Everyone—you, me, politicians, Wall Street, tech people, philosophers—is a little bit scared and dazzled by AI. This is natural. Nobody quite knows if it will kill humanity or cure cancer, destroy all jobs or create infinite prosperity, be a pivotal moment of planetary history or a glorified office assistant. All of the above? Finding out is the fun part!
Unfortunately, this very normal uncertainty about the technological future of AI has bled into some unnecessary political uncertainty about how to approach AI regulation. Some of this, of course, is the usual consequence of corporate influence-buying to insulate themselves from all restraint. But another factor is that the flashy unpredictability of AI seems to be convincing political leaders and a concerned public alike that our current tool set is not up to the task—that this exotic new industry will require an equally exotic new regulatory approach.
This is untrue. It is also damaging the public interest, because it causes a delay in getting some solid regulations on the books. That delay is not only wildly unsafe, but also allows AI companies to race ahead with little oversight, laying the groundwork for the classic “Well, we’ve already come this far, so we can’t go back” style of corporate regulation-dodging.
Perhaps the best example of this somewhat confused approach to AI oversight is Bernie Sanders’ proposal for the federal government to take a 50% stake in the biggest AI companies, and use the proceeds to create a sovereign wealth fund. I have written before about why this is not the best idea. I apologize for returning to the same topic again, but, shockingly, my last piece did not cause Bernie to drop his plan and rally to my side. It feels worth one more stab at clarity on this issue. If progressives genuinely want to get ahead of the dangerous consequences of unchecked AI—and I know that they do—it’s important for all of us to know that everything we need to do so is right there waiting for us.
In my lifetime, we have already witnessed the full arc of a powerful new industry growing up and taking over the world with insufficient regulation. That would be the tech industry. The failure of the US government and of organized labor to get ahead of the tech industry’s wild growth has produced: A crisis of economic inequality; an outright oligarchy with centibillionaire tech executives at the top of pyramid; a near-total absence of labor unions in the world’s richest industry; and massive psychological and societal damage as a result of unregulated social media and algorithms assaulting the overmatched human attention span. Collectively, it is safe to say that the way that the tech industry’s growth has played out this century is not something we want to repeat.
AI, for all of its novelty, is the latest iteration of the tech industry. It has the same explosive economic growth, creeping disruption of existing industries, and grandiose claims to being different from all that has come before. Now is the time for us to recognize that we have three tried and true tools at our disposal to prevent AI from becoming a socioeconomic rerun of the devastating example of the tech industry.
Taxes
Taxes. Yes! The government can levy taxes on companies. Are you afraid that unaccountable and soulless corporations are growing too powerful because they are making too much money? Tax them. This leaves them with less money. It leaves their potentially oligarchical owners with less money, too. Great.
Compare “raising corporate taxes” to “having the federal government take 50% of the equity of select large AI companies.” The first option has a number of advantages over the second. For one thing, taxes fall on all companies across the board, or of a certain size. There is no need to guess at which companies will be successful, or for the government to risk locking in the current dominance of a handful of companies by becoming shareholders in them. The companies can do what they do, and their position in the markets can rise and fall as it will. Don’t want ultra-powerful AI companies coming to dominate the economy? Tax their money away. Don’t want AI company CEOs becoming trillionaires? Tax their money away. Don’t want corrupt Trumpian administrations using their stakes in corporations as a slush fund? Don’t want to put the American public in the absurd position of saying, “Oh no, we must bail out OpenAI because we are all shareholders in it, and we want it to do well?” Don’t make the American public shareholders in these companies, then! Just tax them. And if you want to make a particular outcome like, for example, “automating all the customer service jobs” less economically attractive, you can tax that specific thing as well. It is a simpler, more direct, and more powerful way to accomplish what Bernie is trying to accomplish, without the potential pitfalls.
But, uh oh—what if taxes are not enough to stop these AI companies from doing bad horrible profitable things?
Laws
Laws. Yes! The government can pass laws that prevent the AI companies from doing things. Don’t want them releasing dangerous new models without government oversight? Don’t want them doing autonomous killer drones? Want to make sure there is disclosure if AI is used in media or entertainment? Want to keep AI slop out of public education? Pass a law. Taxes can change the economic incentive for companies to do things, and laws can stop them from doing things at all. Together, taxes and laws are what “regulation” means. I will not go on and on about this, because it is pretty straightforward. It is familiar. It is a regime that already exists.
When you realize that the federal government has the ability to levy taxes and pass laws, it becomes clear how misguided and unnecessary the idea of becoming an equity shareholder in AI companies is. If the government wants half of Anthropic’s money, it does not need to hold stock in the company; it can just tax it away. If the government wants to stop OpenAI from doing something bad, it does not need to go into the boardroom and say “you better listen to me because I’m a big shareholder!” No; it simply needs to pass a law. The federal government is not a person that needs to build up a big war chest of stock in order to throw its weight around with corporate America. The government is an entity charged with protecting the public good, that sits above corporations, and is empowered to take their money and set rules governing their conduct as it sees fit.
My argument is not that it is politically easy to regulate the AI industry; it is that we should pursue a method of regulation that is the best and most effective and has the least downsides. That method is taxes and laws. It is not “having the government go into the AI business.”
