability to moderate personalization in long conversations.

“There are many ways to personalize models without making them overly agreeable. The boundary between personalization and sycophancy is not a fine line, but separating personalization from sycophancy is an important area of future work,” Jain says.

“At the end of the day, we need better ways of capturing the dynamics and complexity of what goes on during long conversations with LLMs, and how things can misalign during that long-term process,” Wilson adds.

A broken motor in an automated machine can bring production on a busy factory floor to a halt. If engineers can’t find a replacement part, they may have to order one from a distributor hundreds of miles away, leading to costly production delays.

It would be easier, faster, and cheaper to make a new motor onsite, but fabricating electric machines typically requires specialized equipment and complicated processes, which restricts production to a few manufacturing centers.

In an effort to democratize the manufacturing of complex devices, MIT researchers have developed a multimaterial 3D-printing platform that could be used to fully print electric machines in a single step.

They designed their system to process multiple functional materials, including electrically conductive materials and magnetic materials, using four extrusion tools that can handle varied forms of printable material. The printer switches between extruders, which deposit material by squeezing it through a nozzle as it fabricates a device one layer at a time.

The researchers used this system to produce a fully 3D-printed electric linear motor in a matter of hours using five materials. They only needed to perform one post-processing step for the motor to be fully functional.

The assembled device performed as well or better than similar motors that require more complex fabrication methods or additional post-processing steps.

In the long run, this 3D printing platform could be used to rapidly fabricate customizable electronic components for robots, vehicles, or medical equipment with much less waste.

“This is a great feat, but it is just the beginning. We have an opportunity to fundamentally change the way things are made by making hardware onsite in one step, rather than relying on a global supply chain. With this demonstration, we’ve shown that this is feasible,” says Luis Fernando Velásquez-García, a principal research scientist in MIT’s Microsystems Technology Laboratories (MTL) and senior author of a paper describing the 3D-printing platform, which appears today in Virtual and Physical Prototyping.

He is joined on the paper by electrical engineering and computer science (EECS) graduate students Jorge Cañada, who is the lead author, and Zoey Bigelow.

More materials

The researchers focused on extrusion 3D printing, a tried-and-true method that involves squirting material through a nozzle to fabricate an object one layer at a time.

To fabricate an electric machine, the researchers needed to be able to switch between multiple materials that offer different functionalities. For instance, the device would need an electrically conductive material to carry electric current and hard magnetic materials to generate magnetic fields for efficient energy conversion.

Most multimaterial extrusion 3D printing systems can only switch between two materials that come in the same form, such as filament or pellets, so the researchers had to design their own. They retrofit an existing printer with four extruders that can each handle a different form of feedstock.

They carefully designed each extruder to balance the requirements and limitations of the material. For instance, the electrically conductive material must be able to harden without the use of too much heat or UV light because this can degrade the dielectric material.

At the same time, the best-performing electrically conductive materials come in the form of inks which are extruded using a pressure system. This process has vastly different requirements than standard extruders that use heated nozzles to squirt melted filament or pellets.

“There were significant engineering challenges. We had to figure out how to marry together many different expressions of the same printing method — extrusion — seamlessly into one platform,” Velásquez-García says.

The researchers utilized strategically placed sensors and a novel control framework so each tool is picked up and put down consistently by the platform’s robotic arms, and so each nozzle moves precisely and predictably.

This ensures each layer of material lines up properly — even a slight misalignment can derail the performance of the finished machine.

Making a motor

After perfecting the printing platform, the researchers fabricated a linear motor, which generates straight-line motion (as opposed to a rotating motor, like the one in a car). Linear motors are used in applications like pick-and-place robotics, optical systems, and baggage conveyers.

They fabricated the motor in about three hours and only needed to magnetize the hard magnetic materials after printing to enable full functionality. The researchers estimate total material costs would be about 50 cents per device. Their 3D-printed motor was able to generate several times more actuation than a common type of linear engine that relies on complex hydraulic amplifiers. 

“Even though we are excited by this engine and its performance, we are equally inspired because this is just an example of so many other things to come that could dramatically change how electronics are manufactured,” says Velásquez-García.

In the future, the researchers want to integrate the magnetization step into the multimaterial extrusion process, demonstrate the fabrication of fully 3D-printed rotary electrical motors, and add more tools to the platform to enable monolithic fabrication of more complex electronic devices.

This research is funded, in part, by Empiriko Corporation and the La Caixa Foundation.

An earthquake typically sets off ruptures that ripple out from its underground origins. But on rare occasions, seismologists have observed quakes that reverse course, further shaking up areas that they passed through only seconds before. These “boomerang” earthquakes often occur in regions with complex fault systems. But a new study by MIT researchers predicts that such ricochet ruptures can occur even along simple faults.

The study, which appears today in the journal AGU Advances, reports that boomerang earthquakes can happen along a simple fault under several conditions: if the quake propagates out in just one direction, over a large enough distance, and if friction along the rupturing fault builds and subsides rapidly during the quake. Under these conditions, even a simple straight fault, like some segments of the San Andreas fault in California, could experience a boomerang quake.

These newly identified conditions are relatively common, suggesting that many earthquakes that have occurred along simple faults may have experienced a boomerang effect, or what scientists term “back-propagating fronts.”

“Our work suggests that these boomerang quakes may have been undetected in a number of cases,” says study author Yudong Sun, a graduate student in MIT’s Department of Earth, Atmospheric and Planetary Sciences (EAPS). “We do think this behavior may be more common than we have seen so far in the seismic data.”

