Kategorie: KI Use Cases

Using a type of artificial intelligence known as deep learning, MIT researchers have discovered a class of compounds that can kill a drug-resistant bacterium that causes more than 10,000 deaths in the United States every year.

In a study appearing today in Nature, the researchers showed that these compounds could kill methicillin-resistant Staphylococcus aureus (MRSA) grown in a lab dish and in two mouse models of MRSA infection. The compounds also show very low toxicity against human cells, making them particularly good drug candidates.

A key innovation of the new study is that the researchers were also able to figure out what kinds of information the deep-learning model was using to make its antibiotic potency predictions. This knowledge could help researchers to design additional drugs that might work even better than the ones identified by the model.

“The insight here was that we could see what was being learned by the models to make their predictions that certain molecules would make for good antibiotics. Our work provides a framework that is time-efficient, resource-efficient, and mechanistically insightful, from a chemical-structure standpoint, in ways that we haven’t had to date,” says James Collins, the Termeer Professor of Medical Engineering and Science in MIT’s Institute for Medical Engineering and Science (IMES) and Department of Biological Engineering.

Felix Wong, a postdoc at IMES and the Broad Institute of MIT and Harvard, and Erica Zheng, a former Harvard Medical School graduate student who was advised by Collins, are the lead authors of the study, which is part of the Antibiotics-AI Project at MIT. The mission of this project, led by Collins, is to discover new classes of antibiotics against seven types of deadly bacteria, over seven years.

Explainable predictions

MRSA, which infects more than 80,000 people in the United States every year, often causes skin infections or pneumonia. Severe cases can lead to sepsis, a potentially fatal bloodstream infection.

Over the past several years, Collins and his colleagues in MIT’s Abdul Latif Jameel Clinic for Machine Learning in Health (Jameel Clinic) have begun using deep learning to try to find new antibiotics. Their work has yielded potential drugs against Acinetobacter baumannii, a bacterium that is often found in hospitals, and many other drug-resistant bacteria.

These compounds were identified using deep learning models that can learn to identify chemical structures that are associated with antimicrobial activity. These models then sift through millions of other compounds, generating predictions of which ones may have strong antimicrobial activity.

These types of searches have proven fruitful, but one limitation to this approach is that the models are “black boxes,” meaning that there is no way of knowing what features the model based its predictions on. If scientists knew how the models were making their predictions, it could be easier for them to identify or design additional antibiotics.

“What we set out to do in this study was to open the black box,” Wong says. “These models consist of very large numbers of calculations that mimic neural connections, and no one really knows what’s going on underneath the hood.”

First, the researchers trained a deep learning model using substantially expanded datasets. They generated this training data by testing about 39,000 compounds for antibiotic activity against MRSA, and then fed this data, plus information on the chemical structures of the compounds, into the model.

“You can represent basically any molecule as a chemical structure, and also you tell the model if that chemical structure is antibacterial or not,” Wong says. “The model is trained on many examples like this. If you then give it any new molecule, a new arrangement of atoms and bonds, it can tell you a probability that that compound is predicted to be antibacterial.”

To figure out how the model was making its predictions, the researchers adapted an algorithm known as Monte Carlo tree search, which has been used to help make other deep learning models, such as AlphaGo, more explainable. This search algorithm allows the model to generate not only an estimate of each molecule’s antimicrobial activity, but also a prediction for which substructures of the molecule likely account for that activity.

Potent activity

To further narrow down the pool of candidate drugs, the researchers trained three additional deep learning models to predict whether the compounds were toxic to three different types of human cells. By combining this information with the predictions of antimicrobial activity, the researchers discovered compounds that could kill microbes while having minimal adverse effects on the human body.

Using this collection of models, the researchers screened about 12 million compounds, all of which are commercially available. From this collection, the models identified compounds from five different classes, based on chemical substructures within the molecules, that were predicted to be active against MRSA.

The researchers purchased about 280 compounds and tested them against MRSA grown in a lab dish, allowing them to identify two, from the same class, that appeared to be very promising antibiotic candidates. In tests in two mouse models, one of MRSA skin infection and one of MRSA systemic infection, each of those compounds reduced the MRSA population by a factor of 10.

Experiments revealed that the compounds appear to kill bacteria by disrupting their ability to maintain an electrochemical gradient across their cell membranes. This gradient is needed for many critical cell functions, including the ability to produce ATP (molecules that cells use to store energy). An antibiotic candidate that Collins’ lab discovered in 2020, halicin, appears to work by a similar mechanism but is specific to Gram-negative bacteria (bacteria with thin cell walls). MRSA is a Gram-positive bacterium, with thicker cell walls.

“We have pretty strong evidence that this new structural class is active against Gram-positive pathogens by selectively dissipating the proton motive force in bacteria,” Wong says. “The molecules are attacking bacterial cell membranes selectively, in a way that does not incur substantial damage in human cell membranes. Our substantially augmented deep learning approach allowed us to predict this new structural class of antibiotics and enabled the finding that it is not toxic against human cells.”

The researchers have shared their findings with Phare Bio, a nonprofit started by Collins and others as part of the Antibiotics-AI Project. The nonprofit now plans to do more detailed analysis of the chemical properties and potential clinical use of these compounds. Meanwhile, Collins’ lab is working on designing additional drug candidates based on the findings of the new study, as well as using the models to seek compounds that can kill other types of bacteria.

“We are already leveraging similar approaches based on chemical substructures to design compounds de novo, and of course, we can readily adopt this approach out of the box to discover new classes of antibiotics against different pathogens,” Wong says.

In addition to MIT, Harvard, and the Broad Institute, the paper’s contributing institutions are Integrated Biosciences, Inc., the Wyss Institute for Biologically Inspired Engineering, and the Leibniz Institute of Polymer Research in Dresden, Germany. The research was funded by the James S. McDonnell Foundation, the U.S. National Institute of Allergy and Infectious Diseases, the Swiss National Science Foundation, the Banting Fellowships Program, the Volkswagen Foundation, the Defense Threat Reduction Agency, the U.S. National Institutes of Health, and the Broad Institute. The Antibiotics-AI Project is funded by the Audacious Project, Flu Lab, the Sea Grape Foundation, the Wyss Foundation, and an anonymous donor.

Imagine you are scrolling through the photos on your phone and you come across an image that at first you can’t recognize. It looks like maybe something fuzzy on the couch; could it be a pillow or a coat? After a couple of seconds it clicks — of course! That ball of fluff is your friend’s cat, Mocha. While some of your photos could be understood in an instant, why was this cat photo much more difficult?

MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) researchers were surprised to find that despite the critical importance of understanding visual data in pivotal areas ranging from health care to transportation to household devices, the notion of an image’s recognition difficulty for humans has been almost entirely ignored. One of the major drivers of progress in deep learning-based AI has been datasets, yet we know little about how data drives progress in large-scale deep learning beyond that bigger is better.

In real-world applications that require understanding visual data, humans outperform object recognition models despite the fact that models perform well on current datasets, including those explicitly designed to challenge machines with debiased images or distribution shifts. This problem persists, in part, because we have no guidance on the absolute difficulty of an image or dataset. Without controlling for the difficulty of images used for evaluation, it’s hard to objectively assess progress toward human-level performance, to cover the range of human abilities, and to increase the challenge posed by a dataset.

To fill in this knowledge gap, David Mayo, an MIT PhD student in electrical engineering and computer science and a CSAIL affiliate, delved into the deep world of image datasets, exploring why certain images are more difficult for humans and machines to recognize than others. „Some images inherently take longer to recognize, and it’s essential to understand the brain’s activity during this process and its relation to machine learning models. Perhaps there are complex neural circuits or unique mechanisms missing in our current models, visible only when tested with challenging visual stimuli. This exploration is crucial for comprehending and enhancing machine vision models,” says Mayo, a lead author of a new paper on the work.

This led to the development of a new metric, the “minimum viewing time” (MVT), which quantifies the difficulty of recognizing an image based on how long a person needs to view it before making a correct identification. Using a subset of ImageNet, a popular dataset in machine learning, and ObjectNet, a dataset designed to test object recognition robustness, the team showed images to participants for varying durations from as short as 17 milliseconds to as long as 10 seconds, and asked them to choose the correct object from a set of 50 options. After over 200,000 image presentation trials, the team found that existing test sets, including ObjectNet, appeared skewed toward easier, shorter MVT images, with the vast majority of benchmark performance derived from images that are easy for humans.

The project identified interesting trends in model performance — particularly in relation to scaling. Larger models showed considerable improvement on simpler images but made less progress on more challenging images. The CLIP models, which incorporate both language and vision, stood out as they moved in the direction of more human-like recognition.

“Traditionally, object recognition datasets have been skewed towards less-complex images, a practice that has led to an inflation in model performance metrics, not truly reflective of a model’s robustness or its ability to tackle complex visual tasks. Our research reveals that harder images pose a more acute challenge, causing a distribution shift that is often not accounted for in standard evaluations,” says Mayo. “We released image sets tagged by difficulty along with tools to automatically compute MVT, enabling MVT to be added to existing benchmarks and extended to various applications. These include measuring test set difficulty before deploying real-world systems, discovering neural correlates of image difficulty, and advancing object recognition techniques to close the gap between benchmark and real-world performance.”

“One of my biggest takeaways is that we now have another dimension to evaluate models on. We want models that are able to recognize any image even if — perhaps especially if — it’s hard for a human to recognize. We’re the first to quantify what this would mean. Our results show that not only is this not the case with today’s state of the art, but also that our current evaluation methods don’t have the ability to tell us when it is the case because standard datasets are so skewed toward easy images,” says Jesse Cummings, an MIT graduate student in electrical engineering and computer science and co-first author with Mayo on the paper.

From ObjectNet to MVT

A few years ago, the team behind this project identified a significant challenge in the field of machine learning: Models were struggling with out-of-distribution images, or images that were not well-represented in the training data. Enter ObjectNet, a dataset comprised of images collected from real-life settings. The dataset helped illuminate the performance gap between machine learning models and human recognition abilities, by eliminating spurious correlations present in other benchmarks — for example, between an object and its background. ObjectNet illuminated the gap between the performance of machine vision models on datasets and in real-world applications, encouraging use for many researchers and developers — which subsequently improved model performance.

Fast forward to the present, and the team has taken their research a step further with MVT. Unlike traditional methods that focus on absolute performance, this new approach assesses how models perform by contrasting their responses to the easiest and hardest images. The study further explored how image difficulty could be explained and tested for similarity to human visual processing. Using metrics like c-score, prediction depth, and adversarial robustness, the team found that harder images are processed differently by networks. “While there are observable trends, such as easier images being more prototypical, a comprehensive semantic explanation of image difficulty continues to elude the scientific community,” says Mayo.

In the realm of health care, for example, the pertinence of understanding visual complexity becomes even more pronounced. The ability of AI models to interpret medical images, such as X-rays, is subject to the diversity and difficulty distribution of the images. The researchers advocate for a meticulous analysis of difficulty distribution tailored for professionals, ensuring AI systems are evaluated based on expert standards, rather than layperson interpretations.

Mayo and Cummings are currently looking at neurological underpinnings of visual recognition as well, probing into whether the brain exhibits differential activity when processing easy versus challenging images. The study aims to unravel whether complex images recruit additional brain areas not typically associated with visual processing, hopefully helping demystify how our brains accurately and efficiently decode the visual world.

Toward human-level performance

Looking ahead, the researchers are not only focused on exploring ways to enhance AI’s predictive capabilities regarding image difficulty. The team is working on identifying correlations with viewing-time difficulty in order to generate harder or easier versions of images.

Despite the study’s significant strides, the researchers acknowledge limitations, particularly in terms of the separation of object recognition from visual search tasks. The current methodology does concentrate on recognizing objects, leaving out the complexities introduced by cluttered images.

“This comprehensive approach addresses the long-standing challenge of objectively assessing progress towards human-level performance in object recognition and opens new avenues for understanding and advancing the field,” says Mayo. “With the potential to adapt the Minimum Viewing Time difficulty metric for a variety of visual tasks, this work paves the way for more robust, human-like performance in object recognition, ensuring that models are truly put to the test and are ready for the complexities of real-world visual understanding.”

“This is a fascinating study of how human perception can be used to identify weaknesses in the ways AI vision models are typically benchmarked, which overestimate AI performance by concentrating on easy images,” says Alan L. Yuille, Bloomberg Distinguished Professor of Cognitive Science and Computer Science at Johns Hopkins University, who was not involved in the paper. “This will help develop more realistic benchmarks leading not only to improvements to AI but also make fairer comparisons between AI and human perception.” 

“It’s widely claimed that computer vision systems now outperform humans, and on some benchmark datasets, that’s true,” says Anthropic technical staff member Simon Kornblith PhD ’17, who was also not involved in this work. “However, a lot of the difficulty in those benchmarks comes from the obscurity of what’s in the images; the average person just doesn’t know enough to classify different breeds of dogs. This work instead focuses on images that people can only get right if given enough time. These images are generally much harder for computer vision systems, but the best systems are only a bit worse than humans.”