Unions
Unions. Yes! Labor unions. One woefully underdiscussed reason why the tech industry has created such profound economic inequality and has operated with such grotesque impunity is that virtually none of the major tech companies are unionized. This fact has allowed the company owners to hoover up bazillions in profits that would have otherwise been shared widely with hundreds of thousands of employees. It has allowed people like Mark Zuckerberg to pursue decisions that range from wasteful (the Metaverse) to morally horrific (enabling ethnic violence) with little effective internal pushback. The government clearly had no ability to or interest in preventing these and other bad tech company actions. Do you know what could have, though? A union.
Unions create a power center inside of companies that is separate from both the government and the management of those companies. Unions are not perfect, but they are the single most effective check on corporate power when they are strong. Had the tech industry been widely unionized, America’s economic inequality today would be drastically less, and the power of tech billionaires would be vastly reduced. Also, the labor movement itself, which has been weakening since the mid-20th century, would be significantly more influential both politically and economically, which would change our entire political landscape and could well have prevented the Trump era from happening altogether. It’s a big deal. I won’t go on about this forever, but there’s a book you can read that expands this argument.
Recognizing the potential of unions inside the major AI companies to mitigate some of the worst outcomes of the industry’s evolution, we can make reforms that make it easier to organize and sustain unions a bedrock part of our overall approach to AI regulation. If we really think it’s important, we can even pursue mandatory measures like reserving a significant chunk of board seats for worker representatives at these companies, as Germany does. It’s ridiculous for the federal government to be involved in boardroom agitation—but it is incredibly healthy for workers to do so.
AI… so new! So powerful! So unpredictable! However shall we regulate it, make it safe, channel it in the best interests of humanity? My friends, we can do it. In the most boring ways possible. In ways that already have centuries-long track records of being effective. We tax the companies. We pass laws to regulate their conduct. And we empower the workers at the companies to unionize and act as a check on the power of the cutthroat bosses. This stuff works. My plea—to Bernie, to progressives, to everyone with common sense who is nervous about how the AI industry is going to fuck us all up—is for us to focus on these powerful tools that we already have, rather than wasting more time brainstorming needlessly elaborate new tactics.
When Donald Trump and Xi Jinping walked side by side through the Temple of Heaven in May, the headlines went to Taiwan, to tariffs, to Nvidia’s chips. The thing that may prove to matter most barely registered. Only afterward, almost in passing, did mention it to reporters on Air Force One: he and Xi had talked about “possibly working together” on guardrails for artificial intelligence – and, he acknowledged when pressed, on the kinds of risks neither side may be able to contain once the technology is built.
That conversation, however elliptical, marks a genuine inflection point. Not because it produced an agreement, but because it revealed something that the “AI arms race” narrative has obscured: both Washington and Beijing are more uncertain about what they’ve built than about what the other side has built.
Beijing’s 2-trillion-yuan blueprint to build a nationwide AI infrastructure network, mandating 80 percent reliance on domestic suppliers like Huawei, looks at first like a declaration of technological total war – a calculated squeeze on Nvidia and AMD. The framing practically writes itself. Rival superpowers. Digital hegemony. A new Cold War fought in silicon rather than steel.
But the architecture of the plan tells a different story. The emphasis on domestic supply chains is not an offensive posture. It is a bunker. China is not building toward dominance so much as insulating itself against dependency – hedging against the possibility that the technology it is racing to develop might be turned against it, or might simply get away from it. That is a different kind of fear than we are used to hearing about.
The same anxiety is audible in Washington, if you know where to listen. Treasury Secretary Scott Bessent – whose portfolio is the financial system – has warned that frontier models could be turned to malicious ends by nonstate actors, and has pressed for a US-China protocol to guard against it. The Pentagon’s guidance on military AI was notable less for what it proposed than for what it insisted on preserving: human beings in the decision loop, especially where nuclear command and control is concerned. Chinese and American officials, it turns out, are unnerved by the same scenario – a crisis in which neither side is confident it can tell whether an autonomous system has made a decision, or simply made an error.
This convergence does not mean the two countries trust each other. They don’t. Chip export controls remain in place. Chinese labs are advancing rapidly on models designed to route around American restrictions. The competition is real, and in many domains it is intensifying.
But competition and restraint are not opposites, and the emerging reality is more complicated than either hawks or optimists tend to acknowledge. What is taking shape between Washington and Beijing is not d̩tente Рno treaty was signed in Beijing, and the first bilateral AI dialogue in Geneva in 2024 collapsed almost immediately under the weight of incompatible expectations. What is taking shape is something more provisional and more fragile: a shared reluctance to be the party that tips an already unstable system into genuine chaos.
Call it a Cold Peace. It is not a doctrine. It is not even an agreement. It is a posture – maintained not by trust but by mutual anxiety about what comes next if restraint breaks down.
The Cold Peace has already begun reshaping markets and industries, often invisibly. Venture capital firms in Silicon Valley are quietly recalibrating portfolios to hedge against regulatory slowdown, betting that governments will impose friction on frontier model deployment before the technology fully matures. In Beijing and Shenzhen, compliance reviews have lengthened development cycles for autonomous systems – not because regulators are opposed to the technology, but because no one is confident yet about who bears liability when an algorithm fails at scale. These are not dramatic reversals. They are adjustments, the kind that accumulate before anyone has named what is happening.
