Category: Researchers

The study, published in Scientific Reports (Nature Portfolio), points to a significant shift. Generative AI systems have now reached a level where they can outperform the average human on certain creativity measures. At the same time, the most creative people still show a clear and consistent advantage over even the strongest AI models.

AI Reaches Average Human Creativity Levels

Researchers evaluated several leading large language models, including ChatGPT, Claude, Gemini, and others, and compared their performance with results from more than 100,000 human participants. The findings highlight a clear turning point. Some AI systems, including GPT-4, exceeded average human scores on tasks designed to measure divergent linguistic creativity.

“Our study shows that some AI systems based on large language models can now outperform average human creativity on well-defined tasks,” explains Professor Karim Jerbi. “This result may be surprising — even unsettling — but our study also highlights an equally important observation: even the best AI systems still fall short of the levels reached by the most creative humans.”

Further analysis by the study’s co-first authors, postdoctoral researcher Antoine Bellemare-Pépin (Université de Montréal) and PhD candidate François Lespinasse (Université Concordia), revealed a striking pattern. While some AI models now outperform the average person, peak creativity remains firmly human.

In fact, when researchers examined the most creative half of participants, their average scores surpassed those of every AI model tested. The gap grew even larger among the top 10 percent of the most creative individuals.

“We developed a rigorous framework that allows us to compare human and AI creativity using the same tools, based on data from more than 100,000 participants, in collaboration with Jay Olson from the University of Toronto,” says Professor Karim Jerbi, who is also an associate professor at Mila.

How Scientists Measure Creativity in Humans and AI

To evaluate creativity fairly across humans and machines, the research team used multiple methods. The primary tool was the Divergent Association Task (DAT), a widely used psychological test that measures divergent creativity, or the ability to generate diverse and original ideas from a single prompt.

Created by study co-author Jay Olson, the DAT asks participants, whether human or AI, to list ten words that are as unrelated in meaning as possible. An example of a highly creative response includes words like “galaxy, fork, freedom, algae, harmonica, quantum, nostalgia, velvet, hurricane, photosynthesis.”

Performance on this task is strongly linked to results on other established creativity tests used in writing, idea generation, and creative problem solving. Although the task is language-based, it goes well beyond vocabulary. It engages broader cognitive processes involved in creative thinking across many domains. The DAT also has practical advantages, as it takes only two to four minutes to complete and can be accessed online by the general public.

From Word Lists to Real Creative Writing

The researchers then explored whether AI success on this simple word association task could extend to more complex and realistic creative activities. To test this, they compared AI systems and human participants on creative writing challenges such as composing haiku (a short three-line poetic form), writing movie plot summaries, and producing short stories.

The results followed a familiar pattern. While AI systems sometimes exceeded the performance of average humans, the most skilled human creators consistently delivered stronger and more original work.

Can AI Creativity Be Adjusted?

These findings raised another important question. Is AI creativity fixed, or can it be shaped? The study shows that creativity in AI can be adjusted by changing technical settings, particularly the model’s temperature. This parameter controls how predictable or adventurous the generated responses are.

At lower temperature settings, AI produces safer and more conventional outputs. At higher temperatures, responses become more varied, less predictable, and more exploratory, allowing the system to move beyond familiar ideas.

The researchers also found that creativity is strongly influenced by how instructions are written. For example, prompts that encourage models to think about word origins and structure using etymology lead to more unexpected associations and higher creativity scores. These results emphasize that AI creativity depends heavily on human guidance, making interaction and prompting a central part of the creative process.

Will AI Replace Human Creators?

The study offers a balanced perspective on fears that artificial intelligence could replace creative professionals. While AI systems can now match or exceed average human creativity on certain tasks, they still have clear limitations and rely on human direction.

“Even though AI can now reach human-level creativity on certain tests, we need to move beyond this misleading sense of competition,” says Professor Karim Jerbi. “Generative AI has above all become an extremely powerful tool in the service of human creativity: it will not replace creators, but profoundly transform how they imagine, explore, and create — for those who choose to use it.”

Rather than signaling the end of creative careers, the findings suggest a future where AI serves as a creative assistant. By expanding ideas and opening new paths for exploration, AI may help amplify human imagination rather than replace it.

“By directly confronting human and machine capabilities, studies like ours push us to rethink what we mean by creativity,” concludes Professor Karim Jerbi.

About the Study

The paper titled “Divergent creativity in humans and large language models” was published in Scientific Reports on January 21, 2026. The research brought together scientists from Université de Montréal, Université Concordia, University of Toronto Mississauga, Mila (Quebec AI Institute), and Google DeepMind.