The new results could help scientists better assess future hazards in simple fault zones where boomerang quakes could potentially strike twice.

“In most cases, it would be impossible for a person to tell that an earthquake has propagated back just from the ground shaking, because ground motion is complex and affected by many factors,” says co-author Camilla Cattania, the Cecil and Ida Green Career Development Professor of Geophysics at MIT. “However, we know that shaking is amplified in the direction of rupture, and buildings would shake more in response. So there is a real effect in terms of the damage that results. That’s why understanding where these boomerang events could occur matters.”

Keep it simple

There have been a handful of instances where scientists have recorded seismic data suggesting that a quake reversed direction. In 2016, an earthquake in the middle of the Atlantic Ocean rippled eastward, and then seconds later richocheted back west. Similar return rumblers may have occurred in 2011 during the magnitude 9 earthquake in Tohoku, Japan, and in 2023 during the destructive magnitude 7.8 quake in Turkey and Syria, among others.

These events took place in various fault regions, from complex zones of multiple intersecting fault lines to regions with just a single, straight fault. While seismologists have assumed that such complex quakes would be more likely to occur in multifault systems, the rare examples along simple faults got Sun and Cattania wondering: Could an earthquake reverse course along a simple fault? And if so, what could cause such a bounce-back in a seemingly simple system?

“When you see this boomerang-like behavior, it is tempting to explain this in terms of some complexity in the Earth,” Cattania says. “For instance, there may be many faults that interact, with earthquakes jumping between fault segments, or fault surfaces with prominent kinks and bends. In many cases, this could explain back-propagating behavior. But what we found was, you could have a very simple fault and still get this complex behavior.”

Faulty friction

In their new study, the team looked to simulate an earthquake along a simple fault system. In geology, a fault is a crack or fracture that runs through the Earth’s crust. An earthquake begins when the stress between rocks on either side of the fault, suddenly decreases, and one side slides against the other, setting off seismic waves that rupture rocks all along the fault. This seismic activity, which initiates deep in the crust, can sometimes reach and shake up the surface.

Cattania and Sun used a computer model to represent the fundamental physics at play during an earthquake along a simple fault. In their model, they simulated the Earth’s crust as a simple elastic material, in which they embedded a single straight fault. They then simulated how the fault would exhibit an earthquake under different scenarios. For instance, the team varied the length of the fault and the location of the quake’s initation point below the surface, as well as whether the quake traveled in one versus two directions.

Over multiple simulations, they observed that only the unilateral quakes — those that traveled in one direction — exhibited a boomerang effect. Specifically, these quakes seemed to include a type that seismologists term “back-propagating” events, in which the rumbler splits at some point along the fault, partly continuing in the same direction and partly reversing back the way it came.

“When you look at a simulation, sometimes you don’t fully understand what causes a given behavior,” Cattania says. “So we developed mathematical models to understand it. And we went back and forth, to ultimately develop a simple theory that tells you should only see this back-propagation under these certain conditions.”

Those conditions, as the team’s new theory lays out, have to do with the friction along the fault. In standard earthquake physics, it’s generally understood that an earthquake is triggered when the stress built up between rocks on either side of a fault, is suddenly released. Rocks slide against each other in response, decreasing a fault’s friction. The reduction in fault friction creates a positive feedback that facilitates further sliding, sustaining the earthquake.

However, in their simulations, the team observed that when a quake travels along a fault in one direction, it can back-propagate when friction along the fault goes down, then up, and then down again.

“When the quake propagates in one direction, it produces a “breaking’’ effect that reduces the sliding velocity, increases friction, and allows only a narrow section of the fault to slide at a time,” Cattania says. “The region behind the quake, which stops sliding, can then rupture again, because it has accumulated more stress to slide again.”

The team found that, in addition to traveling in one direction and along a fault with changing friction, a boomerang is likely to occur if a quake has traveled over a large enough distance.

“This implies that large earthquakes are not simply ‘scaled-up’ versions of small earthquakes, but instead they have their own unique rupture behavior,” Sun says.

The team suspects that back-propagating quakes may be more common than scientists have thought, and they may occur along simple, straight faults, which are typically older than more complex fault systems.

“You shouldn’t only expect this complex behavior on a young, complex fault system. You can also see it on mature, simple faults,” Cattania says. “The key open question now is how often rupture reversals, or ‘boomerang’ earthquakes, occur in nature. Many observational studies so far have used methods that can’t detect back-propagating fronts. Our work motivates actively looking for them, to further advance our understanding of earthquake physics and ultimately mitigate seismic risk.”

Seven MIT researchers are among the 130 new members and 28 international members recently elected to the National Academy of Engineering (NAE) for 2026. Twelve additional MIT alumni were also elected as new members.

One of the highest professional distinctions for engineers, membership in the NAE is given to individuals who have made outstanding contributions to “engineering research, practice, or education,” and to “the pioneering of new and developing fields of technology, making major advancements in traditional fields of engineering, or developing/implementing innovative approaches to engineering education.”

The seven MIT electees this year include:

Moungi Gabriel Bawendi, the Lester Wolfe Professor of Chemistry in the Department of Chemistry, was honored for the synthesis and characterization of semiconductor quantum dots and their applications in displays, photovoltaics, and biology.

Charles Harvey, a professor in the Department of Civil and Environmental Engineering, was honored for contributions to hydrogeology regarding groundwater arsenic contamination, transport, and consequences.