The rare earth ban on ten American companies announced in late June – targeting MP Materials and USA Rare Earth alongside US defense contractors – illustrates precisely this dynamic. Analysts called it symbolic retaliation for the Pentagon’s own blacklisting of Chinese tech giants like Alibaba and Baidu. But “symbolic” is doing a lot of work. The move came just five weeks after the Temple of Heaven summit, and it exposed the fragility of whatever goodwill was generated there. China controls over 90 percent of rare earth refining – the same materials that go into the data center hardware, autonomous systems, and defense electronics at the center of the AI competition. The Cold Peace does not prevent these moves. It only ensures they are calibrated carefully enough not to tip into something worse. That is a thin margin.
The deeper irony is that the AI race may be the first great-power competition in modern history where the rivals are more uncertain about the systems they’ve built than about the intentions of the other side. The nuclear age produced a similar reckoning, eventually – Eisenhower’s warning about the danger of a garrison state, Kennedy’s confrontation with the logic of mutual assured destruction. But that reckoning took decades and several near-catastrophes to arrive at. The AI reckoning is arriving faster, partly because the technology moves faster, and partly because its failure modes are less predictable than a missile trajectory.
The question now is whether fear is a durable enough foundation for stability. History suggests it is not – or not for long. The absence of escalation today is not a guarantee of equilibrium tomorrow. A single cyberattack convincingly attributed to an AI system, a market crash triggered by algorithmic mispricing, or a military incident involving autonomous weapons could unravel the current equilibrium with a speed that traditional diplomatic channels are not built to handle.
What would make this moment more durable is not a grand bargain – neither side is ready for that – but the slow accumulation of precedents: shared definitions of what constitutes a destabilizing AI capability, back-channel communication protocols for autonomous military incidents, agreed norms around AI in nuclear command systems. These are not romantic ambitions. They are the minimum infrastructure for managing a technology neither side fully understands.
Trump and Xi did not produce any of that in Beijing. But they did, apparently, produce something: Trump’s acknowledgment, however offhand, that the two might “work together” on AI guardrails – a concession that the machines they are building are outpacing their ability to govern them, and that some informal floor on recklessness is in both their interests. In the history of great-power competition, that is not nothing.
“People should not wake up to discover their face has become raw material for someone else’s AI experiment. This is another invasion of consumers’ privacy.”
Meta CEO Mark Zuckerberg shows a prototype of computer glasses that can display digital objects in transparent lenses at the Meta Connect developers conference in Menlo Park, California on September 25, 2024.
(Photo: Andrej Sokolow/picture alliance via Getty Images)
Brad Reed
Jul 08, 2026
COMMON DREAMS
Tech giant Meta on Tuesday introduced an artificial intelligence image generation model that critics say is a major potential risk to users’ personal privacy.
Meta, the parent company of social networks including Facebook and Instagram, described its new Muse Image model as a “creative partner that knows your world, making it easy to turn your ideas into high-quality visuals that you can download and share anywhere, including directly to your feed, story, or chat.”
In its announcement, Meta explained how users can either alter existing images or create new ones from scratch using AI prompts.
“You can describe what you want in simple, conversational language, and Meta AI handles the rest thanks to Muse Image,” the company said. “Ask it to mock up an image of you in front of a historical landmark, cleanly erase a photobomber from the background of a shot, or write a custom prompt to build a functional QR code.”
However, tech publication The Verge on Tuesday flagged a potentially troublesome feature that could compromise user privacy, noting that “users can... mention other Instagram accounts in Muse Image prompts,” which will let the AI model “incorporate their likeness into its output.”
According to a Tuesday report from Wired, the feature will let users snatch photos from any public Instagram and Facebook accounts unless those accounts’ owners specifically choose to opt out of the system.
What’s more, opting out of the system is not a simple one-click operation.
“If you want to avoid these AI generations of your Instagram posts without switching your account to private, you’ll have to dig into the app’s settings,” reported Wired. “Open the Instagram app, tap your profile, and then tap the three lines in the top-right corner of the screen. Then, scroll down to the Sharing and reuse tab. Here is where you should see a section labeled ‘Allow people to use your content on Instagram and with AI features on Meta,’ with a toggle for Posts and one for Reels.”
JB Branch, director of federal AI governance and technology policy at Public Citizen, blasted Meta for being careless with its users’ privacy by making them jump through hoops to stop others from swiping their photos.
“Meta has once again chosen the creepiest possible path,” said Branch. “People should not wake up to discover their face has become raw material for someone else’s AI experiment. This is another invasion of consumers’ privacy. Instead of asking for meaningful consent, Meta quietly defaults users into the system and buries the opt-out in account settings.”
Branch added that while Meta had a long history of violating user privacy, forcing them to opt out of its new AI image generation model “crosses what should be a bright line.”
“If our faces can be repurposed for AI simply because we posted a public photo, then very little remains off limits,” Branch emphasized. “Congress should establish clear privacy protections that require affirmative consent before companies can use a person’s image or likeness for AI products.”
The commission’s upcoming first meeting will focus on “strengthening AI infrastructure, accelerating AI’s impact on health, education, food security, and disaster response, and ensuring trust and safety,” said its CEO co-chair.
