AI isn’t as good at recognizing objects as people are
Artificial intelligence (AI) has gradually made its way into many areas of life with abilities that often match or surpass those of humans. But a study in the Cell Press journal iScience publishing on September 17 finds that AI-powered machines have trouble recognizing objects from their overall shapes when aspects of an image are distorted. The findings offer a new way to evaluate and compare computer vision to that of people.
“Current AI models do not accomplish visual object perception in the same way that humans do,” says author Biyu J. He of New York University. “We tested more than 200 deep neural networks, and no models fully reproduced humans’ object recognition patterns. Humans’ ability to leverage the global shape cue for visual object recognition remains unparalleled.”
When you’re presented with an object, you can recognize it based on multiple features, including its texture, small internal details, and overall shape or silhouette. But, compared to people, machines can get tripped up by even small image distortions.
Earlier studies have shown that computational deep neural network (DNN) models, which are developed by training computers on large amounts of data and images, can rival the abilities of human observers on certain tasks. Over time, the machines “learn” complex patterns that allow them to visually recognize and name objects based on their appearance.
To find out if existing DNN models for machine vision work the same way human vision does, He and her colleagues created a set of images to put the AI models to the test. They systematically altered the global shapes, internal parts, and textures of many images, such as a cat, butterfly, car, and corn on the cob, and then compared humans’ ability to correctly identify those objects to that of dozens of the best DNN models.
They found that humans were better at recognizing objects from their overall shapes than the AI algorithms. The computer algorithms consistently underperformed compared to people anytime object recognition depended only on global shape recognition.
“These results show that current computer vision models are not quite human-aligned, despite being trained on a massive number of pictures that humans have taken,” says He.
The study offers a new way to evaluate and compare computer vision to that of people. The researchers say that their results may lead to improvements in computer vision in the future. Such improvements could aid in the development of improved assistive devices, including brain-computer interfaces that hold promise for enabling people with disabilities to better see and act in the real world.
The team says that they’ll continue making comparisons between AI and the human brain, with the goal to “help build more human-aligned AI.”
###
This work was supported by the W. M. Keck Foundation, National Institutes of Health, Astellas Foundation, and Uehara Memorial Foundation.
iScience, Kato & He, “Systematic image perturbations reveal persistent gaps between human and machine vision” https://www.cell.com/iscience/fulltext/S2589-0042(26)02748-3
iScience (@iScience_CP) is an open access journal from Cell Press that provides a platform for original research and interdisciplinary thinking in the life, physical, health, and earth sciences. The primary criterion for publication in iScience is a significant contribution to a relevant field combined with robust results and underlying methodology. Visit: http://www.cell.com/iscience. To receive Cell Press media alerts, contact press@cell.com.
Artificial intelligence (AI) has gradually made its way into many areas of life with abilities that often match or surpass those of humans. But a study in the Cell Press journal iScience publishing on September 17 finds that AI-powered machines have trouble recognizing objects from their overall shapes when aspects of an image are distorted. The findings offer a new way to evaluate and compare computer vision to that of people.
“Current AI models do not accomplish visual object perception in the same way that humans do,” says author Biyu J. He of New York University. “We tested more than 200 deep neural networks, and no models fully reproduced humans’ object recognition patterns. Humans’ ability to leverage the global shape cue for visual object recognition remains unparalleled.”
When you’re presented with an object, you can recognize it based on multiple features, including its texture, small internal details, and overall shape or silhouette. But, compared to people, machines can get tripped up by even small image distortions.
Earlier studies have shown that computational deep neural network (DNN) models, which are developed by training computers on large amounts of data and images, can rival the abilities of human observers on certain tasks. Over time, the machines “learn” complex patterns that allow them to visually recognize and name objects based on their appearance.
To find out if existing DNN models for machine vision work the same way human vision does, He and her colleagues created a set of images to put the AI models to the test. They systematically altered the global shapes, internal parts, and textures of many images, such as a cat, butterfly, car, and corn on the cob, and then compared humans’ ability to correctly identify those objects to that of dozens of the best DNN models.
They found that humans were better at recognizing objects from their overall shapes than the AI algorithms. The computer algorithms consistently underperformed compared to people anytime object recognition depended only on global shape recognition.
“These results show that current computer vision models are not quite human-aligned, despite being trained on a massive number of pictures that humans have taken,” says He.