Piotr Indyk, the Thomas D. and Virginia W. Cabot Professor in the Department of Electrical Engineering and Computer Science and a member of the Computer Science and Artificial Intelligence Laboratory, was honored for contributions to approximate nearest neighbor search, streaming, and sketching algorithms for massive data processing.

John Henry Lienhard, the Abdul Latif Jameel Professor of Water and Mechanical Engineering in the Department of Mechanical Engineering, was honored for advances and technological innovations in desalination.

Ram Sasisekharan, the Alfred H. Caspary Professor of Biological Physics and Physics in the Department of Biological Engineering, was honored for discovering the U.S. heparin contaminant in 2008 and creating clinical antibodies for Zika, dengue, SARS-CoV-2, and other diseases.

Frances Ross, the TDK Professor in the Department of Materials Science and Engineering, was honored for ultra-high vacuum and liquid-cell transmission electron microscopies and their worldwide adoptions for materials research and semiconductor technology development.

Zoltán Sandor Spakovszky SM ’99, PhD ’01, the T. Wilson (1953) Professor in Aeronautics in the Department of Aeronautics and Astronautics, was honored for contributions, through rigorous discoveries and advancements, in aeroengine aerodynamic and aerostructural stability and acoustics.

“Each of the MIT faculty and alumni elected to the National Academy of Engineering has made extraordinary contributions to their fields through research, education, and innovation,” says Paula T. Hammond, dean of the School of Engineering and Institute Professor in the Department of Chemical Engineering. "They represent the breadth of excellence we have here at MIT. This honor reflects the impact of their work, and I’m proud to celebrate their achievement and offer my warmest congratulations.”

Twelve additional alumni were elected to the National Academy of Engineering this year. They are: Anne Hammons Aunins PhD ’91; Lars James Blackmore PhD ’07; John-Paul Clarke ’91, SM ’92, SCD ’97; Michael Fardis SM ’77, SM ’78, PhD ’79; David Hays PhD ’98; Stephen Thomas Kent ’76, EE ’78, ENG ’78, PhD ’81; Randal D. Koster SM ’85, SCD ’88; Fred Mannering PhD ’83; Peyman Milanfar SM ’91, EE ’93, ENG ’93, PhD ’93; Amnon Shashua PhD ’93; Michael Paul Thien SCD ’88; and Terry A. Winograd PhD ’70.

Dean of Engineering Paula Hammond ’84 PhD ’93 made a resounding call for the MIT community to “embrace endless hope” and “never stop looking forward,” in a keynote address at the Institute’s annual MLK Celebration on Wednesday, Feb. 11.

“We each have a role to play in contributing to our future, and we each must embrace endless hope and continuously renew our faith in ourselves to accomplish that dream,” Hammond said, to an audience of hundreds at the event.

She added: “Whether it is through caring for those in our community, teaching others, providing inspiration, leadership, or critical support to others in their moment of need, we provide support for one another on our journey … It is that future that will feed the optimism and faith that we need to move forward, to inspire and encourage, and to never stop looking forward.”

The MLK Celebration is an annual tribute to the life and legacy of Martin Luther King Jr., and is always thematically organized around a quotation of King’s. This year, that passage was, “We must accept finite disappointment, but never lose infinite hope.”

Hammond and multiple other speakers at the event organized their remarks around that idea, while weaving in personal reflections about the importance of community, family, and mentorship.

As Hammond noted, “We can lay the path toward a better, greater time with the steps that we take today even in the face of incredible disappointment, shock and disruption.” She added: “Principles founded in fear, ignorance, or injustice ultimately fail because they do not meet the needs of a growing and prosperous nation and world.”

The event, which took place in MIT’s Walker Memorial (Building 50), featured remarks by students, staff, and campus leaders, as well as musical performances by the recently reconstituted MIT Gospel Choir. (Listen to one of those performances by clicking on the player at the end of this article.)

MIT President Sally A. Kornbluth provided introductory remarks, noting that this year’s event was occurring during “a time when feeling fractured, isolated, and pitted against each other feels exhaustingly routine. A time when it’s easy to feel discouraged.” As such, she added, “the solace we take from [coming together at this event] couldn’t be more relevant now.”

Kornbluth also offered laudatory thoughts about Hammond, a highly accomplished research scientist who has held numerous leadership roles at MIT and elsewhere. Hammond, a chemical engineer, was named dean of the MIT School of Engineering in December. Prior to that, she has served as vice provost for faculty, from 2023 to 2025, and head of the Department of Chemical Engineering, from 2015 to 2023. In honor of her accomplishments, Hammond was named an Institute Professor, MIT’s highest faculty honor. A member of MIT’s Koch Institute for Integrative Cancer Research, Hammond has developed polymers and nanoscale materials with multiple applications, including drug delivery, imaging, and even battery advances.

Hammond was awarded the National Medal of Technology and Innovation in 2024. That year she also received MIT’s Killian Award, for faculty achievement. And she has earned the rare distinction of having been elected to all three national academies — the National Academy of Engineering, the National Academy of Medicine, and the National Academy of Sciences.

“I’ve never met anyone who better represents MIT’s highest values and aspirations than Paula Hammond,” Kornbluth said, citing both Hammond’s record of academic excellence and Institute service.

Among other things, Kornbluth observed, “Paula has been a longtime champion of MIT’s culture of openness to people and ideas from everywhere. In fact, it’s hard to think of anyone more open to sharing what she knows — and more interested in hearing your point of view. And the respect she shows to everyone — no matter their job or background — is an example for us all.”

Michael Ewing ’27, a mechanical engineering major, provided welcoming remarks while introducing the speakers as well as the MLK Celebration planning committee.