The CEO of the software company Salesforce, Marc Benioff, attends the 55th annual meeting of the World Economic Forum in Davos, Switzerland on January 23, 2025.
(Photo by Halil Sagirkaya/Anadolu via Getty Images)
Jessica Corbett
Jul 01, 2026
COMMON DREAMS
A week after United Nations Secretary-General António Guterres called on artificial intelligence companies to “come clean” about the full costs of power-sucking data centers, and as a UN panel on Wednesday released a report detailing the risks and impacts of AI, Axios revealed the creation of a related commission that’s full of Big Tech executives.
“The UN and its International Telecommunication Union (ITU) are convening the AI for Good Global Commission, which will hold its first meeting on July 8 in Geneva, Switzerland,” according to the outlet. It will be co-chaired by Salesforce CEO Marc Benioff and Rwandan President Paul Kagame, with other tech and policy leaders joining as members.
So far, Axios reported, they include ITU Secretary-General Doreen Bogdan-Martin, Estonian President Alar Karis, Amazon CEO Andy Jassy, Anthropic co-founder Jack Clark, Cohere co-founder Aidan Gomez, Microsoft president Brad Smith, Nvidia founder and CEO Jensen Huang, and AI and tech policymakers from Kazakhstan, Namibia, Nigeria, Saudi Arabia, and Singapore.
“AI is the most profound technological transition in history. And our values have to guide every step, because responsibility is the core of AI ethics,” Benioff said. The commission will bring together “the people who build AI, deploy it, shape policy, and represent communities.”
He added that “our inaugural meeting will focus on where this group is uniquely positioned to act together: strengthening AI infrastructure, accelerating AI’s impact on health, education, food security, and disaster response, and ensuring trust and safety.”
However, given recent polls showing that the public has limited confidence in large technology companies, opposes constructing data centers for artificial intelligence in their local area, is wary of AI’s impact on daily life, and has concerns about politicians having a “cozy relationship” with Big Tech, the commission may be met with skepticism or even backlash.
In the lead-up to the commission’s meeting next week at the ITU’s AI for Good Global Summit, the UN plans to hold the inaugural Global Dialogue on AI Governance, featuring a presentation of the “Preliminary Report of the Independent International Scientific Panel on Artificial Intelligence,” published Wednesday.
Established with a UN resolution last August, the panel is the first global scientific body on AI—and, as Guterres said in a statement about its new report, “the panel is intended to help the world separate fact from fakes, and science from slop.”
“We are looking to them to provide an authoritative reference point at a moment when reliable, unbiased understanding of AI has never been more critical,” the UN chief explained. “I am pleased to say that they have delivered a down payment on that commitment—in record time.”
The panel’s co-chair, Canadian computer scientist Yoshua Bengio, noted that “AI capabilities are outpacing both scientific understanding and governments’ ability to adapt. With growing evidence of deceptive AI behavior, science currently cannot guarantee that as capabilities continue to increase, AI will not cause catastrophic harm, either on its own or due to malicious users.”
“To act effectively, global policymakers must understand these systems,” he asserted. “This panel provides exactly that: a rigorous, shared scientific foundation to guide our collective way forward.”
The report discusses AI’s recent advances and expected trajectories; societal applications, from agriculture to education to healthcare; economic implications; security and environmental concerns; impacts on democracy, human rights, and information; potential harms to child safety and culture; and governance of the rapidly developing technology.
“The technology is transformative, but if the world keeps moving along this trajectory, humanity will fail to realize the gains it promises. The risks—to societies, to security, and to our species—are too high, and the forces driving AI forward are not the forces that will deliver its benefits,” said Maria Ressa, a panel’s co-chair and Nobel Peace Prize-winning Filipino-American journalist.
Guterres, whose term ends this year, similarly stressed the need for urgent action on a global scale. He said that the “single lesson” he wanted to highlight from the multifaceted report is that “the more AI advances without shared rules, the less say governments and people will have in the outcome. So my message to governments is simple: Do not wait.”
“Next week in Geneva, the first Global Dialogue on AI Governance will begin to turn science into shared action—with every nation at the same table,” he said. “I look forward to joining member states there to help carry this work forward. And soon, I will set out proposals to help countries build the capacity to adequately deal with this technology—and share in its rewards.”
Guterres’ Wednesday comments came after he publicly took aim at artificial intelligence companies last week, proposing the AI Environmental Transparency Initiative during London Climate Action Week, as the second heatwave in as many months scorched the United Kingdom and various other European countries, killing at least hundreds of people.
“I am calling on every major AI company to measure and publicly disclose the full environmental impact of its systems—carbon, water, and land footprints—and to commit to power every data center with renewable energy by 2030,” he declared. “No more hidden costs. No more shifting the burden onto those least able to bear it.”
AI system developed by UC Irvine physicists helps explain why neutrinos have mass
Innovation allows researchers to explore uncharted areas of particle physics theory
Irvine, Calif., July 9, 2026 — Physicists at the University of California, Irvine have developed an artificial intelligence system that can autonomously design theoretical physics models, a task traditionally carried out by human theorists. The approach allows researchers to explore large, uncharted areas of particle physics theory, helping identify promising new explanations for the behavior of neutrinos.
