Daniel Hook, chief scientific officer at Holtzbrinck Group, argues that AI chatbots are pulling researchers away from one another, and that the cure belongs with funders and institutions rather than with individual scientists. He made the case in an essay for Research Agenda, published on 7 September 2026, so the argument has been circulating for almost four weeks and is not new today. It is an opinion piece with no new data of its own.
The idea starts in a driverless Waymo taxi in San Francisco. Hook and Susan Winslow, CEO of Macmillan Learning, both loved the ride because nobody expected small talk. From that, Hook builds what he calls the Waymo effect: a technology strips out the hassle of dealing with another person, and we register only the gain, because the cost of the hassle was always visible while its benefit never was. The taxi driver, he notes, may be the last stranger who gives you a view you did not ask for.
Hook says language models now fill that gap in research. A human collaborator is busy, brings an agenda, and may tell you the premise of your whole project is flawed. A chatbot is awake at 2am on a Sunday and wants nothing from you. Ask it for a critique and it delivers one, but only of the argument as you framed it. In Hook’s telling, the collaborator’s inconvenience is not a flaw in the collaboration. It is most of the value. He proposes a name for the slow loss of that habit: “decollaboration.”
The sharper part of the essay is about incentives. Hook points to three forces pushing scientists toward working alone. Tight budgets cut travel, workshops and visiting posts first. Evaluation systems still reward the speed of output, which is exactly what chatbots promise. And a model never fights over author order, which Hook calls arbitrage in a system where credit is survival.
He leans on research about research to back this up, though he cites it in general terms rather than by study. Small teams, he says, tend to disrupt while large ones develop, and unusual pairings across fields produce the most novel work. Early evidence on generative AI, he adds, shows individual productivity rising while the range of ideas narrows. Readers should treat those as his summary of the literature, not as findings he tested.
The essay also makes a case that writing is thinking. Hook invokes the psychologist Robert Bjork’s “desirable difficulties” and argues that handing off the drafting of a paper skips the step where weak arguments get found. He adds Lisanne Bainbridge’s 1983 work on the ironies of automation, whose point is that a system which rarely fails leaves the person supervising it with little chance to keep their skills sharp.
Hook does not call for a ban. He cites a March 2026 Nature comment by Dashun Wang, who describes AI agents as aeroplanes for the mind and urges “pilot-in-command science,” with the researcher keeping authority over the question and the conclusions. Wang also warns that outputs converge unless people deliberately build in dissent. Hook’s reading is that even the optimists want a human in the front seat.
His prescription is to fund the friction: workshops, visits, co-location and unstructured time, with evaluation based on contribution rather than velocity. That is a hard sell, since each of those items is a line a finance office can cut without anyone noticing the same quarter.
The weak point is evidence. Hook offers no count of how many researchers have swapped colleagues for chatbots, and his opening scene is a personal anecdote about a taxi. The market angle is also awkward: vendors now pitch AI as a tireless research partner, which is the frictionless colleague Hook is worried about, and a lab that buys seats for it is spending from the same pot as the workshop budget.
For a lab director setting next year’s budget, the question is concrete: if the tool licenses are approved, who is still in the room to tell the team it is solving the wrong problem?
Reported by Research Agenda, in an essay by Daniel Hook published on 7 September 2026, nearly four weeks before this summary.