Ewing noted that the event remains “extremely and vitally important” to the MIT community, and reflected on the meaning of this year’s motif, for individuals and larger communities.

“Dr. King’s hope constitutes the belief that one can make things better, even when current conditions are poor,” Ewing said. “In the face of adversity, we must remain connected to what’s most important, be grateful for both the challenges and the opportunities, and hold on to the long-term belief that no matter what, no matter what, there’s an opportunity for us to learn, grow, and improve.”

The annual MLK Celebration also highlighted further reflections from students and staff on King’s life and legacy and the value of his work.

“Everyone that has fought for a greater good in this world has left the battle without something that they came with,” said Oluwadara Deru, a senior in mechanical engineering and the featured undergraduate speaker. “But what they gained is invaluable.”

Ekua Beneman, a graduate student in chemistry, offered thoughts relating matters of academic achievement, and helping others in a university setting, to the larger themes of the celebration.

“Hope is not pretending disappointment doesn’t exist,” Beneman said. “Hope is choosing to pass forward what was once given to you. At a place like MIT, infinite hope looks like mentorship. It looks like making space. It looks like sharing knowledge instead of guarding or gatekeeping it. If we truly want to honor Dr. King’s legacy, beyond this beautiful celebration today, we do it by choosing community, mentorship, and hope in action.”

Denzil Streete, associate dean and director of the Office of Graduate Education, related the annual theme to everyday life at the Institute, as well as social life everywhere.

“Hope lies in small, often uncelebrated acts,” Streete said. “Showing up. Being present. Responding with patience. Translating complicated processes into next steps. Making one more call. Sending one more email.”

He concluded: “See your daily work as moral work … Every day, through joy and care, we choose infinite hope, for our students, and for one another.”

Reverend Thea Keith-Lucas, chaplain to the Institute and associate dean in the Office of Religious, Spiritual, and Ethical Life, offered both an invocation and a benediction at the event.

The annual celebration includes the Dr. Martin Luther King Jr. Leadership Awards Recipients, given this year to Melissa Smith PhD ’12, Fred Harris, Carissma McGee, Janine Medrano, and Edwin Marrero.

For all the turbulence in the world, Hammond said toward the conclusion of her address, people can continue to make progress in their own communities, and can be intentional about focusing, in part, on the possibilities of progress ahead.

At MIT, Hammond noted, “The commitment of our faculty, students, and staff to continuously learn, to ask deep questions and to apply our knowledge, our perspectives and our insights to the biggest world problems is something that gives me infinite hope and optimism for the future.”

In many academic circles, innovation is imagined as a lab-to-market pipeline that travels through patent filings, venture rounds, and coastal research hubs. But a growing movement inside U.S. universities is pushing students toward a different frontier: solving real engineering problems alongside rural communities whose challenges directly shape national food security. 

A compelling example of this shift can be found in the story of Kiyoko “Kik” Hayano, a second-year mechanical engineering student at MIT, and her work through MIT D-Lab with Keo Fish Farms, a commercial aquaculture operation in the Arkansas Delta.

Hayano’s journey — from a small, windswept town in rural Wyoming to MIT’s campus in Cambridge, Massachusetts, and on to a working Arkansas fish farm — offers a tangible glimpse into how applied engineering, academic partnerships, and on-the-ground innovation can create new models for regenerative agriculture in the United States.

Wyoming childhood and an engineering dream

Hayano grew up in Powell, Wyoming (population ~6,400), a community defined by agriculture, water scarcity, and long distances. Her early interests in gardening with her grandmother and tinkering with irrigation projects through her high school’s agricultural center formed the foundation for a more ambitious goal: studying mechanical engineering at MIT.

That ambition paid off. Shortly after arriving in Cambridge, Hayano connected with MIT D-Lab, a program founded to co-create engineering solutions with communities, rather than for them — especially in regions facing poverty, resource constraints, or climate-related disruptions. For many MIT students, D-Lab is their entry point into field-based development work across Africa, Latin America, and Southeast Asia. Increasingly, however, the program has expanded its domestic mission to include rural areas of the United States experiencing food, water, and energy insecurity.

MIT D-Lab meets the Arkansas Delta

That domestic shift set the stage for a new joint effort. In 2024, Keo Fish Farms — a commercial aquaculture farm near Keo, Arkansas — contacted D-Lab seeking technical collaboration on a growing water quality challenge. The farm had begun to observe elevated iron levels in its groundwater, leading to fish mortality events during peak summer conditions. The problem was both biological and mechanical: Aquaculture species like hybrid striped bass and triploid grass carp require consistent, clean water inputs, and well systems tapping iron-rich geologic layers were compromising fish health, hatchery performance, and long-term viability.

Kendra Leith, MIT D-Lab associate director for research, saw an opportunity. The Delta region represents a collision of three major realities that matter deeply to both public policy and academic research: high-value protein production, aging or inadequate water infrastructure, and generational rural decline.

For Hayano, the chance to work on an important engineering problem with environmental, agricultural, and economic implications was exactly why she chose mechanical engineering in the first place.

Applied engineering in a living laboratory

When Hayano arrived at Keo Fish Farms, the project was structured as a co-creative engineering engagement — D-Lab’s core model. She documented the existing water intake system, analyzed the well depth relative to geological iron strata, and evaluated filtration options including aeration, sedimentation, and emerging biochar-based media.