The system is called Autonomous Model Builder, or AMBer, and was developed by a research team led by UC Irvine doctoral candidates Victoria Knapp- Pérez and Jake Rudolph in the Department of Physics and Astronomy. The work is described in a study published in Nature Communications Physics.
AMBer uses reinforcement learning, a form of artificial intelligence that learns through trial and error rather than by following predefined instructions. As it explores possible particle physics theories, the system evaluates its own choices and improves over time.
“Reinforcement learning is different from other kinds of machine learning, in which models predict labels or find patterns in data,” Rudolph said. “AMBer’s RL framework allows it to learn about the space of theoretical models as it explores, effectively creating its own training data as it searches for promising models.”
The system constructs particle physics models by selecting mathematical symmetry groups, determining which particles to include and assigning how those particles behave under the chosen symmetries. Each proposed model is evaluated based on how well it matches experimental data while minimizing the number of adjustable parameters, a key measure of a theory’s predictive power.
The researchers tested AMBer on well-studied classes of neutrino theories and demonstrated that it could reproduce known results. They then applied the system to previously unexplored mathematical frameworks, identifying new candidate models that may merit further investigation.
Neutrinos are subatomic particles with extremely small but non-zero mass – a property not explained by the Standard Model of particle physics. Developing theories that explain neutrino mass remains one of the field’s major challenges.
The researchers emphasized that the system is designed to assist, not replace, human physicists by narrowing vast theory spaces down to the most promising candidates.
“AMBer functions as a filter, giving human physicists a better-informed starting point from which to study more complex behavior of neutrino models,” Knapp-Perez said.
Additional collaborators include Max Fieg, a former UC Irvine doctoral student now a postdoctoral fellow at Fermilab; Aishik Ghosh, a former UC Irvine postdoctoral scholar now a professor at the Georgia Institute of Technology; and Daniel Whiteson, UC Irvine professor of physics, who supervised the research. Jason Baretz, UC Irvine doctoral student in Whiteson’s group, also contributed to the research.
This research used computing resources from the National Energy Research Scientific Computing Center. Funding was provided in part by the National Science Foundation, UC-MEXUS-CONACyT and the Department of Energy’s Office of High Energy Physics.
About the University of California, Irvine: Founded in 1965, UC Irvine is a member of the prestigious Association of American Universities and is ranked among the nation’s top 10 public universities by U.S. News & World Report. The campus has produced five Nobel laureates and is known for its academic achievement, premier research, innovation and anteater mascot. Led by Chancellor Howard Gillman, UC Irvine has more than 36,000 students and offers 224 degree programs. It’s located in one of the world’s safest and most economically vibrant communities and is Orange County’s second-largest employer, contributing $7 billion annually to the local economy and $8 billion statewide. For more on UC Irvine, visit www.uci.edu.
Media access: Radio programs/stations may, for a fee, use an on-campus studio with a Comrex IP audio codec to interview UC Irvine faculty and experts, subject to availability and university approval. For more UC Irvine news, visit news.uci.edu. Additional resources for journalists may be found at https://news.uci.edu/media-resources.
Journal
Communications Physics
Method of Research
Computational simulation/modeling
Subject of Research
Not applicable
Article Title
Towards AI-assisted neutrino flavor theory design
Meet Biomni – an AI-powered biomedical co-scientist
Researchers at Stanford University have announced the debut of Biomni – an AI-powered multi-skilled biomedical research agent. Biomni is no mere chatbot. It is a full-fledged “co-scientist” capable of designing and developing complex research workflows, said Jure Leskovec, the Alfred and Rebecca Lin Professor and professor of computer science in the School of Engineering and senior author of the paper introducing Biomni in the journal Science.
“If you think of an agent as a carpenter, a carpenter without tools is just a carpenter who can talk,” Leskovec said, explaining what sets Biomni apart from popular generative AI chatbots. “With Biomni, we give the carpenter a set of tools, so it can build.”
Born for impact
Biomni was born from the notion that, when working with an AI agent, a scientist should be able to describe a research problem in simple, natural language. With that in mind, the researchers designed Biomni to read the literature, form hypotheses, choose datasets and
tools, write code, interpret results, and suggest next-stage experiments in a complete research workflow.
“Biomni is able to understand a simple question like, ‘Why are these patients responding differently to the drug?’” explained Kexin Huang, a former doctoral student in Leskovec’s lab who recently earned his PhD and now heads a startup to bring this technology to market. “Then it digs in, doing a lot of the scientific legwork.”
The researchers have chosen biomedical sciences for its potential to improve the lives of everyday people. From basic understanding of life to new cures for myriad diseases, scientific breakthroughs in biomedical research cannot come fast enough.
In a real-world example, one Biomni user uploaded more than 450 files of continuous glucose monitoring, food intake, and physical activity data and asked a simple question: “Analyze this data, find interesting and plausible hypotheses.” In just 40 minutes, Biomni cleaned and unified the data, generated visualizations, and identified patterns relating food intake and body temperature. Leskovec estimates that work would have taken 60 or more hours for a human to complete.
Biomni offers one more advantage the chatbots can’t claim: It provides full citations and tracking of its work. In its traceability, the researchers argue, Biomni makes the science more rigorous and more reproducible.
