This recording is limited to attendees of the event, but you can read a recap and watch clips below.
Setting the stage
Ward traced his interest in AI back to a 2016 dinner in San Francisco, where two newly minted psychologists with deep expertise in addiction described an experiment in which former cocaine users, placed in a simulated nightclub and shown a mirror of baking soda, rated their craving at “11 out of 10” even though they knew exactly what it was. The brain takes its cues from its environment, and the psychologists’ pitch was to offer that insight to app makers through a consultancy called Dopamine. Ward realized then that his beat was no longer just technology but the intersection of technology, behavioral science, and the companies applying one to the other.Drawing on his PBS series Hacking Your Mind and his book The Loop, Ward explained the difference between our fast, automatic “system one” thinking and our slower, more deliberate “system two.” The promises AI companies are making to science, including OpenAI’s stated goal of an AI research intern this year, speak directly to the first, and as Ward put it, “We love this idea, unconsciously, that we will be absolved of all tedious work.”
A narrower songbook
Ward questioned the idea that AI gives people creative abilities they wouldn’t otherwise have, comparing generative AI to a tribute band playing a hotel ballroom. “AI is just that but from a much bigger songbook,” he said, adding that its medleys are harder to recognize only because we don’t know all the songs. He pointed to “Artificial Hivemind,” an award-winning NeurIPS 2025 paper in which researchers gave about 70 models the same open-ended creative prompts and found that the responses converged toward the same narrow middle, both within each model and across all of them.That narrowing is already showing up in research, as Ward illustrated with a Nature paper published this year that analyzed 41.3 million papers. Scientists who used AI published three times as many papers, received 4.8 times the citations, and became project leaders more than a year earlier, yet they worked across fewer topics, and engagement between scientists fell by 22%. A separate study of 2.1 million preprints found that output rose between 23% and 90% after authors adopted LLMs, while polished prose stopped being a reliable signal of a paper’s quality, a finding with obvious implications for editors and reviewers.
Confidence without competence
The research Ward found most concerning deals with how people judge their own work when AI is involved. In one study, endoscopists who used AI for several months and then lost access to it saw their adenoma detection rate drop from 28% to 22%. In another, medical students answering difficult diagnostic questions scored 21% on their own and 22.6% with an expert-checked AI explanation, but only 9.2% when given a convincing explanation that was wrong, and they felt just as sure of those wrong answers.Ward was clear that he isn’t arguing against the technology. “I believe that foundational science deserves every token it can be given,” he said, but realizing that potential means understanding how easily we misjudge these tools. His example was an AI model credited with solving a long-standing Erdős problem, which Fields Medalists initially celebrated. It turned out the model had surfaced an overlooked solution from the existing literature without crediting anyone, prompting mathematicians to warn of “severe attribution and plagiarism problems.” Finding that paper was a real service, Ward said, but “it’s not the same thing when the librarian says, oh, I know the answer to that.”
Why friction matters
Ward argued that much of today’s AI is designed to replace human experience rather than extend it, with the industry focused on reducing friction rather than giving credit or supporting learning. He compared industry leaders’ resistance to asking creators for permission to General Motors’ 1972 claim that requiring catalytic converters could halt production, noting that every automaker installed them anyway. He also noted that LibGen, a pirated dataset heavy with academic books, was approved for AI training, and he cited Silverchair data, shared via The Scholarly Kitchen, showing that half of platform traffic in the first ten weeks of 2026 came from bots, crawlers, and automated systems.As a counterweight, Ward shared an idea from former California Supreme Court Justice Mariano-Florentino Cuéllar: important legal decisions are deliberately made difficult so they engage our most active thinking. A guilty plea could easily be entered with a swipe on a phone, but requiring someone to stand before a judge and say it aloud creates the conditions for real reflection. Good judgment, Ward said, is like a concert pianist’s skill, a muscle that has to stay in use to stay strong.
That idea has real consequences for our workforce, since many people in the room started in entry-level roles such as copy editing before going on to lead journal programs. The current vision for AI leaves little room for that kind of formative work or the institutional memory it builds, and Ward pointed to Brookings research by Molly Kinder showing that administrative and clerical roles are among the first exposed.
An uncontaminated record
Ward sees our community as holding something especially valuable, and he compared this moment to the years after 1945, when nuclear testing contaminated newly produced steel and scientists had to salvage older metal for sensitive instruments. In his view, anything written since late 2021 carries a similar risk of contamination, and publishers are the stewards of a corpus that predates it. The Erdős solution, after all, was found in that record, and in Ward’s words, “There’s something to be protected there.”Looking ahead
In the Q&A, Ward suggested that AI leaders’ public warnings about the technology serve several purposes at once, from marketing to regulatory advantage, and cautioned against letting companies talk about their products’ behavior “as if it’s the weather.” He said he relies on AI every day to run his independent outlet, and that his most useful safeguard is understanding how predictable we all are. “If a company can see your behavior, it’s going to try and predict your behavior. And if it can predict your behavior, it’s going to try and shape your behavior,” he said. He also expressed hope that, as with tobacco and auto emissions, litigation and moral frameworks will move us toward the right things, even if it takes time.Ward’s message to our community was that the values scholarly publishing stands for, including attribution, peer review, and careful judgment, won’t be protected by the market alone. “Friction is actually extremely valuable,” he said, and for an industry built on verification and credit, that is a strength worth carrying forward as we decide how AI fits into the future of research.
View the Platform Strategies 2026 session recordings here.