The collaboration generated three immediate academic values. First, the team reviewed real constraints, a process known as ground truthing. Constraints in this situation included iron levels that shift seasonally, capital budgets that do not assume infinite funding, and labor cycles tied to harvest seasons. The team then scoped out the technology that might be used to mitigate problem areas. Iron-reduction solutions ranged from drilling deeper wells to incorporating biochar and other regenerative filtration mediums capable of binding contaminants while improving soil and plant health elsewhere on the farm. Finally, they reviewed policy relevance: Water quality in aquaculture sits at the intersection of U.S. Department of Agriculture (USDA) conservation, Environmental Protection Agency (EPA) water standards, climate-driven aquifer variability, and domestic protein security — issues central to U.S. food systems.

Leith notes that “the most transformative experiences happen when students and communities learn from one another.” The Keo project, she adds, is an example of how domestic food production systems can act as test beds for innovation that previously would have been deployed exclusively abroad.

Regenerative agriculture as a national opportunity

While Keo Fish Farms played a supporting role in the narrative, the project highlighted a broader challenge and opportunity: Can U.S. aquaculture transition toward regenerative agriculture principles?

Regenerative agriculture — long associated with row crops, grazing systems, and soil carbon — rarely includes aquaculture in the national conversation. Yet aquaculture sits at the nexus of water chemistry, nutrient cycling, renewable energy integration, biochar and filtration research, protein production, and greenhouse gas mitigation.

Hayano’s work helped illuminate that regenerative aquaculture will likely depend on regenerative water systems, where filtration, biochar, solar energy, and nutrient reuse form a closed-loop infrastructure, rather than a linear extract–use–discharge model.

D-Lab’s domestic projects increasingly intersect with this space, creating pathways for MIT students and faculty to collaborate with USDA, the U.S. Department of Energy (DoE), and National Science Foundation (NSF) priorities around rural innovation, renewable energy, and water systems engineering.

The role of industry partners: less spotlight, more signal

Keo Fish Farms’ involvement served as a platform — not a spotlight — for the engineering and policy implications emerging from the project. The farm provided three critical ingredients academic institutions often lack: a real commercial engineering problem with economic consequences, a living laboratory for field research and prototyping, and a pathway for future regenerative adoption at scale.

The farm’s leadership has stated that its long-term goal is to become a first-in-class demonstration site for regenerative aquaculture in the United States, combining advanced iron and sediment filtration, biochar production from local rice hull waste streams, renewable solar energy systems, water recycling and nutrient recovery, reduced chemical inputs, and habitat and biodiversity considerations.

To be sure, the D-Lab collaboration did not solve that entire puzzle, but it created the blueprint for a pathway, showing how academic partnerships can accelerate regenerative transitions in rural U.S. agriculture and aquaculture systems.

Lessons for universities and policymakers

For universities, the Keo–MIT D-Lab partnership offers a replicable model for experiential learning for STEM students, field-based regenerative research, technology validation in live agricultural systems, and cross-disciplinary collaboration. And for federal and state policymakers, it illustrates how rural communities can serve as innovation sites, why water infrastructure modernization matters to food security, how regenerative agriculture can expand beyond soil and grazing, and why public-private-academic partnerships deserve new funding pathways.

All of this aligns with emerging priorities at the USDA, DoE, NSF, and EPA around sustainability, climate resilience, and domestic protein systems.

For Hayano, the experience reinforced that engineering careers can be rooted not only in Silicon Valley labs or aerospace firms, but also in overlooked rural systems that feed the country. 

“I’m really grateful for the experience,” she reflected after the project. “It opened my eyes to how engineering can support sustainable food systems and rural communities.”

The sentiment echoes a broader trend among students seeking careers at the intersection of technology, environment, and public good. Whether Hayano returns to the Arkansas Delta or not, her path captures something deeply relevant to America’s innovation story: talent emerging from rural places, innovating at world-class institutions, and returning engineering capacity back into the country’s agricultural heartland.

It is, in many ways, a modern form of the American dream — one grounded not in abstraction, but in water, food, soil, and the systems that will define our next century.

Industrial yeasts are a powerhouse of protein production, used to manufacture vaccines, biopharmaceuticals, and other useful compounds. In a new study, MIT chemical engineers have harnessed artificial intelligence to optimize the development of new protein manufacturing processes, which could reduce the overall costs of developing and manufacturing these drugs.

Using a large language model (LLM), the MIT team analyzed the genetic code of the industrial yeast Komagataella phaffii — specifically, the codons that it uses. There are multiple possible codons, or three-letter DNA sequences, that can be used to encode a particular amino acid, and the patterns of codon usage are different for every organism.

The new MIT model learned those patterns for K. phaffii and then used them to predict which codons would work best for manufacturing a given protein. This allowed the researchers to boost the efficiency of the yeast’s production of six different proteins, including human growth hormone and a monoclonal antibody used to treat cancer.

“Having predictive tools that consistently work well is really important to help shorten the time from having an idea to getting it into production. Taking away uncertainty ultimately saves time and money,” says J. Christopher Love, the Raymond A. and Helen E. St. Laurent Professor of Chemical Engineering at MIT, a member of the Koch Institute for Integrative Cancer Research, and faculty co-director of the MIT Initiative for New Manufacturing (MIT INM).

Love is the senior author of the new study, which appears this week in the Proceedings of the National Academy of Sciences. Former MIT postdoc Harini Narayanan is the paper’s lead author.

Codon optimization

Yeast such as K. phaffii and Saccharomyces cerevisiae (baker’s yeast) are the workhorses of the biopharmaceutical industry, producing billions of dollars of protein drugs and vaccines every year.