Professor Karim Jerbi led the study, with Antoine Bellemare-Pépin (Université de Montréal) and François Lespinasse (Université Concordia) serving as co-first authors. The research team also included Yoshua Bengio, founder of Mila and LoiZéro, and a pioneer of deep learning, the technology behind modern AI systems such as ChatGPT.

“What is special about these synthetically designed elements is that they show remarkable specificity to the target cell type they were designed for,” said Ryan Tewhey, PhD, an associate professor at The Jackson Laboratory and co-senior author of the work. “This creates the opportunity for us to turn the expression of a gene up or down in just one tissue without affecting the rest of the body.”

In recent years, genetic editing technologies and other gene therapy approaches have given scientists the ability to alter the genes inside living cells. However, affecting genes only in selected cell types or tissues, rather than across an entire organism, has been difficult. That is in part because of the ongoing challenge of understanding the DNA switches, called cis-regulatory elements (CREs), that control the expression and repression of genes.

In a paper published in Oct. 23 advanced online issue of Nature, Tewhey and his collaborators not only designed new, never-before-seen synthetic CREs, but used the CREs to successfully activate genes in brain, liver or blood cells without turning on those genes in other cell types.

Tissue- and time-specific instructions

Although every cell in an organism contains the same genes, not all the genes are needed in every cell, or at all times. CREs help ensure that genes needed in the brain are not used by skin cells, for instance, or that genes required during early development are not activated in adults. CREs themselves are not part of genes, but are separate, regulatory DNA sequences — often located near the genes they control.

Scientists know that there are thousands of different CREs in the human genome, each with slightly different roles. But the grammar of CREs has been hard to figure out, “with no straightforward rules that control what each CRE does,” explained Rodrigo Castro, PhD, a computational scientist in the Tewhey lab at JAX and co-first author of the new paper. “This limits our ability to design gene therapies that only effect certain cell types in the human body.”

“This project essentially asks the question: ‘Can we learn to read and write the code of these regulatory elements?’” said Steven Reilly, PhD, assistant professor of genetics at Yale and one of the senior authors of the study. “If we think about it in terms of language, the grammar and syntax of these elements is poorly understood. And so, we tried to build machine learning methods that could learn a more complex code than we could do on our own.”

Using a form of artificial intelligence (AI) called deep learning, the group trained a model using hundreds of thousands of DNA sequences from the human genome that they measured in the laboratory for CRE activity in three types of cells: blood, liver and brain. The AI model allowed the researchers to predict the activity for any sequence from the almost infinite number of possible combinations. By analyzing these predictions, the researchers discovered new patterns in the DNA, learning how the grammar of CRE sequences in the DNA impact how much RNA would be made — a proxy for how much a gene is activated.

The team, including Pardis Sabeti, MD, DPhil, co-senior author of the study and a core institute member at the Broad Institute and professor at Harvard, then developed a platform called CODA (Computational Optimization of DNA Activity), which used their AI model to efficiently design thousands of completely new CREs with requested characteristics, like activating a particular gene in human liver cells but not activating the same gene in human blood or brain cells. Through an iterative combination of ‘wet’ and ‘dry’ investigation, using experimental data to first build and then validate computational models, the researchers refined and improved the program’s ability to predict the biological impact of each CRE and enabled the design of specific CREs never before seen in nature.

“Natural CREs, while plentiful, represent a tiny fraction of possible genetic elements and are constrained in their function by natural selection,” said study co-first author Sager Gosai, PhD, a postdoctoral fellow in Sabeti’s lab. “These AI tools have immense potential for designing genetic switches that precisely tune gene expression for novel applications, such as biomanufacturing and therapeutics, that lie outside the scope of evolutionary pressures.”

Pick-and-choose your organ

Tewhey and his colleagues tested the new, AI-designed synthetic CREs by adding them into cells and measuring how well they activated genes in the desired cell type, as well as how good they were at avoiding gene expression in other cells. The new CREs, they discovered, were even more cell-type-specific than naturally occurring CREs known to be associated with the cell types.

“The synthetic CREs semantically diverged so far from natural elements that predictions for their effectiveness seemed implausible,” said Gosai. “We initially expected many of the sequences would misbehave inside living cells.”

“It was a thrilling surprise to us just how good CODA was at designing these elements,” said Castro.

Tewhey and his collaborators studied why the synthetic CREs were able to outperform naturally occurring CREs and discovered that the cell-specific synthetic CREs contained combinations of sequences responsible for expressing genes in the target cell types, as well as sequences that repressed or turned off the gene in the other cell types.