Mayo, Cummings, and Xinyu Lin MEng ’22 wrote the paper alongside CSAIL Research Scientist Andrei Barbu, CSAIL Principal Research Scientist Boris Katz, and MIT-IBM Watson AI Lab Principal Researcher Dan Gutfreund. The researchers are affiliates of the MIT Center for Brains, Minds, and Machines.

The team is presenting their work at the 2023 Conference on Neural Information Processing Systems (NeurIPS).

Cathy Wu has always delighted in systems that run smoothly. In high school, she designed a project to optimize the best route for getting to class on time. Her research interests and career track are evidence of a propensity for organizing and optimizing, coupled with a strong sense of responsibility to contribute to society instilled by her parents at a young age.

As an undergraduate at MIT, Wu explored domains like agriculture, energy, and education, eventually homing in on transportation. “Transportation touches each of our lives,” she says. “Every day, we experience the inefficiencies and safety issues as well as the environmental harms associated with our transportation systems. I believe we can and should do better.”

But doing so is complicated. Consider the long-standing issue of traffic systems control. Wu explains that it is not one problem, but more accurately a family of control problems impacted by variables like time of day, weather, and vehicle type — not to mention the types of sensing and communication technologies used to measure roadway information. Every differentiating factor introduces an exponentially larger set of control problems. There are thousands of control-problem variations and hundreds, if not thousands, of studies and papers dedicated to each problem. Wu refers to the sheer number of variations as the curse of variety — and it is hindering innovation.

“To prove that a new control strategy can be safely deployed on our streets can take years. As time lags, we lose opportunities to improve safety and equity while mitigating environmental impacts. Accelerating this process has huge potential,” says Wu.  

Which is why she and her group in the MIT Laboratory for Information and Decision Systems are devising machine learning-based methods to solve not just a single control problem or a single optimization problem, but families of control and optimization problems at scale. “In our case, we’re examining emerging transportation problems that people have spent decades trying to solve with classical approaches. It seems to me that we need a different approach.”

Optimizing intersections

Currently, Wu’s largest research endeavor is called Project Greenwave. There are many sectors that directly contribute to climate change, but transportation is responsible for the largest share of greenhouse gas emissions — 29 percent, of which 81 percent is due to land transportation. And while much of the conversation around mitigating environmental impacts related to mobility is focused on electric vehicles (EVs), electrification has its drawbacks. EV fleet turnover is time-consuming (“on the order of decades,” says Wu), and limited global access to the technology presents a significant barrier to widespread adoption.

Wu’s research, on the other hand, addresses traffic control problems by leveraging deep reinforcement learning. Specifically, she is looking at traffic intersections — and for good reason. In the United States alone, there are more than 300,000 signalized intersections where vehicles must stop or slow down before re-accelerating. And every re-acceleration burns fossil fuels and contributes to greenhouse gas emissions.

Highlighting the magnitude of the issue, Wu says, “We have done preliminary analysis indicating that up to 15 percent of land transportation CO2 is wasted through energy spent idling and re-accelerating at intersections.”

To date, she and her group have modeled 30,000 different intersections across 10 major metropolitan areas in the United States. That is 30,000 different configurations, roadway topologies (e.g., grade of road or elevation), different weather conditions, and variations in travel demand and fuel mix. Each intersection and its corresponding scenarios represents a unique multi-agent control problem.

Wu and her team are devising techniques that can solve not just one, but a whole family of problems comprised of tens of thousands of scenarios. Put simply, the idea is to coordinate the timing of vehicles so they arrive at intersections when traffic lights are green, thereby eliminating the start, stop, re-accelerate conundrum. Along the way, they are building an ecosystem of tools, datasets, and methods to enable roadway interventions and impact assessments of strategies to significantly reduce carbon-intense urban driving.

Their collaborator on the project is the Utah Department of Transportation, which Wu says has played an essential role, in part by sharing data and practical knowledge that she and her group otherwise would not have been able to access publicly.

“I appreciate industry and public sector collaborations,” says Wu. “When it comes to important societal problems, one really needs grounding with practitioners. One needs to be able to hear the perspectives in the field. My interactions with practitioners expand my horizons and help ground my research. You never know when you’ll hear the perspective that is the key to the solution, or perhaps the key to understanding the problem.”

Finding the best routes

In a similar vein, she and her research group are tackling large coordination problems. For example, vehicle routing. “Every day, delivery trucks route more than a hundred thousand packages for the city of Boston alone,” says Wu. Accomplishing the task requires, among other things, figuring out which trucks to use, which packages to deliver, and the order in which to deliver them as efficiently as possible. If and when the trucks are electrified, they will need to be charged, adding another wrinkle to the process and further complicating route optimization.

The vehicle routing problem, and therefore the scope of Wu’s work, extends beyond truck routing for package delivery. Ride-hailing cars may need to pick up objects as well as drop them off; and what if delivery is done by bicycle or drone? In partnership with Amazon, for example, Wu and her team addressed routing and path planning for hundreds of robots (up to 800) in their warehouses.

Every variation requires custom heuristics that are expensive and time-consuming to develop. Again, this is really a family of problems — each one complicated, time-consuming, and currently unsolved by classical techniques — and they are all variations of a central routing problem. The curse of variety meets operations and logistics.

By combining classical approaches with modern deep-learning methods, Wu is looking for a way to automatically identify heuristics that can effectively solve all of these vehicle routing problems. So far, her approach has proved successful.

“We’ve contributed hybrid learning approaches that take existing solution methods for small problems and incorporate them into our learning framework to scale and accelerate that existing solver for large problems. And we’re able to do this in a way that can automatically identify heuristics for specialized variations of the vehicle routing problem.” The next step, says Wu, is applying a similar approach to multi-agent robotics problems in automated warehouses.

Wu and her group are making big strides, in part due to their dedication to use-inspired basic research. Rather than applying known methods or science to a problem, they develop new methods, new science, to address problems. The methods she and her team employ are necessitated by societal problems with practical implications. The inspiration for the approach? None other than Louis Pasteur, who described his research style in a now-famous article titled “Pasteur’s Quadrant.” Anthrax was decimating the sheep population, and Pasteur wanted to better understand why and what could be done about it. The tools of the time could not solve the problem, so he invented a new field, microbiology, not out of curiosity but out of necessity.

For a small percentage of cancer patients, doctors are unable to determine where their cancer originated. This makes it much more difficult to choose a treatment for those patients, because many cancer drugs are typically developed for specific cancer types.

A new approach developed by researchers at MIT and Dana-Farber Cancer Institute may make it easier to identify the sites of origin for those enigmatic cancers. Using machine learning, the researchers created a computational model that can analyze the sequence of about 400 genes and use that information to predict where a given tumor originated in the body.