To engineer yeast for industrial protein production, researchers take a gene from another organism, such as the insulin gene, and modify it so that the microbe will produce it in large quantities. This requires coming up with an optimal DNA sequence for the yeast cells, integrating it into the yeast’s genome, devising favorable growth conditions for it, and finally purifying the end product.

For new biologic drugs — large, complex drugs produced by living organisms — this development process might account for 15 to 20 percent of the overall cost of commercializing the drug.

“Today, those steps are all done by very laborious experimental tasks,” Love says. “We have been looking at the question of where could we take some of the concepts that are emerging in machine learning and apply them to make different aspects of the process more reliable and simpler to predict.”

In this study, the researchers wanted to try to optimize the sequence of DNA codons that make up the gene for a protein of interest. There are 20 naturally occurring amino acids, but 64 possible codon sequences, so most of these amino acids can be encoded by more than one codon. Each codon corresponds to a unique transfer RNA (tRNA) molecule, which carries the correct amino acid to the ribosome, where amino acids are strung together into proteins.

Different organisms use each of these codons at different rates, and designers of engineered proteins often optimize the production of their proteins by choosing the codons that occur the most frequently in the host organism. However, this doesn’t necessarily produce the best results. If the same codon is always used to encode arginine, for example, the cell may run low on the tRNA molecules that correspond to that codon.

To take a more nuanced approach, the MIT team deployed a type of large language model known as an encoder-decoder. Instead of analyzing text, the researchers used it to analyze DNA sequences and learn the relationships between codons that are used in specific genes.

Their training data, which came from a publicly available dataset from the National Center for Biotechnology Information, consisted of the amino acid sequences and corresponding DNA sequences for all of the approximately 5,000 proteins naturally produced by K. phaffii.

“The model learns the syntax or the language of how these codons are used,” Love says. “It takes into account how codons are placed next to each other, and also the long-distance relationships between them.”

Once the model was trained, the researchers asked it to optimize the codon sequences of six different proteins, including human growth hormone, human serum albumin, and trastuzumab, a monoclonal antibody used to treat cancer.

They also generated optimized sequences of these proteins using four commercially available codon optimization tools. The researchers inserted each of these sequences into K. phaffii cells and measured how much of the target protein each sequence generated. For five of the six proteins, the sequences from the new MIT model worked the best, and for the sixth, it was the second-best.

“We made sure to cover a variety of different philosophies of doing codon optimization and benchmarked them against our approach,” Narayanan says. “We’ve experimentally compared these approaches and showed that our approach outperforms the others.”

Learning the language of proteins

K. phaffii, formerly known as Pichia pastoris, is used to produce dozens of commercial products, including insulin, hepatitis B vaccines, and a monoclonal antibody used to treat chronic migraines. It is also used in the production of nutrients added to foods, such as hemoglobin.

Researchers in Love’s lab have started using the new model to optimize proteins of interest for K. phaffii, and they have made the code available for other researchers who wish to use it for K. phaffii or other organisms.

The researchers also tested this approach on datasets from different organisms, including humans and cows. Each of the resulting models generated different predictions, suggesting that species-specific models are needed to optimize codons of target proteins.

By looking into the inner workings of the model, the researchers found that it appeared to learn some of the biological principles of how the genome works, including things that the researchers did not teach it. For example, it learned not to include negative repeat elements — DNA sequences that can inhibit the expression of nearby genes. The model also learned to categorize amino acids based on traits such as hydrophobicity and hydrophilicity.

“Not only was it learning this language, but it was also contextualizing it through aspects of biophysical and biochemical features, which gives us additional confidence that it is learning something that’s actually meaningful and not simply an optimization of the task that we gave it,” Love says.

The research was funded by the Daniel I.C. Wang Faculty Research Innovation Fund at MIT, the MIT AltHost Research Consortium, the Mazumdar-Shaw International Oncology Fellowship, and the Koch Institute.

America has been making steel from iron ore the same way for hundreds of years. Unfortunately, it hasn’t been making enough of it. Today the U.S. is the world’s largest steel importer, relying on other countries to produce a material that serves as the backbone of our society.

That’s not to say the U.S. is alone: Globally, most steel today is made in enormous, multi-billion-dollar plants using a coal-based process that hasn’t changed much in 300 years.

Now Hertha Metals, founded by CEO Laureen Meroueh SM ’18, PhD ’20, is scaling up a new steel production system powered by natural gas and electricity. The process, which can also run on hydrogen, uses a continuous electric arc furnace within which iron ore of any grade and format is reduced and carburized into molten steel in a single step. It also eliminates the need for coking and sintering plants, along with other dangerous and expensive components of traditional systems. As a result, the company says its process uses 30 percent less energy and costs less to operate than conventional steel mills in America.

“The real headline is the fact that we can make steel from iron ore more cost-competitive by 25 percent in the United States, while also reducing emissions.” Meroueh says. “The United States hasn’t been competitive in steelmaking in decades. Now we’re enabling that.”

Since late 2024, Hertha has been operating a 1-tonne-per-day pilot plant at its first production facility outside Houston, Texas. The company calls it the world’s largest demonstration of a single-step steelmaking process. This year, the company will begin construction of a plant that will be able to produce 10,000 tons of steel each year. That plant, which Hertha expects to reach full production capacity at the end 2027, will also produce high-purity iron for the magnet industry, helping America onshore another critical material.

“By importing so much of our pig iron and steel, we are completely reliant on global trade mechanisms and geopolitics remaining the way they are today for us to continue making the materials that are critical for our infrastructure, our defense systems, and our energy systems,” Meroueh says. “Steel is the most foundational material to our society. It is simply irreplaceable.”