Finally, the group tested several of the synthetic CRE sequences in zebrafish and mice, with good results. One CRE, for instance, was able to activate a fluorescent protein in developing zebrafish livers but not in any other areas of the fish.

“This technology paves the way toward the writing of new regulatory elements with pre-defined functions,” said Tewhey. “Such tools will be valuable for basic research but also could have significant biomedical implications where you could use these elements to control gene expression in very specific cell types for therapeutic purposes.”

Although each condition occurs in a small number of individuals, collectively these diseases exert a staggering human and economic toll because they affect some 300 million people worldwide.

Yet, with a mere 5 to 7 percent of these conditions having an FDA-approved drug, they remain largely untreated or undertreated.

Developing new medicines represents a daunting challenge, but a new artificial intelligence tool can propel the discovery of new therapies from existing medicines, offering hope for patients with rare and neglected conditions and for the clinicians who treat them.

The AI model, called TxGNN, is the first one developed specifically to identify drug candidates for rare diseases and conditions with no treatments.

It identified drug candidates from existing medicines for more than 17,000 diseases, many of them without any existing treatments. This represents the largest number of diseases that any single AI model can handle to date. The researchers note that the model could be applied to even more diseases beyond the 17,000 it worked on in the initial experiments.

The work, described Sept. 25 in Nature Medicine, was led by scientists at Harvard Medical School. The researchers have made the tool available for free and want to encourage clinician-scientists to use it in their search for new therapies, especially for conditions with no or with limited treatment options.

“With this tool we aim to identify new therapies across the disease spectrum but when it comes to rare, ultrarare, and neglected conditions, we foresee this model could help close, or at least narrow, a gap that creates serious health disparities,” said lead researcher Marinka Zitnik, assistant professor of biomedical informatics in the Blavatnik Institute at HMS.

“This is precisely where we see the promise of AI in reducing the global disease burden, in finding new uses for existing drugs, which is also a faster and more cost-effective way to develop therapies than designing new drugs from scratch,” added Zitnik, who is an associate faculty member at the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University.

The new tool has two central features — one that identifies treatment candidates along with possible side effects and another one that explains the rationale for the decision.

In total, the tool identified drug candidates from nearly 8,000 medicines (both FDA-approved medicines and experimental ones now in clinical trials) for 17,080 diseases, including conditions with no available treatments. It also predicted which drugs would have side effects and contraindications for specific conditions — something that the current drug discovery approach identifies mostly by trial and error during early clinical trials focused on safety.

Compared against the leading AI models for drug repurposing, the new tool was nearly 50 percent better, on average, at identifying drug candidates. It was also 35 percent more accurate in predicting what drugs would have contraindications.

Advantages of using already approved drugs

Repurposing existing drugs is an alluring way to develop new treatments because it relies on medicines that have been studied, have well-understood safety profiles, and have gone through the regulatory approval process.

Most medicines have multiple effects beyond the specific targets they were originally developed and approved for. But many of these effects remain undiscovered and understudied during initial testing, clinical trials, and review, only emerging after years of use by millions of people. Indeed, nearly 30 percent of FDA- approved drugs have acquired at least one additional indication for treatment following initial approval, and many have acquired tens of additional treatment indications over the years.

This approach to drug repurposing is haphazard at best. It relies on patient reports of unexpected beneficial side effects or on physicians’ intuition about whether to use a drug for a condition that it was not intended for, a practice known as off-label use.

“We’ve tended to rely on luck and serendipity rather than on strategy, which limits drug discovery to diseases for which drugs already exist,” Zitnik said.

The benefits of drug repurposing extend beyond diseases without treatments, Zitnik noted.

“Even for more common diseases with approved treatments, new drugs could offer alternatives with fewer side effects or replace drugs that are ineffective for certain patients,” she said.

What makes the new AI tool better than existing models

Most current AI models used for drug discovery are trained on a single disease or a handful of conditions. Rather than focusing on specific diseases, the new tool was trained in a manner that enables it to use existing data to make new predictions. It does so by identifying shared features across multiple diseases, such as shared genomic aberrations.

For example, the AI model pinpoints shared disease mechanisms based on common genomic underpinnings, which allows it to extrapolate from a well-understood disease with known treatments to a poorly understood one with no treatments.

This capacity, the research team said, brings the AI tool closer to the type of a reasoning a human clinician might use to generate novel ideas if they had access to all the preexisting knowledge and raw data that the AI model does but that the human brain cannot possibly access or store.