Innovation apace
Biomni is specifically trained in biomedical sciences. It incorporates the breadth of full-text, publicly available papers, code, and data stored on bioRxiv, a service for prepublishing early versions of promising scientific findings, to identify common software, tools, and databases that are used in biomedical research. Biomni layers in 150 specialized biomedical tools, 105 software packages, and 59 databases spanning all 25 biomedical subdomains defined by bioRxiv, ranging from genetics to neurology.
Biomni speeds the process of scientific ideation and innovation. Leskovec explained there is an inverse relationship between scientific information and the pace of discovery. As the volume of knowledge, data, and tools has grown, innovation has slowed.
The reason for that slowdown is simple. Behind every breakthrough lies years of study that all begin with a hypothesis. Even just developing a hypothesis requires substantial investment from scientists – reading literature, ingesting and homogenizing datasets, writing code, and looking for unexplored patterns that then become the basis for groundbreaking work. This process can take weeks or even months.
“The hurdle in biomedical science is not intelligence or ideas; it is mechanics,” Leskovec emphasized. “It’s this laborious stuff that slows innovation. Biomni can do this work in minutes.”
Human in the loop
Leskovec and Huang are quick to point out that Biomni will not replace humans, but it frees them to concentrate on the value of the scientist – ideation and judgment. While Biomni can synthesize vast amounts of information and data very quickly and is adept at pattern recognition, the choice to pursue a scientific trajectory demands human experience and reasoning.
“And it always will,” said Huang. “This is not about machines taking over science, but more about machines becoming a powerful new partner to augment human researchers. With Biomni, scientists have a fast and tireless collaborator that empowers them to focus on the important work of science.”
A prototype Biomni is already in use by more than 10,000 labs in academia and industry, making it the most widely used AI co-scientist system in biomedicine.
“Biomni is my first research project that has gained wide use by real biologists,” Huang said. “To have that impact on how biologists are doing their work has been rewarding. I look forward to seeing where Biomni goes from here.”
Contributing authors include: graduate students Serena Zhang, Ryan Li and Gavin Li; postdoctoral scholars Hanchen Wang, Yusuf Roohani, Yuanhao Qu, Junze Zhang, Xin Zhou, Yingzhou Lu and Di Yin; visiting Professor Xin Zhou; Michael Snyder, the Stanford W. Ascherman Professor of Genetics in the School of Medicine (Stanford Medicine); Le Cong, associate professor of pathology and of genetics in Stanford Medicine; research engineer Shruti Marwaha; genetic counselor Jennefer N. Carter; professor Matthew Wheeler, associate professor of medicine (cardiovascular medicine) in Stanford Medicine; and professor Jonathan A. Bernstein, professor of pediatrics (genetics) in Stanford Medicine. Researchers at the University of Washington, the Arc Institute, Genentech, Princeton University, and the University of San Francisco also contributed.
Bernstein is also a member of Stanford Bio-X, the Maternal & Child Health Research Institute (MCHRI), and the Stanford Medicine Children’s Health Center for IBD and Celiac Disease, and a faculty affiliate of the Institute for Human-Centered Artificial Intelligence (HAI). Le Cong is also a member of Stanford Bio-X, the Cardiovascular Institute, the Stanford Medicine Children’s Health Center for IBD and Celiac Disease, the Stanford Cancer Institute, and the Wu Tsai Neurosciences Institute. Leskovec is also a member of Bio-X and the Wu Tsai Neurosciences Institute, and a faculty affiliate of HAI. Snyder is also a member of Stanford Bio-X, the Cardiovascular Institute, the Wu Tsai Human Performance Alliance, the MCHRI, the Stanford Medicine Children’s Health Center for IBD and Celiac Disease, the Stanford Cancer Institute, and the Wu Tsai Neurosciences Institute. Wheeler is also a member of Bio-X, the Cardiovascular Institute, the Wu Tsai Human Performance Alliance, MCHRI, and the Wu Tsai Neurosciences Institute.
Funding was provided by the National Science Foundation, Stanford Data Science Applications, Wu Tsai Neurosciences Institute, Stanford Institute for Human-Centered AI, Chan Zuckerberg Initiative, Amazon, Genentech, GSK, Hitachi, and SAP.
Journal
Science
Article Title
Autonomous biomedical research with an artificial intelligence agent
Article Publication Date
9-Jul-2026
Doctors may have trouble learning from experience that contradicts AI advice
New study highlights potential challenges for using automated tools in healthcare
PLOS
In experiments in which physicians made decisions about treating hypothetical patients, the physicians tended to trust incorrect advice presented as being generated by artificial intelligence (AI), even after given the opportunity to notice that patient recovery data contradicted the recommendations. Aranzazu Vinas of the University of the Basque Country, Spain, and colleagues present these findings in the open-access journal PLOS Digital Health.
AI systems can help physicians categorize patients according to their different care needs, such as whether a patient is more or less likely to benefit from a certain treatment. Since these systems are not perfect, they are meant to be used as suggestions, with potential errors caught and corrected by physicians.
Prior research has shown that, in general, people struggle to notice and correct mistakes made by AI. To explore how this challenge may extend to physicians, Vinas and colleagues analyzed data from 223 physicians who anonymously participated in online experiments.