Streamlining steelmaking

Meroueh earned her master’s degree in the lab of Gang Chen, MIT’s Carl Richard Soderberg Professor of Power Engineering. She studied thermal energy storage and the fundamental physics of heat transfer, eventually getting her first taste of entrepreneurship when she explored commercializing some of that research. Meroueh received a grant from the MIT Sandbox Innovation Fund and considers Executive Director Jinane Abounadi a close mentor today.

The experience taught Meroueh a lot about startups, but she ultimately decided to stay at MIT to pursue her PhD in metallurgy and hydrogen production in the lab of Douglas Hart, MIT professor of mechanical engineering. After earning her PhD in 2020, she was recruited to lead a hydrogen production startup for a year and a half.

“After that experience, I was looking at all of the hard-to-abate, high-emissions sectors of the economy to find the one receiving the least attention,” Meroueh says. “I stumbled onto steel and fell in love.”

Meroueh became an Innovators Fellow at the climate and energy startup investment firm Breakthrough Energy and officially founded Hertha Metals in 2022.

The company is named after Hertha Ayrton, a 19th-century physicist and inventor who advanced our understanding of electric arcs, which the company uses in its furnaces.

Globally, most steel today is made by combining iron ore with coke (from coal) and limestone in a blast furnace to make molten iron. That “pig iron” is then sent to another furnace to burn off excess carbon and impurities. Alloying elements are then added, and the steel is sent for casting and finishing, requiring additional machinery.

The U.S. makes most of its steel from recycled scrap metal, but it still must import iron made from a blast furnace to reach useful grades of steel.

“The United States has a massive need to make steel from iron ore, not just scrap, so we can stop relying on importing so much,” Meroueh explains. “We only have about 11 operational blast furnaces in the U.S., so we end up importing about 90 percent of the pig iron needed to feed into domestic scrap steel furnaces.”

To solve the problem, Meroueh leveraged a fuel America has in abundance: natural gas. Hertha’s system uses natural gas (the process also works with hydrogen) to reduce iron ore while using electricity to melt it in a single step. She says the closest competing technology requires scarce and expensive pelletized, high-grade iron ore and multiple furnaces to produce liquid steel. Meroueh’s process uses iron ore of any format or grade, producing refined liquid steel in a single furnace, cutting both cost and emissions.

“Many reactions that were previously run sequentially though a conventional steelmaking process are now occurring simultaneously, within a single furnace,” Meroueh explains. “We’re melting, we’re reducing, and we’re carburizing the steel to the exact amount we need. What exits our furnace is a refined molten steel. We can process any grade and format of iron ore because everything is occurring in the molten phase. It doesn’t matter whether the ore came in as a pellet or clumps and fines out of the ground.”

Meroueh says the company’s biggest innovation is performing the gaseous reduction when the iron oxide is a molten liquid using proprietary gas technologies.     

“All of the conventional steelmaking technologies perform reduction while the iron ore is in a solid state, and they use gas — whether that’s combusted coke or natural gas — to perform that reduction,” Meroueh says. “We saw the inefficiency in doing that and how it restricted the grade and form of usable iron ore, because at the end of the day you have to melt the ore anyway.”

Hertha’s system is modular and uses standard off-gas handling equipment, steam turbines, and heat exchangers. It also recycles natural gas to regenerate electricity from the hot off-gas leaving the furnace.

“Our steel mill has its own little power plant attached that leads to 35 percent recovery in energy and minimizes grid power demand in an age in which we are competing with data centers,” Meroueh says.

Onshoring critical materials

Today’s steel mills are the result of enormous investments and are designed to run for at least 50 years. Hertha Metals doesn’t envision replacing those entirely — at least not anytime soon.

“You’re not just going to shut off a steel mill in the middle of its life,” Meroueh says. “Sure, you can build new steel mills, but we really want to be able to displace the blast furnace and the basic oxygen furnace while still utilizing all the mill’s downstream equipment.”

The company’s Houston plant began producing one ton of steel per day just two years after Hertha’s founding and less than one year after Meroueh opened up Hertha’s headquarters. She calls it an important first step.

“This is the largest-scale demonstration of a single-step steelmaking company,” Meroueh says. “It’s a true breakthrough in terms of scalability, pace of progress, and capital efficiency.”

The company’s next plant, which will be capable of producing 10,000 tons of steel each year, will also be producing high-purity iron for permanent magnets, which are used in electric motors, robotics, consumer electronics, aerospace and military hardware.

“It’s insane that we don’t make rare earth magnets domestically,” Meroueh says. “It’s insane that any country doesn’t make their own rare earth magnets. Most rare earth magnets are permanent magnets, so neodymium magnets. What’s interesting is that by weight, 70 percent of that magnet is not a rare earth, it’s high-purity iron. America doesn’t currently make any high-purity iron, but Hertha has already made it in our pilot plant.”

Hertha plans to quickly scale up its production of high-purity iron so that, by 2030, it will be able to meet about a quarter of total projected demand for magnets in the U.S.

After that, the company plans to run a full-scale commercial steel plant in partnership with a steel manufacturer in America. Meroueh says that plant, which will be able to produce around half a million tons of steel each year, should be operational by 2030.