Using this model, the researchers showed that they could accurately classify at least 40 percent of tumors of unknown origin with high confidence, in a dataset of about 900 patients. This approach enabled a 2.2-fold increase in the number of patients who could have been eligible for a genomically guided, targeted treatment, based on where their cancer originated.

“That was the most important finding in our paper, that this model could be potentially used to aid treatment decisions, guiding doctors toward personalized treatments for patients with cancers of unknown primary origin,” says Intae Moon, an MIT graduate student in electrical engineering and computer science who is the lead author of the new study.

Alexander Gusev, an associate professor of medicine at Harvard Medical School and Dana-Farber Cancer Institute, is the senior author of the paper, which appears today in Nature Medicine.

Mysterious origins

In 3 to 5 percent of cancer patients, particularly in cases where tumors have metastasized throughout the body, oncologists don’t have an easy way to determine where the cancer originated. These tumors are classified as cancers of unknown primary (CUP).

This lack of knowledge often prevents doctors from being able to give patients “precision” drugs, which are typically approved for specific cancer types where they are known to work. These targeted treatments tend to be more effective and have fewer side effects than treatments that are used for a broad spectrum of cancers, which are commonly prescribed to CUP patients.

“A sizeable number of individuals develop these cancers of unknown primary every year, and because most therapies are approved in a site-specific way, where you have to know the primary site to deploy them, they have very limited treatment options,” Gusev says.

Moon, an affiliate of the Computer Science and Artificial Intelligence Laboratory who is co-advised by Gusev, decided to analyze genetic data that is routinely collected at Dana-Farber to see if it could be used to predict cancer type. The data consist of genetic sequences for about 400 genes that are often mutated in cancer. The researchers trained a machine-learning model on data from nearly 30,000 patients who had been diagnosed with one of 22 known cancer types. That set of data included patients from Memorial Sloan Kettering Cancer Center and Vanderbilt-Ingram Cancer Center, as well as Dana-Farber.

The researchers then tested the resulting model on about 7,000 tumors that it hadn’t seen before, but whose site of origin was known. The model, which the researchers named OncoNPC, was able to predict their origins with about 80 percent accuracy. For tumors with high-confidence predictions, which constituted about 65 percent of the total, its accuracy rose to roughly 95 percent.

After those encouraging results, the researchers used the model to analyze a set of about 900 tumors from patients with CUP, which were all from Dana-Farber. They found that for 40 percent of these tumors, the model was able to make high-confidence predictions.

The researchers then compared the model’s predictions with an analysis of the germline, or inherited, mutations in a subset of tumors with available data, which can reveal whether the patients have a genetic predisposition to develop a particular type of cancer. The researchers found that the model’s predictions were much more likely to match the type of cancer most strongly predicted by the germline mutations than any other type of cancer.

Guiding drug decisions

To further validate the model’s predictions, the researchers compared data on the CUP patients’ survival time with the typical prognosis for the type of cancer that the model predicted. They found that CUP patients who were predicted to have cancer with a poor prognosis, such as pancreatic cancer, showed correspondingly shorter survival times. Meanwhile, CUP patients who were predicted to have cancers that typically have better prognoses, such as neuroendocrine tumors, had longer survival times.

Another indication that the model’s predictions could be useful came from looking at the types of treatments that CUP patients analyzed in the study had received. About 10 percent of these patients had received a targeted treatment, based on their oncologists’ best guess about where their cancer had originated. Among those patients, those who received a treatment consistent with the type of cancer that the model predicted for them fared better than patients who received a treatment typically given for a different type of cancer than what the model predicted for them.

Using this model, the researchers also identified an additional 15 percent of patients (2.2-fold increase) who could have received an existing targeted treatment, if their cancer type had been known. Instead, those patients ended up receiving more general chemotherapy drugs.

“That potentially makes these findings more clinically actionable because we’re not requiring a new drug to be approved. What we’re saying is that this population can now be eligible for precision treatments that already exist,” Gusev says.

The researchers now hope to expand their model to include other types of data, such as pathology images and radiology images, to provide a more comprehensive prediction using multiple data modalities. This would also provide the model with a comprehensive perspective of tumors, enabling it to predict not just the type of tumor and patient outcome, but potentially even the optimal treatment.

The research was funded by the National Institutes of Health, the Louis B. Mayer Foundation, the Doris Duke Charitable Foundation, the Phi Beta Psi Sorority, and the Emerson Collective.

As we enter a new era where technologies powered by artificial intelligence can craft and manipulate images with a precision that blurs the line between reality and fabrication, the specter of misuse looms large. Recently, advanced generative models such as DALL-E and Midjourney, celebrated for their impressive precision and user-friendly interfaces, have made the production of hyper-realistic images relatively effortless. With the barriers of entry lowered, even inexperienced users can generate and manipulate high-quality images from simple text descriptions — ranging from innocent image alterations to malicious changes. Techniques like watermarking pose a promising solution, but misuse requires a preemptive (as opposed to only post hoc) measure. 

In the quest to create such a new measure, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) developed “PhotoGuard,” a technique that uses perturbations — minuscule alterations in pixel values invisible to the human eye but detectable by computer models — that effectively disrupt the model’s ability to manipulate the image.

PhotoGuard uses two different “attack” methods to generate these perturbations. The more straightforward “encoder” attack targets the image’s latent representation in the AI model, causing the model to perceive the image as a random entity. The more sophisticated “diffusion” one defines a target image and optimizes the perturbations to make the final image resemble the target as closely as possible.

“Consider the possibility of fraudulent propagation of fake catastrophic events, like an explosion at a significant landmark. This deception can manipulate market trends and public sentiment, but the risks are not limited to the public sphere. Personal images can be inappropriately altered and used for blackmail, resulting in significant financial implications when executed on a large scale,” says Hadi Salman, an MIT graduate student in electrical engineering and computer science (EECS), affiliate of MIT CSAIL, and lead author of a new paper about PhotoGuard. 

“In more extreme scenarios, these models could simulate voices and images for staging false crimes, inflicting psychological distress and financial loss. The swift nature of these actions compounds the problem. Even when the deception is eventually uncovered, the damage — whether reputational, emotional, or financial — has often already happened. This is a reality for victims at all levels, from individuals bullied at school to society-wide manipulation.”

PhotoGuard in practice

AI models view an image differently from how humans do. It sees an image as a complex set of mathematical data points that describe every pixel’s color and position — this is the image’s latent representation. The encoder attack introduces minor adjustments into this mathematical representation, causing the AI model to perceive the image as a random entity. As a result, any attempt to manipulate the image using the model becomes nearly impossible. The changes introduced are so minute that they are invisible to the human eye, thus preserving the image’s visual integrity while ensuring its protection.