The tool was trained on vast amounts of data, including DNA information, cell signaling, levels of gene activity, clinical notes, and more. The researchers tested and refined the model by asking it to perform various tasks. Finally, the tool’s performance was validated on 1.2 million patient records and asked to identify drug candidates for various diseases.

The researchers also asked the tool to predict what patient characteristics would render the identified drug candidates contraindicated for certain patient populations.

Another task involved asking the tool to identify existing small molecules that might effectively block the activity of certain proteins implicated in disease-causing pathways and processes.

In a test designed to gauge the model’s ability to reason as a human clinician might, the researchers prompted the model to find drugs for three rare conditions it had not seen as part of its training — a neurodevelopmental disorder, a connective-tissue disease, and a rare genetic condition that causes water imbalance.

The researchers then compared the model’s recommendations for drug therapy against current medical knowledge about how the suggested drugs work. In every example, the tool’s recommendations aligned with current medical knowledge.

Moreover, the model not only identified medicines for all three diseases but also provided the rationale behind its decision. This explainer feature allows for transparency and can increase physician confidence.

The researchers caution that any therapies identified by the model would require additional evaluation for dosing and timing of delivery. But, they add, with this unprecedented capacity, the new AI model would expedite drug repurposing in a manner not possible until now. The team is already collaborating with several rare disease foundations to help identify possible treatments.

So, he decided to try an experiment.

He tested six forms of generative, large-language AI models — for example ChatGPT — in an online version of his popular introductory course in biomedical and health informatics to see how they performed compared with living, thinking students. A study published in the journal npj Digital Medicine, revealed the answer: Better than as many as three-quarters of his human students.

“This does raise concern about cheating, but there is a larger issue here,” Hersh said. “How do we know that our students are actually learning and mastering the knowledge and skills they need for their future professional work?”

As a professor of medical informatics and clinical epidemiology in the OHSU School of Medicine, Hersh is especially attuned to new technologies. The role of technology in education is nothing new, Hersh said, recalling his own experience as a high school student in the 1970s during the transition from slide rules to calculators.

Yet, the shift to generative AI represents an exponential leap forward.

“Clearly, everyone should have some kind of foundation of knowledge in their field,” Hersh said. “What is the foundation of knowledge you expect people to have to be able to think critically?”

Large-language models

Hersh and co-author Kate Fultz Hollis, an OHSU informatician, pulled the knowledge assessment scores of 139 students who took the introductory course in biomedical and health informatics in 2023. They prompted six generative AI large language models with student assessment materials from the course. Depending on the model, AI scored in the top 50^th to 75^th percentile on multiple-choice questions that were used in quizzes and a final exam that required short written responses to questions.

“The results of this study raise significant questions for the future of student assessment in most, if not all, academic disciplines,” the authors write.

The study is the first to compare large-language models to students for a full academic course in the biomedical field. Hersh and Fultz Hollis noted that a knowledge-based course such as this one may be especially ripe for generative, large-language models, in contrast to more participatory academic courses that help students develop more complex skills and abilities.

Hersh remembers his experience in medical school.

“When I was a medical student, one of my attending physicians told me I needed to have all the knowledge in my head,” he said. “Even in the 1980s, that was a stretch. The knowledge base of medicine has long surpassed the capacity of the human brain to memorize it all.”

Maintaining the human touch

Yet, he believes there’s a fine line between making sensible use of technical resources to advance learning and over-reliance to the point that it inhibits learning. Ultimately, the goal of an academic health center like OHSU is to educate health care professionals capable of caring for patients and optimizing the use of data and information about them in the real world.

In that sense, he said, medicine will always require the human touch.

“There are a lot of things that health care professionals do that are pretty straightforward, but there are those instances where it gets more complicated and you have to make judgment calls,” he said. “That’s when it helps to have that broader perspective, without necessarily needing to have every last fact in your brain.”

With fall classes starting soon, Hersh said he’s not worried about cheating.

“I update the course each year,” he said. “In any scientific field, there are new advancements all the time and large-language models aren’t necessarily up to date on all of it. This just means we’ll have to look at newer or more nuanced tests where you won’t get the answer out of ChatGPT.”

The tool, called Deep Predictor of Binding Specificity (DeepPBS), is a geometric deep learning model designed to predict protein-DNA binding specificity from protein-DNA complex structures. DeepPBS allows scientists and researchers to input the data structure of a protein-DNA complex into an online computational tool.