The study offers a new way to evaluate and compare computer vision to that of people. The researchers say that their results may lead to improvements in computer vision in the future. Such improvements could aid in the development of improved assistive devices, including brain-computer interfaces that hold promise for enabling people with disabilities to better see and act in the real world.
The team says that they’ll continue making comparisons between AI and the human brain, with the goal to “help build more human-aligned AI.”
###
This work was supported by the W. M. Keck Foundation, National Institutes of Health, Astellas Foundation, and Uehara Memorial Foundation.
iScience, Kato & He, “Systematic image perturbations reveal persistent gaps between human and machine vision” https://www.cell.com/iscience/fulltext/S2589-0042(26)02748-3
iScience (@iScience_CP) is an open access journal from Cell Press that provides a platform for original research and interdisciplinary thinking in the life, physical, health, and earth sciences. The primary criterion for publication in iScience is a significant contribution to a relevant field combined with robust results and underlying methodology. Visit: http://www.cell.com/iscience. To receive Cell Press media alerts, contact press@cell.com.
Journal
iScience
iScience
DOI
Method of Research
Experimental study
Experimental study
Subject of Research
Not applicable
Not applicable
Article Title
Systematic image perturbations reveal persistent gaps between human and machine vision
Systematic image perturbations reveal persistent gaps between human and machine vision
Article Publication Date
17-Sep-2026
17-Sep-2026
How a simple game of 20 questions could help make AI fit for the future
Artificial intelligence programmes used to classify images could be trained much cheaper using a surprisingly simple method inspired by the childhood game of 20 Questions, new research has found.
The research, led by the University of Bristol and presented at the Allerton Conference on Communication, Control, and Computing in Illinois, in the US on Wednesday 16 September, shows how simple binary classifiers, each trained in a matter of minutes on a standard laptop, can be combined using a string of yes/no questions to perform complex classification tasks. These tasks normally require tens of thousands of Graphics Processing Units to train, costing millions of dollars.
Prof Sidharth Jaggi , Professor of Mathematics at the University of Bristol School of Mathematics, explained: “If you were out on a walk and wanted to identify an unusual species of plant you spotted using an AI app, a current classifier programme would likely be designed to carefully separate out millions, if not billions, of different types of objects from each other.
“In this new piece of work, we are able to mathematically prove – and back up by empirical validation – that even if we ask ‘simple random questions’ the answers can be combined to perform complex tasks.
“The key observation is that no complex coordination of the simple binary classifiers is required. There just need to be enough of them, and that number is surprisingly small. The implication is that our algorithms have a far lower computational cost, are more robust, and are easier to deploy at scale – all properties that are critical for real-world use.”
The approach works especially well for AI systems that operate across multiple devices or directly within smart devices, such as sensors, robots, and edge devices which process data close to where it is generated. Because each question is answered independently, the overall system can still produce reliable results even if some individual answers are incorrect.
Lead author Dr Ioannis Papageorgiou, who carried out the research while working as a Senior Research Associate at the University of Bristol, said: “What is exciting about this approach is that a very large and difficult classification problem can be broken down into lots of much simpler yes-or-no decisions, chosen at random. Each individual classifier only needs to answer one of these simple questions, but together they can identify from millions of possibilities.
“As artificial intelligence becomes increasingly embedded in our lives, from healthcare and transport to manufacturing and national infrastructure, the challenge is shifting. The key question is no longer just how to make AI systems more powerful, but how to make them efficient, trustworthy, and resilient in real‑world conditions.”
The team’s work forms part of the Informed AI research hub at the University of Bristol, which tackles a range of foundational problems traversing mathematics, information theory, and AI safety.
By providing these theoretical foundations for efficiency, robustness, and reliability, Informed AI research aims to ensure that future UK AI systems are not only innovative, but also safe, trustworthy, and socially deployable. As AI continues to move into everyday environments, these principles will be essential for maintaining public confidence and delivering long‑term value for the economy and society at large.
Paper:
‘Fundamental limits of distributed multiclass classification from simple binary decisions’, by I.Papageorgiou et al. in Arxiv
Notes to editors:
For further information or to arrange an interview with Sidharth Jaggi or Ioannis Papageorgiou, please contact Steve Salter, email steven.salter@bristol.ac.uk, mobile +44 (0)7964 022596 in the University of Bristol News and Content team.
Artificial intelligence programmes used to classify images could be trained much cheaper using a surprisingly simple method inspired by the childhood game of 20 Questions, new research has found.