The second and decidedly more intricate “diffusion” attack strategically targets the entire diffusion model end-to-end. This involves determining a desired target image, and then initiating an optimization process with the intention of closely aligning the generated image with this preselected target.

In implementing, the team created perturbations within the input space of the original image. These perturbations are then used during the inference stage, and applied to the images, offering a robust defense against unauthorized manipulation.

“The progress in AI that we are witnessing is truly breathtaking, but it enables beneficial and malicious uses of AI alike,” says MIT professor of EECS and CSAIL principal investigator Aleksander Madry, who is also an author on the paper. “It is thus urgent that we work towards identifying and mitigating the latter. I view PhotoGuard as our small contribution to that important effort.”

The diffusion attack is more computationally intensive than its simpler sibling, and requires significant GPU memory. The team says that approximating the diffusion process with fewer steps mitigates the issue, thus making the technique more practical.

To better illustrate the attack, consider an art project, for example. The original image is a drawing, and the target image is another drawing that’s completely different. The diffusion attack is like making tiny, invisible changes to the first drawing so that, to an AI model, it begins to resemble the second drawing. However, to the human eye, the original drawing remains unchanged.

By doing this, any AI model attempting to modify the original image will now inadvertently make changes as if dealing with the target image, thereby protecting the original image from intended manipulation. The result is a picture that remains visually unaltered for human observers, but protects against unauthorized edits by AI models.

As far as a real example with PhotoGuard, consider an image with multiple faces. You could mask any faces you don’t want to modify, and then prompt with “two men attending a wedding.” Upon submission, the system will adjust the image accordingly, creating a plausible depiction of two men participating in a wedding ceremony.

Now, consider safeguarding the image from being edited; adding perturbations to the image before upload can immunize it against modifications. In this case, the final output will lack realism compared to the original, non-immunized image.

All hands on deck

Key allies in the fight against image manipulation are the creators of the image-editing models, says the team. For PhotoGuard to be effective, an integrated response from all stakeholders is necessary. “Policymakers should consider implementing regulations that mandate companies to protect user data from such manipulations. Developers of these AI models could design APIs that automatically add perturbations to users’ images, providing an added layer of protection against unauthorized edits,” says Salman.

Despite PhotoGuard’s promise, it’s not a panacea. Once an image is online, individuals with malicious intent could attempt to reverse engineer the protective measures by applying noise, cropping, or rotating the image. However, there is plenty of previous work from the adversarial examples literature that can be utilized here to implement robust perturbations that resist common image manipulations.

“A collaborative approach involving model developers, social media platforms, and policymakers presents a robust defense against unauthorized image manipulation. Working on this pressing issue is of paramount importance today,” says Salman. “And while I am glad to contribute towards this solution, much work is needed to make this protection practical. Companies that develop these models need to invest in engineering robust immunizations against the possible threats posed by these AI tools. As we tread into this new era of generative models, let’s strive for potential and protection in equal measures.”

“The prospect of using attacks on machine learning to protect us from abusive uses of this technology is very compelling,” says Florian Tramèr, an assistant professor at ETH Zürich. “The paper has a nice insight that the developers of generative AI models have strong incentives to provide such immunization protections to their users, which could even be a legal requirement in the future. However, designing image protections that effectively resist circumvention attempts is a challenging problem: Once the generative AI company commits to an immunization mechanism and people start applying it to their online images, we need to ensure that this protection will work against motivated adversaries who might even use better generative AI models developed in the near future. Designing such robust protections is a hard open problem, and this paper makes a compelling case that generative AI companies should be working on solving it.”

Salman wrote the paper alongside fellow lead authors Alaa Khaddaj and Guillaume Leclerc MS ’18, as well as Andrew Ilyas ’18, MEng ’18; all three are EECS graduate students and MIT CSAIL affiliates. The team’s work was partially done on the MIT Supercloud compute cluster, supported by U.S. National Science Foundation grants and Open Philanthropy, and based upon work supported by the U.S. Defense Advanced Research Projects Agency. It was presented at the International Conference on Machine Learning this July.

Researchers from MIT and Stanford University have devised a new machine-learning approach that could be used to control a robot, such as a drone or autonomous vehicle, more effectively and efficiently in dynamic environments where conditions can change rapidly.

This technique could help an autonomous vehicle learn to compensate for slippery road conditions to avoid going into a skid, allow a robotic free-flyer to tow different objects in space, or enable a drone to closely follow a downhill skier despite being buffeted by strong winds.

The researchers’ approach incorporates certain structure from control theory into the process for learning a model in such a way that leads to an effective method of controlling complex dynamics, such as those caused by impacts of wind on the trajectory of a flying vehicle. One way to think about this structure is as a hint that can help guide how to control a system.

“The focus of our work is to learn intrinsic structure in the dynamics of the system that can be leveraged to design more effective, stabilizing controllers,” says Navid Azizan, the Esther and Harold E. Edgerton Assistant Professor in the MIT Department of Mechanical Engineering and the Institute for Data, Systems, and Society (IDSS), and a member of the Laboratory for Information and Decision Systems (LIDS). “By jointly learning the system’s dynamics and these unique control-oriented structures from data, we’re able to naturally create controllers that function much more effectively in the real world.”

Using this structure in a learned model, the researchers’ technique immediately extracts an effective controller from the model, as opposed to other machine-learning methods that require a controller to be derived or learned separately with additional steps. With this structure, their approach is also able to learn an effective controller using fewer data than other approaches. This could help their learning-based control system achieve better performance faster in rapidly changing environments.

“This work tries to strike a balance between identifying structure in your system and just learning a model from data,” says lead author Spencer M. Richards, a graduate student at Stanford University. “Our approach is inspired by how roboticists use physics to derive simpler models for robots. Physical analysis of these models often yields a useful structure for the purposes of control — one that you might miss if you just tried to naively fit a model to data. Instead, we try to identify similarly useful structure from data that indicates how to implement your control logic.”

Additional authors of the paper are Jean-Jacques Slotine, professor of mechanical engineering and of brain and cognitive sciences at MIT, and Marco Pavone, associate professor of aeronautics and astronautics at Stanford. The research will be presented at the International Conference on Machine Learning (ICML).

Learning a controller

Determining the best way to control a robot to accomplish a given task can be a difficult problem, even when researchers know how to model everything about the system.

A controller is the logic that enables a drone to follow a desired trajectory, for example. This controller would tell the drone how to adjust its rotor forces to compensate for the effect of winds that can knock it off a stable path to reach its goal.

This drone is a dynamical system — a physical system that evolves over time. In this case, its position and velocity change as it flies through the environment. If such a system is simple enough, engineers can derive a controller by hand. 