“Structures of protein-DNA complexes contain proteins that are usually bound to a single DNA sequence. For understanding gene regulation, it is important to have access to the binding specificity of a protein to any DNA sequence or region of the genome,” said Remo Rohs, professor and founding chair in the department of Quantitative and Computational Biology at the USC Dornsife College of Letters, Arts and Sciences. “DeepPBS is an AI tool that replaces the need for high-throughput sequencing or structural biology experiments to reveal protein-DNA binding specificity.”

AI analyzes, predicts protein–DNA structures

DeepPBS employs a geometric deep learning model, a type of machine-learning approach that analyzes data using geometric structures. The AI tool was designed to capture the chemical properties and geometric contexts of protein-DNA to predict binding specificity.

Using this data, DeepPBS produces spatial graphs that illustrate protein structure and the relationship between protein and DNA representations. DeepPBS can also predict binding specificity across various protein families, unlike many existing methods that are limited to one family of proteins.

“It is important for researchers to have a method available that works universally for all proteins and is not restricted to a well-studied protein family. This approach allows us also to design new proteins,” Rohs said.

Major advance in protein-structure prediction

The field of protein-structure prediction has advanced rapidly since the advent of DeepMind’s AlphaFold, which can predict protein structure from sequence. These tools have led to an increase in structural data available to scientists and researchers for analysis. DeepPBS works in conjunction with structure prediction methods for predicting specificity for proteins without available experimental structures.

Rohs said the applications of DeepPBS are numerous. This new research method may lead to accelerating the design of new drugs and treatments for specific mutations in cancer cells, as well as lead to new discoveries in synthetic biology and applications in RNA research.

About the study: In addition to Rohs, other study authors include Raktim Mitra of USC; Jinsen Li of USC; Jared Sagendorf of University of California, San Francisco; Yibei Jiang of USC; Ari Cohen of USC; and Tsu-Pei Chiu of USC; as well as Cameron Glasscock of the University of Washington.

This research was primarily supported by NIH grant R35GM130376.

In a recent article in Nature Methods, researchers at Carnegie Mellon University’s School of Computer Science propose guidelines that outline pitfalls and opportunities for using interpretable machine learning methods to tackle computational biology problems. The Perspectives article, “Applying Interpretable Machine Learning in Computational Biology — Pitfalls, Recommendations and Opportunities for New Developments,” is featured in the journal’s August special issue on AI.

“Interpretable machine learning has generated significant excitement as machine learning and artificial intelligence tools are being applied to increasingly important problems,” said Ameet Talwalkar, an associate professor in CMU’s Machine Learning Department (MLD). “As these models grow in complexity, there is great promise not only in developing highly predictive models but also in creating tools that help end users understand how and why these models make certain predictions. However, it is crucial to acknowledge that interpretable machine learning has yet to deliver turnkey solutions to this interpretability problem.”

The paper is a collaboration between doctoral students Valerie Chen in MLD and Muyu (Wendy) Yang in the Ray and Stephanie Lane Computational Biology Department. Chen’s earlier work critiquing the interpretable machine learning community’s lack of grounding in downstream use cases inspired the article, and the idea was developed through discussions with Yang and Jian Ma, the Ray and Stephanie Lane Professor of Computational Biology.

“Our collaboration began with a deep dive into computational biology papers to survey the application of interpretable machine learning methods,” Yang said. “We noticed that many applications used these methods in a somewhat ad hoc manner. Our goal with this paper was to provide guidelines for more robust and consistent use of interpretable machine learning methods in computational biology.”

One major pitfall the paper addresses is the reliance on a single interpretable machine learning method. Instead, the researchers recommend using multiple interpretable machine learning methods with diverse sets of hyperparameters and comparing their results to obtain a more comprehensive understanding of the model behavior and its underlying interpretations.

“While some machine learning models seem to work surprisingly well, we often do not fully understand why,” Ma said. “In scientific domains like biomedicine, understanding why models work is crucial for discovering fundamental biological mechanisms.”

The paper also warns against cherry-picking results when evaluating interpretable machine learning methods, as this can lead to incomplete or biased interpretations of scientific findings.

Chen emphasized that the guidelines may have broader implications for a wider audience of researchers interested in applying interpretable machine-learning methods to their work.

“We hope that machine learning researchers developing new interpretable machine learning methods and tools — particularly those working on explaining large language models — will carefully consider the human-centric aspects of interpretable machine learning,” Chen said. “This includes understanding who their target user is and how the method will be used and evaluated.”

While understanding model behavior remains crucially important for scientific discovery and a fundamentally unsolved machine learning problem, the authors hope these challenges spur further interdisciplinary collaborations to facilitate the broader use of AI for scientific impact.