The research, led by the University of Bristol and presented at the Allerton Conference on Communication, Control, and Computing in Illinois, in the US on Wednesday 16 September, shows how simple binary classifiers, each trained in a matter of minutes on a standard laptop, can be combined using a string of yes/no questions to perform complex classification tasks. These tasks normally require tens of thousands of Graphics Processing Units to train, costing millions of dollars.
Prof Sidharth Jaggi , Professor of Mathematics at the University of Bristol School of Mathematics, explained: “If you were out on a walk and wanted to identify an unusual species of plant you spotted using an AI app, a current classifier programme would likely be designed to carefully separate out millions, if not billions, of different types of objects from each other.
“In this new piece of work, we are able to mathematically prove – and back up by empirical validation – that even if we ask ‘simple random questions’ the answers can be combined to perform complex tasks.
“The key observation is that no complex coordination of the simple binary classifiers is required. There just need to be enough of them, and that number is surprisingly small. The implication is that our algorithms have a far lower computational cost, are more robust, and are easier to deploy at scale – all properties that are critical for real-world use.”
The approach works especially well for AI systems that operate across multiple devices or directly within smart devices, such as sensors, robots, and edge devices which process data close to where it is generated. Because each question is answered independently, the overall system can still produce reliable results even if some individual answers are incorrect.
Lead author Dr Ioannis Papageorgiou, who carried out the research while working as a Senior Research Associate at the University of Bristol, said: “What is exciting about this approach is that a very large and difficult classification problem can be broken down into lots of much simpler yes-or-no decisions, chosen at random. Each individual classifier only needs to answer one of these simple questions, but together they can identify from millions of possibilities.
“As artificial intelligence becomes increasingly embedded in our lives, from healthcare and transport to manufacturing and national infrastructure, the challenge is shifting. The key question is no longer just how to make AI systems more powerful, but how to make them efficient, trustworthy, and resilient in real‑world conditions.”
The team’s work forms part of the Informed AI research hub at the University of Bristol, which tackles a range of foundational problems traversing mathematics, information theory, and AI safety.
By providing these theoretical foundations for efficiency, robustness, and reliability, Informed AI research aims to ensure that future UK AI systems are not only innovative, but also safe, trustworthy, and socially deployable. As AI continues to move into everyday environments, these principles will be essential for maintaining public confidence and delivering long‑term value for the economy and society at large.
Paper:
‘Fundamental limits of distributed multiclass classification from simple binary decisions’, by I.Papageorgiou et al. in Arxiv
Notes to editors:
For further information or to arrange an interview with Sidharth Jaggi or Ioannis Papageorgiou, please contact Steve Salter, email steven.salter@bristol.ac.uk, mobile +44 (0)7964 022596 in the University of Bristol News and Content team.
Article Title
Fundamental limits of distributed multiclass classification from simple binary decisions
Fundamental limits of distributed multiclass classification from simple binary decisions
Virtual biotech company puts thousands of AI scientist agents to work on drug discovery
Virtual biotech company
The latest company to spin out of a Stanford Medicine lab is a biotech undertaking with 37,000 employees — and none of them are human. There’s no lab space, no lunch breaks and no payroll. It’s an artificial intelligence-powered virtual biotech company that’s the brainchild of associate professor of biomedical data science James Zou, PhD, who is also the principal investigator of a virtual lab that launched in 2025, and graduate student Harrison Zhang.
The idea, Zou said, was to build on the virtual lab, in which AI scientists emulate an academic research laboratory. They created an entire company with tens of thousands of AI agents, all trained to support the full pipeline of drug development.
“Our idea was to see how far we could push this. Could we create a biotech company that takes on everything from looking for drug targets all the way to designing clinical trials?” Zou said. “Could we have a fully agentic company that tackles the extremely complex challenges of drug discovery?”
One advantage of an AI company is that you can skip the startup phase. Zou’s virtual company more or less mirrors the organizational chart of an established brick-and-mortar biotech: A chief science officer agent leads the research teams, which are broken into multiple specialized divisions that work in parallel to tackle the core elements of drug design, such as identifying molecular targets and designing clinical trials.
The resulting effort, while intangible, so far seems powerful. The virtual biotech company has been able to uncover a biological signal that predicts which drug candidates are more likely to succeed and has designed a cancer therapy that a major pharmaceutical company later independently built.
A paper describing the virtual biotech company will publish on Sept. 17 in Science. Zou is the senior author, and Zhang is the lead author.