Modeling a system by hand intrinsically captures a certain structure based on the physics of the system. For instance, if a robot were modeled manually using differential equations, these would capture the relationship between velocity, acceleration, and force. Acceleration is the rate of change in velocity over time, which is determined by the mass of and forces applied to the robot.

But often the system is too complex to be exactly modeled by hand. Aerodynamic effects, like the way swirling wind pushes a flying vehicle, are notoriously difficult to derive manually, Richards explains. Researchers would instead take measurements of the drone’s position, velocity, and rotor speeds over time, and use machine learning to fit a model of this dynamical system to the data. But these approaches typically don’t learn a control-based structure. This structure is useful in determining how to best set the rotor speeds to direct the motion of the drone over time.

Once they have modeled the dynamical system, many existing approaches also use data to learn a separate controller for the system.

“Other approaches that try to learn dynamics and a controller from data as separate entities are a bit detached philosophically from the way we normally do it for simpler systems. Our approach is more reminiscent of deriving models by hand from physics and linking that to control,” Richards says.

Identifying structure

The team from MIT and Stanford developed a technique that uses machine learning to learn the dynamics model, but in such a way that the model has some prescribed structure that is useful for controlling the system.

With this structure, they can extract a controller directly from the dynamics model, rather than using data to learn an entirely separate model for the controller.

“We found that beyond learning the dynamics, it’s also essential to learn the control-oriented structure that supports effective controller design. Our approach of learning state-dependent coefficient factorizations of the dynamics has outperformed the baselines in terms of data efficiency and tracking capability, proving to be successful in efficiently and effectively controlling the system’s trajectory,” Azizan says. 

When they tested this approach, their controller closely followed desired trajectories, outpacing all the baseline methods. The controller extracted from their learned model nearly matched the performance of a ground-truth controller, which is built using the exact dynamics of the system.

“By making simpler assumptions, we got something that actually worked better than other complicated baseline approaches,” Richards adds.

The researchers also found that their method was data-efficient, which means it achieved high performance even with few data. For instance, it could effectively model a highly dynamic rotor-driven vehicle using only 100 data points. Methods that used multiple learned components saw their performance drop much faster with smaller datasets.

This efficiency could make their technique especially useful in situations where a drone or robot needs to learn quickly in rapidly changing conditions.

Plus, their approach is general and could be applied to many types of dynamical systems, from robotic arms to free-flying spacecraft operating in low-gravity environments.

In the future, the researchers are interested in developing models that are more physically interpretable, and that would be able to identify very specific information about a dynamical system, Richards says. This could lead to better-performing controllers.

“Despite its ubiquity and importance, nonlinear feedback control remains an art, making it especially suitable for data-driven and learning-based methods. This paper makes a significant contribution to this area by proposing a method that jointly learns system dynamics, a controller, and control-oriented structure,” says Nikolai Matni, an assistant professor in the Department of Electrical and Systems Engineering at the University of Pennsylvania, who was not involved with this work. “What I found particularly exciting and compelling was the integration of these components into a joint learning algorithm, such that control-oriented structure acts as an inductive bias in the learning process. The result is a data-efficient learning process that outputs dynamic models that enjoy intrinsic structure that enables effective, stable, and robust control. While the technical contributions of the paper are excellent themselves, it is this conceptual contribution that I view as most exciting and significant.”

This research is supported, in part, by the NASA University Leadership Initiative and the Natural Sciences and Engineering Research Council of Canada.

Imagine purchasing a robot to perform household tasks. This robot was built and trained in a factory on a certain set of tasks and has never seen the items in your home. When you ask it to pick up a mug from your kitchen table, it might not recognize your mug (perhaps because this mug is painted with an unusual image, say, of MIT’s mascot, Tim the Beaver). So, the robot fails.

“Right now, the way we train these robots, when they fail, we don’t really know why. So you would just throw up your hands and say, ‘OK, I guess we have to start over.’ A critical component that is missing from this system is enabling the robot to demonstrate why it is failing so the user can give it feedback,” says Andi Peng, an electrical engineering and computer science (EECS) graduate student at MIT.

Peng and her collaborators at MIT, New York University, and the University of California at Berkeley created a framework that enables humans to quickly teach a robot what they want it to do, with a minimal amount of effort.

When a robot fails, the system uses an algorithm to generate counterfactual explanations that describe what needed to change for the robot to succeed. For instance, maybe the robot would have been able to pick up the mug if the mug were a certain color. It shows these counterfactuals to the human and asks for feedback on why the robot failed. Then the system utilizes this feedback and the counterfactual explanations to generate new data it uses to fine-tune the robot.

Fine-tuning involves tweaking a machine-learning model that has already been trained to perform one task, so it can perform a second, similar task.

The researchers tested this technique in simulations and found that it could teach a robot more efficiently than other methods. The robots trained with this framework performed better, while the training process consumed less of a human’s time.

This framework could help robots learn faster in new environments without requiring a user to have technical knowledge. In the long run, this could be a step toward enabling general-purpose robots to efficiently perform daily tasks for the elderly or individuals with disabilities in a variety of settings.

Peng, the lead author, is joined by co-authors Aviv Netanyahu, an EECS graduate student; Mark Ho, an assistant professor at the Stevens Institute of Technology; Tianmin Shu, an MIT postdoc; Andreea Bobu, a graduate student at UC Berkeley; and senior authors Julie Shah, an MIT professor of aeronautics and astronautics and the director of the Interactive Robotics Group in the Computer Science and Artificial Intelligence Laboratory (CSAIL), and Pulkit Agrawal, a professor in CSAIL. The research will be presented at the International Conference on Machine Learning.

On-the-job training

Robots often fail due to distribution shift — the robot is presented with objects and spaces it did not see during training, and it doesn’t understand what to do in this new environment.

One way to retrain a robot for a specific task is imitation learning. The user could demonstrate the correct task to teach the robot what to do. If a user tries to teach a robot to pick up a mug, but demonstrates with a white mug, the robot could learn that all mugs are white. It may then fail to pick up a red, blue, or “Tim-the-Beaver-brown” mug.

Training a robot to recognize that a mug is a mug, regardless of its color, could take thousands of demonstrations.

“I don’t want to have to demonstrate with 30,000 mugs. I want to demonstrate with just one mug. But then I need to teach the robot so it recognizes that it can pick up a mug of any color,” Peng says.

To accomplish this, the researchers’ system determines what specific object the user cares about (a mug) and what elements aren’t important for the task (perhaps the color of the mug doesn’t matter). It uses this information to generate new, synthetic data by changing these “unimportant” visual concepts. This process is known as data augmentation.

The framework has three steps. First, it shows the task that caused the robot to fail. Then it collects a demonstration from the user of the desired actions and generates counterfactuals by searching over all features in the space that show what needed to change for the robot to succeed.