The physicians were asked to imagine they had the option to treat patients for a rare disease using a not-yet-proven treatment still under development. They were told that an AI system had identified which patients were more or less likely to benefit from the treatment. The physicians then chose which patients to treat, and after being presented with data on patient recovery, rated their perceptions of how reliable the AI was.
Crucially, the actual effectiveness of the hypothetical treatment did not align with the AI recommendations. In one experiment, the treatment was equally moderately effective for all patients, and in a second experiment, it was equally ineffective for all.
However, in both experiments, the physicians tended to rate the AI system as reliable and apparently did not use the patient recovery data to conclude that the AI recommendations were incorrect. In the second experiment, the physicians did not realize that the treatment was entirely ineffective.
These findings highlight potential challenges for incorporating AI-based classification into healthcare. Future research could build on this study, such as by developing and testing strategies and protocols that could increase human critical thinking and detection of AI errors, in order to maximize the benefits of the human-AI collaboration while minimizing potential errors."
Lead author Aranzazu Vinas notes: " In both experiments, physicians mostly trusted the AI's classifications and had trouble learning from the feedback. Furthermore, in the second experiment, professionals did not notice that the treatment was completely ineffective."
Co-author Helena Matute adds, "People tend to say that there is always a human controlling the algorithm, but our experiments show that doctors (as well as anyone else) have problems in learning from the available evidence when it contradicts the suggestions of an algorithm."
Co-author Fernando Blanco summarizes: "It is important to investigate the errors that humans (including doctors) make when working with algorithms, in order to learn how to minimize the problems that arise from them."
In your coverage please use this URL to provide access to the freely available article in PLOS Digital Health: https://plos.io/4blGKHA
Citation: Vinas A, Blanco F, Matute H (2026) Doctors vs. Algorithms: Physicians, too, struggle to learn from evidence that contradicts AI suggestions. PLOS Digit Health 5(7): e0001490. https://doi.org/10.1371/journal.pdig.0001490
Author Countries: Spain
Funding: Support for this research was provided by Grant PID2021-126320NB-I00 funded by MICIU/AEI/10.13039/501100011033 and by ERDF A way of making Europe, as well asGrant IT1696-22 funded by the Basque Government. A.V. was supported by Fellowship FPU20/01009 funded by MICIU. The funders had no role in the study design, data collection, and analysis, decision to publish, or preparation of the manuscript.
Journal
PLOS Digital Health
Method of Research
Experimental study
Subject of Research
People
Article Publication Date
9-Jul-2026
JMIR Report: Investigating AI-based personal training
image:
Anna Zucker, JMIR Correspondent.
view moreCredit: Image provided by the author.
(Toronto, July 9 2026) JMIR Publications released a feature News and Perspectives story on AI fitness advice in its News and Perspectives section. In “Should AI Be Your Personal Trainer?”, JMIR Correspondent Anna Zucker covers the growing use of AI chatbots for personal fitness programs.
The Pros of Using AI for Personal Training…
Personal training can support and motivate physical activity, benefiting clients’ health—but it’s not cheap. Free or low-cost AI chatbots like ChatGPT, which use data from scholarly sources and online content created by human fitness professionals, can be used to generate training programs, “help[ing] democratize fitness knowledge, especially for people in resource-limited settings,” writes Zucker. She reports on two studies which compared ChatGPT and human fitness professionals when, in one study, answering fitness questions; and in the other, creating a personalized fitness program. Both studies found ChatGPT to match or outperform the fitness professionals.
…And The Cons
“Access to fitness knowledge,” writes Zucker, “may not be the same as access to fitness coaching.” The degree of nuance required in helping clients stay motivated, work around injuries, and safely push their limits requires continuous monitoring and communication—not to mention visual and tactical judgment—according to JMIR’s sources, which includes multiple runners; physical therapist and owner of ReMove Rehab and Performance Katy Vieira; and personal trainer and rehab specialist Steven Kane.
Striking A Balance
Can we have it both ways? Zucker reports on a recent validation study of MediaPipe Pose, an AI model which can analyze real-time smartphone video to estimate body position and movement. Though the study found that the model achieved 97.2% accuracy in estimating human movement, its authors cautioned that it can’t fully replace human professionals’ comprehensive clinical judgement and support, instead recommending a hybrid model using “AI for daily guidance and certified professionals for regular check-ins,” writes Zucker. This model—combining the low cost of AI tools with the professional expertise of human trainers—could potentially allow for a wider reach of quality, personalized fitness advice.
Please cite as:
Zucker A. Should AI Be Your Personal Trainer?
J Med Internet Res 2026;28:e106128
URL: https://www.jmir.org/2026/1/e106128
doi: 10.2196/106128
About JMIR Publications News and Perspectives
JMIR Publications is a leading open access publisher of digital health research. The News and Perspectives section is the newest addition to its portfolio, established to bring the rigor and integrity of academic publishing to scientific journalism. The section features well-researched, expert-driven content from the Scientific News Editor, Kayleigh-Ann Clegg, PhD, and a network of specialist JMIR Publications Correspondents to keep the digital health community informed, inspired, and ahead of the curve.