A biological clue
One of the biggest challenges in drug discovery is determining which molecules are most likely to succeed in clinical trials, Zou said. “The end-to-end process of a clinical trial can cost tens — sometimes hundreds — of millions of dollars, and it can take many years.” If there are hidden biological features that could tip scientists off to a drug’s odds of clearing clinical trials, that would be a huge boon to the industry, he added.
He sent the virtual biotech agents hunting for any characteristics that might set successful drugs apart. Instead of loosing a cadre of agents into repositories of scientific literature, Zou took a more meticulous approach, assigning a single agent to analyze a specific clinical trial and to retrieve relevant data about safety and effectiveness. In total, the agents analyzed and catalogued some 50,000 trials in less than a week. That task would have taken human agents years, he said.
Alongside the trial result analysis, the agents were tasked with investigating molecular data collected during the trial. “Here, the virtual agents did something quite interesting,” Zou said. For trials with single-cell gene activity data available, the agents built two scoring systems: one that evaluated how specifically a drug targeted a certain cell type (as opposed to affecting lots of cell types broadly) and one that measured something called bimodality, which indicates whether a targeted gene’s activity is more akin to a light switch (on-off — high bimodality) or a dimmer.
The agents found that trials with high scores in both categories fared better. Drugs that targeted switch-like genes were 40% more likely to advance from phase 1 to phase 2 trials, were 48% more likely to reach market, and had 32% fewer adverse events compared with those that had a broad spectrum of activity. What’s more, these patterns persisted for a variety of conditions, including cancers, brain diseases, heart diseases, kidney and lung conditions, and more. Zou’s theory: A target that behaves like an on-off switch and homes in on a specific cell type may be easier, and therefore safer, to control with a drug as opposed to one with a spectrum of activity.
“The science the agents discovered is really exciting, and it shows that these single-cell features can be used to make better drugs. It points to the importance of collecting this kind of data,” Zou said. “This could help the entire drug discovery industry.”
AI-designed, real world ready
Still, the question remained: Could an all-AI company design a new drug capable of helping people? To test this, Zou and his team turned the agents’ attention to a protein that lung cancer researchers have long eyed — B7-H3. The agents analyzed relevant data from a variety of studies and biomedical data repositories and reported that B7-H3 was highly expressed in a cell type known as fibroblasts, which are found in connective tissue and often live near tumor cells.
The agents looked more closely at communication between cells and at spatial-transcriptomic analyses (which map the activity of certain cells in a specific location); they found that fibroblasts expressing B7-H3 seemed to be signaling to nearby immune cells and suppressing their activity, effectively cloaking the tumor from normal immune defenses. The AI scientists designed something called an antibody-drug conjugate: a protein-based tag team that homes in on cells harboring many of the B7-H3 proteins and delivers a toxic chemotherapy payload directly to them.
The agents proposed this drug design using information available before January 2025. Months later, in August 2025, a private, well-established pharmaceutical company independently arrived at the same antibody-drug conjugate strategy against B7-H3. That therapy went on to receive a Food and Drug Administration breakthrough therapy designation, which helps fast-track promising drugs to market after showing effectiveness in a human study. “This was really exciting as an independent, third-party validation that’s consistent with the effects and the design proposed by the virtual biotech,” Zou said.
Zou isn’t chasing the B7-H3 target any further, as it’s already being shepherded into clinics by another company. But he says the virtual biotech has surfaced other candidate targets designed in a similar way. Humans, physical experimentation and validation will always be the conduit through which AI makes an impact, Zou said. “Our next step is to bring the new findings from the virtual biotech into real labs and test how many hold up in the real world.”
This study was funded by a Knight-Hennessy Scholarship and the National Institutes of Health (grant T32-GM145402), the National Science Foundation, and Chan Zuckerberg Biohub. Stanford’s Department of Biomedical Data Science also supported the work.
Researchers from PHD Biosciences also contributed to this study.
# # #
About Stanford Medicine
Stanford Medicine is an integrated academic health system comprising the Stanford School of Medicine and adult and pediatric health care delivery systems. Together, they harness the full potential of biomedicine through collaborative research, education and clinical care for patients. For more information, please visit med.stanford.edu.