“We are eager to partner with today’s steel producers so that we can collectively leverage the existing infrastructure alongside Hertha’s innovation,” Meroueh says. “That includes the $1.5 billion of capital downstream of a melt shop that Hertha’s process can integrate into. The melt shop is the ore-to-liquid steel portion of the steel mill. That’s just the start.  It’s a smaller scale than a conventional plant in which we still economically out compete traditional production processes. Then we’re going to scale to 2 million tons per year once we build up our balance sheet.”

The Abdul Latif Jameel Poverty Action Lab (J-PAL) at MIT has awarded funding to eight new research studies to understand how artificial intelligence innovations can be used in the fight against poverty through its new Project AI Evidence.

The age of AI has brought wide-ranging optimism and skepticism about its effects on society. To realize AI’s full potential, Project AI Evidence (PAIE) will identify which AI solutions work and for whom, and scale only the most effective, inclusive, and responsible solutions — while scaling down those that may potentially cause harm.

PAIE will generate evidence on what works by connecting governments, tech companies, and nonprofits with world-class economists at MIT and across J-PAL’s global network to evaluate and improve AI solutions to entrenched social challenges.

The new initiative is prioritizing questions policymakers are already asking: Do AI-assisted teaching tools help all children learn? How can early-warning flood systems help people affected by natural disasters? Can machine learning algorithms help reduce deforestation in the Amazon? Can AI-powered chatbots help improve people’s health? In the coming years, PAIE will run a series of funding competitions to invite proposals for evaluations of AI tools that address questions like these, and many more.

PAIE is financially supported by a grant from Google.org, philanthropic support from Community Jameel, a grant from Canada’s International Development Research Centre and UK International Development, and a collaboration agreement with Amazon Web Services. Through a grant from Eric and Wendy Schmidt, awarded by recommendation of Schmidt Sciences, the initiative will also study generative AI in the workplace, particularly in low- and middle-income countries.

Alex Diaz, head of AI for social good at Google.org, says, “we’re thrilled to collaborate with MIT and J-PAL, already leaders in this space, on Project AI Evidence. AI has great potential to benefit all people, but we urgently need to study what works, what doesn’t, and why, if we are to realize this potential.”

“Artificial intelligence holds extraordinary potential, but only if the tools, knowledge, and power to shape it are accessible to all — that includes contextually grounded research and evidence on what works and what does not,” adds Maggie Gorman-Velez, vice president of strategy, regions, and policies at IDRC. “That is why IDRC is proud to be supporting this new evaluation work as part of our ongoing commitment to the responsible scaling of proven safe, inclusive, and locally relevant AI innovations.”

J-PAL is uniquely positioned to help understand AI’s effects on society: Since its inception in 2003, J-PAL’s network of researchers has led over 2,500 rigorous evaluations of social policies and programs around the world. Through PAIE, J-PAL will bring together leading experts in AI technology, research, and social policy, in alignment with MIT president Sally Kornbluth’s focus on generative AI as a strategic priority.

PAIE is chaired by Professor Joshua Blumenstock of the University of California at Berkeley; J-PAL Global Executive Director Iqbal Dhaliwal; and Professor David Yanagizawa-Drott of the University of Zurich.

New evaluations of urgent policy questions

The studies funded in PAIE’s first round of competition explore urgent questions in key sectors like education, health, climate, and economic opportunity.

How can AI be most effective in classrooms, helping both students and teachers?

Existing research shows that personalized learning is important for students, but challenging to implement with limited resources. In Kenya, education social enterprise EIDU has developed an AI tool that helps teachers identify learning gaps and adapt their daily lesson plans. In India, the nongovernmental organization (NGO) Pratham is developing an AI tool to increase the impact and scale of the evidence-informed Teaching at the Right Level approach. J-PAL researchers Daron Acemoglu, Iqbal Dhaliwal, and Francisco Gallego will work with both organizations to study the effects and potential of these different use cases on teachers’ productivity and students’ learning.

Can AI tools reduce gender bias in schools?

Researchers are collaborating with Italy’s Ministry of Education to evaluate whether AI tools can help close gender gaps in students’ performance by addressing teachers’ unconscious biases. J-PAL affiliates Michela Carlana and Will Dobbie, along with Francesca Miserocchi and Eleonora Patacchini, will study the impacts of two AI tools, one that helps teachers predict performance and a second that gives real-time feedback on the diversity of their decisions.

Can AI help career counselors uncover more job opportunities?

In Kenya, researchers are evaluating if an AI tool can identify overlooked skills and unlock employment opportunities, particularly for youth, women, and those without formal education. In collaboration with NGOs Swahilipot and Tabiya, Jasmin Baier and J-PAL researcher Christian Meyer will evaluate how the tool changes people’s job search strategies and employment. This study will shed light on AI as a complement, rather than a substitute, for human expertise in career guidance.

Looking forward

As use of AI in the social sector evolves, these evaluations are a first step in discovering effective, responsible solutions that will go the furthest in alleviating poverty and inequality.

J-PAL’s Dhaliwal notes, “J-PAL has a long history of evaluating innovative technology and its ability to improve people’s lives. While AI has incredible potential, we need to maximize its benefits and minimize possible harms. We’re grateful to our donors, sponsors, and collaborators for their catalytic support in launching PAIE, which will help us do exactly that by continuing to expand evidence on the impacts of AI innovations.”

J-PAL is also seeking new collaborators who share its vision of discovering and scaling up real-world AI solutions. It aims to support more governments and social sector organizations that want to adopt AI responsibly, and will continue to expand funding for new evaluations and provide policy guidance based on the latest research.

To learn more about Project AI Evidence, subscribe to J-PAL's newsletter or contact paie@povertyactionlab.org.