About JMIR Publications
JMIR Publications is a leading open access publisher of digital health research and a champion of open science. With a focus on author advocacy and research amplification, JMIR Publications partners with researchers to advance their careers and maximize the impact of their work. As a technology organization with publishing at its core, we provide innovative tools and resources that go beyond traditional publishing, supporting researchers at every step of the dissemination process. Our portfolio features a range of peer-reviewed journals, including the renowned Journal of Medical Internet Research.
To find out more about JMIR Publications, visit jmirpublications.com or connect with them on Bluesky, X, LinkedIn, YouTube, Facebook, and Instagram.
Media Contact:
Dennis O’Brien, Vice President, Communications & Partnerships
JMIR Publications
communications@jmir.org
+1 416-583-2040
The content of this communication is licensed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, published by JMIR Publications, is properly cited.
Journal
Journal of Medical Internet Research
Method of Research
Commentary/editorial
Subject of Research
People
Article Title
Should AI Be Your Personal Trainer?
Article Publication Date
8-Jul-2026
Researchers hone new AI method to track “smart” vapes with digital screens
New approach helps researchers track emerging e-cigarette products and marketplace trends
Georgia Institute of Technology
E-cigarettes, also known as vapes, are battery-operated devices that heat a liquid that typically contains nicotine, an addictive substance. These devices are continually changing, with new flavors, novel device designs, and digital screens. Some of these e-cigarettes — sometimes called “smart vapes”— include built-in games and Bluetooth connectivity that have the potential to gamify the use of nicotine. Many of these devices are marketed online but cannot be easily monitored with existing data sources and methods.
A new study published July 9 in the journal Nicotine and Tobacco Research demonstrates how artificial intelligence (AI) can be used to automatically detect and classify new e-cigarette devices with screens. The study, led by Georgia Tech Research Institute (GTRI) scientists, in collaboration with the CDC Foundation, analyzed publicly available product images from online tobacco retailers.
“Monitoring online e-cigarette marketing is like a game of Whack-A-Mole, with so many new products and features popping up,” said Kristy Marynak, PhD, Senior Director for Tobacco Control Initiatives at the CDC Foundation and a study author. “This study shows how machine learning techniques can shed light on the online e-cigarette marketplace and the vast quantities and types of e-cigarette products available.”
“Smart” Vapes Attract Youth and Young Adults
Vapes with digital screens are appealing to young people, underscoring the importance of monitoring emerging e-cigarette technologies and product features. According to a CDC Foundation study of a nationally representative cohort of youth and young adults, nearly a third of youth and young adults who use e-cigarettes use “smart” vapes.
AI Helps Classify E-Cigarettes More Efficiently
The GTRI team studied images containing e-cigarettes and other related tobacco products from an open-source dataset and augmented them with images obtained from five online sites selling e-cigarettes. An AI-based object detection model was trained on approximately 7,000 of those images and tested with 3,920 additional images to ensure accuracy. In total, 2,401 images were predicted by the object detection model to contain an e-cigarette.
The researchers then used a vision-language model (VLM), a type of AI that combines large language models with computer vision to process both images and text simultaneously. The VLM analyzed the 2,401 images along with the text descriptions of e-cigarette devices to automatically determine if screens were present. Results were found to be more than 90% accurate.
“There are thousands of e-cigarette devices, and we have currently identified more than 60 websites selling them,” said Charity Hilton, a GTRI research scientist who leads the overall project. “Using AI techniques such as natural language processing, machine learning, and large language models, we’re now able to classify these products much more efficiently and repeatably.”
New Tool Will Provide Real-time Data for Public Health
Using what they’ve learned, GTRI and CDC Foundation researchers now plan to incorporate the AI-based process into a tool that will complement existing techniques as part of the CDC Foundation’s e-cigarette monitoring efforts.
“This tool gives the CDC Foundation a force multiplier to look at a vast swath of new products and keep up with the market trends and changing environment,” said Dr. Hunter Morera, a GTRI researcher and the study’s lead author. “With thousands of products going up monthly, traditional manual coding methods simply can’t keep up.”
The use of AI to capture and rapidly analyze e-cigarette data has the potential to transform how the tobacco product landscape and trends are monitored and understood. This will allow for more informed public health research, surveillance and decision-making.
“This study demonstrates that AI can systematically and efficiently identify novel features of emerging tobacco products, in this case the presence of screens on thousands of e-cigarette devices,” said Elisha Crane, MPH, a public health data scientist at the CDC Foundation and the study’s senior author. “This work serves as a case study of how these methods can be applied to enhance existing tobacco product monitoring and has the potential to provide real-time data to inform public health officials, policymakers, and regulatory agencies.”
Hilton hopes this study will help open the door for other AI applications in public health monitoring. “There’s definitely a lot of AI work going on, but it’s not necessarily being applied to public health issues,” she said. “We’d like to support public health agencies by applying a cutting-edge technology to the critical challenges they are addressing.”
Additional study authors include James Jun and Dianna King of GTRI, and Elizabeth Seaman Jones, PhD, from the CDC Foundation.
Method of Research
Imaging analysis
Subject of Research
Not applicable
Article Title
Automatic detection of e-cigarette screens using object detection and vision language models.
Article Publication Date
9-Jul-2026