Journal
Science
Method of Research
Experimental study
Subject of Research
Not applicable
Article Title
The Virtual Biotech: A Multi-Agent AI Framework for Therapeutic Discovery and Development
Article Publication Date
17-Sep-2026
Introducing RADAR, a generalist AI tool for abdominal CT diagnosis
Summary author: Walter Beckwith
An artificial intelligence (AI) model called RADAR, designed to provide broad diagnostic interpretation of abdominal CT scans, substantially outperformed existing AI systems across a wide range of diseases and clinical settings, researchers report. The findings demonstrate the potential of vision-language models to advance expert-level generalist AI tools for radiology. The goal of AI in radiology is to create an expert-level generalist system capable of providing accurate, comprehensive diagnostic support across a wide range of real-world clinical situations. This, however, remains particularly challenging for contrast-enhanced abdominal computed tomography (CT), which often involves the evaluation of dozens of organs and the identification of hundreds of potential diseases, a task that requires physicians with specialized expertise and training. Although vision-language models offer a promising alternative by learning from associations between medical images and their accompanying text, abdominal CT presents unique challenges because relevant diagnostic information is sparse and highly dependent on complex anatomical context.
Here, Qi Zhang and colleagues present Rapid Abdominal Diagnosis with AI and Radiology (RADAR) – a vision-language AI model designed to provide broad diagnostic interpretation of contrast-enhanced abdominal CT scans. Trained on a dataset of 424,911 examinations, containing 1.5 million image-text pairs and more than 15 million anatomy-specific pairs, RADAR breaks CT volumes into individual anatomical structures and links those structures to corresponding descriptions in radiology reports, using a large-scale contrastive-learning approach to improve the connection between imaging findings and diagnoses. Throughout internal and external evaluations across multiple medical centers and varied clinical scenarios, Zhang et al. found that RADAR substantially outperformed existing specialist AI models, achieving a mean AUC (the average probability that a model can correctly separate positive and negative cases across evaluations) of 0.913 across 146 abdominal CT findings, compared with 0.776 for the best competing vision-language model. The model also performed well in challenging emergency settings, despite not being specifically trained on emergency data, achieving an AUC of 0.904 across more than 27,000 emergency CT cases. According to the authors, its performance remained strong across organs, diseases, and emergency cases, including uncommon findings. Moreover, in testing in cohorts at eight external centers, RADAR maintained high accuracy (AUC 0.895), demonstrating robust generalization across diverse patients, clinical settings, and imaging protocols.
Journal
Science
Article Title
An expert-level generalist AI for abdominal CT diagnosis
Article Publication Date
17-Sep-2026
Introducing the Virtual Biotech: a multi-agent AI for drug discovery
Summary author: Walter Beckwith
Researchers present the Virtual Biotech, a multi-agent AI system to inform drug-development decisions. By combining diverse biomedical and clinical evidence in a unified platform, the system may help identify therapeutic opportunities that would otherwise be missed. Drug discovery is a costly process – in both time and resources – yet about 90% of drug candidates entering clinical trials fail, usually due to inadequate efficacy or safety. Although researchers can now draw on vast amounts of genetic, genomic, molecular, and clinical evidence to improve these efforts, these data are fragmented across disciplines, making them difficult to integrate and interpret at the scale required for modern drug development. To help with this, Harrison Zhang and colleagues introduce the Virtual Biotech, a multi-agent AI platform designed to address bottlenecks by coordinating specialized AI “scientist” agents under the direction of a virtual chief scientific officer. According to the authors, unlike earlier AI systems that typically focus on individual biological analyses or rely on a single agent, the Virtual Biotech is designed to retrieve and analyze diverse biomedical and clinical evidence to support early-stage drug development decisions.
Zhang et al. tested the platform at three key points in drug development. Across 55,984 clinical trials, more than 37,000 AI agents found that drugs targeting genes specific to particular cell types were 48% more likely to reach the market and associated with 32% fewer adverse events. The system also combined diverse biological evidence to suggest a therapeutic strategy for lung cancer and analyzed a failed ulcerative colitis trial to identify possible reasons for its failure, demonstrating how human-guided AI teams could provide transparent, large-scale analysis to inform drug-development decisions. However, the authors note that Virtual Biotech’s conclusions are limited by the quality and breadth of available data, are not yet suitable for poorly studied diseases or targets, and still must be experimentally validated. “The Virtual Biotech illustrates a shift from isolated AI tools toward coordinated systems that reason across biological scales and stages of translation,” write the authors. “Rather than replacing scientists, agentic systems could expand the scope and speed of therapeutic hypothesis exploration while making the reasoning process more transparent and reproducible.”
Journal
Science
Article Title
The virtual biotech: A multi-agent AI framework for therapeutic discovery and development
No comments:
Post a Comment