The system shows these counterfactuals to the user and asks for feedback to determine which visual concepts do not impact the desired action. Then it uses this human feedback to generate many new augmented demonstrations.

In this way, the user could demonstrate picking up one mug, but the system would produce demonstrations showing the desired action with thousands of different mugs by altering the color. It uses these data to fine-tune the robot.

Creating counterfactual explanations and soliciting feedback from the user are critical for the technique to succeed, Peng says.

From human reasoning to robot reasoning

Because their work seeks to put the human in the training loop, the researchers tested their technique with human users. They first conducted a study in which they asked people if counterfactual explanations helped them identify elements that could be changed without affecting the task.

“It was so clear right off the bat. Humans are so good at this type of counterfactual reasoning. And this counterfactual step is what allows human reasoning to be translated into robot reasoning in a way that makes sense,” she says.

Then they applied their framework to three simulations where robots were tasked with: navigating to a goal object, picking up a key and unlocking a door, and picking up a desired object then placing it on a tabletop. In each instance, their method enabled the robot to learn faster than with other techniques, while requiring fewer demonstrations from users.

Moving forward, the researchers hope to test this framework on real robots. They also want to focus on reducing the time it takes the system to create new data using generative machine-learning models.

“We want robots to do what humans do, and we want them to do it in a semantically meaningful way. Humans tend to operate in this abstract space, where they don’t think about every single property in an image. At the end of the day, this is really about enabling a robot to learn a good, human-like representation at an abstract level,” Peng says.

This research is supported, in part, by a National Science Foundation Graduate Research Fellowship, Open Philanthropy, an Apple AI/ML Fellowship, Hyundai Motor Corporation, the MIT-IBM Watson AI Lab, and the National Science Foundation Institute for Artificial Intelligence and Fundamental Interactions.

The MIT Stephen A. Schwarzman College of Computing named Armando Solar-Lezama as the inaugural Distinguished College of Computing Professor, effective July 1. 

Solar-Lezama is the first person appointed to this position generously endowed by Professor Jae S. Lim of the Department of Electrical Engineering and Computer Science (EECS). Established in the MIT Schwarzman College of Computing, the chair is being awarded to Solar-Lezama for being an outstanding faculty member who is recognized as a leader and innovator.

“I’m pleased to make this appointment and recognize Armando for his remarkable contributions to MIT and the scientific community,” says Daniel Huttenlocher, dean of the MIT Schwarzman College of Computing and the Henry Ellis Warren Professor of Electrical Engineering and Computer Science. “I’m greatly appreciative of Professor Lim for his thoughtful gesture in creating this new chair in the college, providing us with the opportunity to acknowledge the accomplishments of our faculty.”

Solar-Lezama, a professor of electrical engineering and computer science, leads the Computer-Aided Programming Group in the Computer Science and Artificial Intelligence Laboratory (CSAIL) that focuses on program synthesis, an area of research that lies at the intersection of programming systems and artificial intelligence. The group’s research ranges from designing new analysis techniques and automated reasoning mechanisms to developing new programming models that automate challenging aspects of programming.

A member of the EECS faculty since 2008, Solar-Lezama, who also serves as the associate director and chief operating officer for CSAIL, is most interested in software synthesis and its applications to particular program domains such as high-performance computing. He first found this niche area of program synthesis as a graduate student at the University of California at Berkeley, for which his thesis project, a language called Sketch, treats program synthesis as a search problem in which the algorithms pare down the search space to make the search faster and more efficient. Since then, program synthesis research has greatly expanded into the active field it is today.

Your brand new household robot is delivered to your house, and you ask it to make you a cup of coffee. Although it knows some basic skills from previous practice in simulated kitchens, there are way too many actions it could possibly take — turning on the faucet, flushing the toilet, emptying out the flour container, and so on. But there’s a tiny number of actions that could possibly be useful. How is the robot to figure out what steps are sensible in a new situation?

It could use PIGINet, a new system that aims to efficiently enhance the problem-solving capabilities of household robots. Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) are using machine learning to cut down on the typical iterative process of task planning that considers all possible actions. PIGINet eliminates task plans that can’t satisfy collision-free requirements, and reduces planning time by 50-80 percent when trained on only 300-500 problems. 

Typically, robots attempt various task plans and iteratively refine their moves until they find a feasible solution, which can be inefficient and time-consuming, especially when there are movable and articulated obstacles. Maybe after cooking, for example, you want to put all the sauces in the cabinet. That problem might take two to eight steps depending on what the world looks like at that moment. Does the robot need to open multiple cabinet doors, or are there any obstacles inside the cabinet that need to be relocated in order to make space? You don’t want your robot to be annoyingly slow — and it will be worse if it burns dinner while it’s thinking.

Household robots are usually thought of as following predefined recipes for performing tasks, which isn’t always suitable for diverse or changing environments. So, how does PIGINet avoid those predefined rules? PIGINet is a neural network that takes in “Plans, Images, Goal, and Initial facts,” then predicts the probability that a task plan can be refined to find feasible motion plans. In simple terms, it employs a transformer encoder, a versatile and state-of-the-art model designed to operate on data sequences. The input sequence, in this case, is information about which task plan it is considering, images of the environment, and symbolic encodings of the initial state and the desired goal. The encoder combines the task plans, image, and text to generate a prediction regarding the feasibility of the selected task plan. 

Keeping things in the kitchen, the team created hundreds of simulated environments, each with different layouts and specific tasks that require objects to be rearranged among counters, fridges, cabinets, sinks, and cooking pots. By measuring the time taken to solve problems, they compared PIGINet against prior approaches. One correct task plan may include opening the left fridge door, removing a pot lid, moving the cabbage from pot to fridge, moving a potato to the fridge, picking up the bottle from the sink, placing the bottle in the sink, picking up the tomato, or placing the tomato. PIGINet significantly reduced planning time by 80 percent in simpler scenarios and 20-50 percent in more complex scenarios that have longer plan sequences and less training data.

“Systems such as PIGINet, which use the power of data-driven methods to handle familiar cases efficiently, but can still fall back on “first-principles” planning methods to verify learning-based suggestions and solve novel problems, offer the best of both worlds, providing reliable and efficient general-purpose solutions to a wide variety of problems,” says MIT Professor and CSAIL Principal Investigator Leslie Pack Kaelbling.

PIGINet’s use of multimodal embeddings in the input sequence allowed for better representation and understanding of complex geometric relationships. Using image data helped the model to grasp spatial arrangements and object configurations without knowing the object 3D meshes for precise collision checking, enabling fast decision-making in different environments. 

One of the major challenges faced during the development of PIGINet was the scarcity of good training data, as all feasible and infeasible plans need to be generated by traditional planners, which is slow in the first place. However, by using pretrained vision language models and data augmentation tricks, the team was able to address this challenge, showing impressive plan time reduction not only on problems with seen objects, but also zero-shot generalization to previously unseen objects.

“Because everyone’s home is different, robots should be adaptable problem-solvers instead of just recipe followers. Our key idea is to let a general-purpose task planner generate candidate task plans and use a deep learning model to select the promising ones. The result is a more efficient, adaptable, and practical household robot, one that can nimbly navigate even complex and dynamic environments. Moreover, the practical applications of PIGINet are not confined to households,” says Zhutian Yang, MIT CSAIL PhD student and lead author on the work. “Our future aim is to further refine PIGINet to suggest alternate task plans after identifying infeasible actions, which will further speed up the generation of feasible task plans without the need of big datasets for training a general-purpose planner from scratch. We believe that this could revolutionize the way robots are trained during development and then applied to everyone’s homes.” 

“This paper addresses the fundamental challenge in implementing a general-purpose robot: how to learn from past experience to speed up the decision-making process in unstructured environments filled with a large number of articulated and movable obstacles,” says Beomjoon Kim PhD ’20, assistant professor in the Graduate School of AI at Korea Advanced Institute of Science and Technology (KAIST). “The core bottleneck in such problems is how to determine a high-level task plan such that there exists a low-level motion plan that realizes the high-level plan. Typically, you have to oscillate between motion and task planning, which causes significant computational inefficiency. Zhutian’s work tackles this by using learning to eliminate infeasible task plans, and is a step in a promising direction.”

Yang wrote the paper with NVIDIA research scientist Caelan Garrett SB ’15, MEng ’15, PhD ’21; MIT Department of Electrical Engineering and Computer Science professors and CSAIL members Tomás Lozano-Pérez and Leslie Kaelbling; and Senior Director of Robotics Research at NVIDIA and University of Washington Professor Dieter Fox. The team was supported by AI Singapore and grants from National Science Foundation, the Air Force Office of Scientific Research, and the Army Research Office. This project was partially conducted while Yang was an intern at NVIDIA Research. Their research will be presented in July at the conference Robotics: Science and Systems.

Amid a huge amount of hype around generative AI, a new study from researchers at MIT sheds light on the technology’s impact on work, finding that it increased productivity for workers assigned tasks like writing cover letters, delicate emails, and cost-benefit analyses.

The tasks in the study weren’t quite replicas of real work: They didn’t require precise factual accuracy or context about things like a company’s goals or a customer’s preferences. Still, a number of the study’s participants said the assignments were similar to things they’d written in their real jobs — and the benefits were substantial. Access to the assistive chatbot ChatGPT decreased the time it took workers to complete the tasks by 40 percent, and output quality, as measured by independent evaluators, rose by 18 percent.

The researchers hope the study, which appears today in open-access form in the journal Science, helps people understand the impact that AI tools like ChatGPT can have on the workforce.

“What we can say for sure is generative AI is going to have a big effect on white collar work,” says Shakked Noy, a PhD student in MIT’s Department of Economics, who co-authored the paper with fellow PhD student Whitney Zhang ’21. “I think what our study shows is that this kind of technology has important applications in white collar work. It’s a useful technology. But it’s still too early to tell if it will be good or bad, or how exactly it’s going to cause society to adjust.”

Simulating work for chatbots

For centuries, people have worried that new technological advancements would lead to mass automation and job loss. But new technologies also create new jobs, and when they increase worker productivity, they can have a net positive effect on the economy.

“Productivity is front of mind for economists when thinking of new technological developments,” Noy says. “The classical view in economics is that the most important thing that technological advancement does is raise productivity, in the sense of letting us produce economic output more efficiently.”

To study generative AI’s effect on worker productivity, the researchers gave 453 college-educated marketers, grant writers, consultants, data analysts, human resource professionals, and managers two writing tasks specific to their occupation. The 20- to 30-minute tasks included writing cover letters for grant applications, emails about organizational restructuring, and plans for analyses helping a company decide which customers to send push notifications to based on given customer data. Experienced professionals in the same occupations as each participant evaluated each submission as if they were encountering it in a work setting. Evaluators did not know which submissions were created with the help of ChatGPT.

Half of participants were given access to the chatbot ChatGPT-3.5, developed by the company OpenAI, for the second assignment. Those users finished tasks 11 minutes faster than the control group, while their average quality evaluations increased by 18 percent.

The data also showed that performance inequality between workers decreased, meaning workers who received a lower grade in the first task benefitted more from using ChatGPT for the second task.

The researchers say the tasks were broadly representative of assignments such professionals see in their real jobs, but they noted a number of limitations. Because they were using anonymous participants, the researchers couldn’t require contextual knowledge about a specific company or customer. They also had to give explicit instructions for each assignment, whereas real-world tasks may be more open-ended. Additionally, the researchers didn’t think it was feasible to hire fact-checkers to evaluate the accuracy of the outputs. Accuracy is a major problem for today’s generative AI technologies.

The researchers said those limitations could lessen ChatGPT’s productivity-boosting potential in the real world. Still, they believe the results show the technology’s promise — an idea supported by another of the study’s findings: Workers exposed to ChatGPT during the experiment were twice as likely to report using it in their real job two weeks after the experiment.

“The experiment demonstrates that it does bring significant speed benefits, even if those speed benefits are lesser in the real world because you need to spend time fact-checking and writing the prompts,” Noy says.

Taking the macro view

The study offered a close-up look at the impact that tools like ChatGPT can have on certain writing tasks. But extrapolating that impact out to understand generative AI’s effect on the economy is more difficult. That’s what the researchers hope to work on next.

“There are so many other factors that are going to affect wages, employment, and shifts across sectors that would require pieces of evidence that aren’t in our paper,” Zhang says. “But the magnitude of time saved and quality increases are very large in our paper, so it does seem like this is pretty revolutionary, at least for certain types of work.”

Both researchers agree that, even if it’s accepted that ChatGPT will increase many workers’ productivity, much work remains to be done to figure out how society should respond to generative AI’s proliferation.

“The policy needed to adjust to these technologies can be very different depending on what future research finds,” Zhang says. “If we think this will boost wages for lower-paid workers, that’s a very different implication than if it’s going to increase wage inequality by boosting the wages of already high earners. I think there’s a lot of downstream economic and political effects that are important to pin down.”

The study was supported by an Emergent Ventures grant, the Mercatus Center, George Mason University, a George and Obie Shultz Fund grant, the MIT Department of Economics, and a National Science Foundation Graduate Research Fellowship Grant.