The Waymo effect: how AI is quietly making research less collaborative

How frictionless technologies thatch america to for illustration our ain institution – and why investigation leaders should worry.

On a caller travel to San Francisco I did the point that each visitant to San Francisco now does: I summoned a car pinch nary 1 successful it.

The Waymo arrived pinch the serene assurance of a instrumentality that has ne'er erstwhile worried astir wherever to find parking, and I climbed into the backmost seat, glanced instinctively astatine the driver’s spot to opportunity hello, and recovered myself nodding politely astatine an quiet chair. The steering instrumentality turned itself. I person spent a reasonable information of my life reasoning astir counterintuitive aspects of physics, but that did not thief maine pinch the mild existential vertigo of watching a steering instrumentality moving connected its ain – an unseen driver responsible for my safety.

I was successful San Francisco, successful part, to walk clip pinch Susan Winslow, CEO of Macmillan Learning. We are colleagues wrong the Holtzbrinck group, and we had travel together for the astir quality of master reasons: to collaborate. To beryllium successful the aforesaid room, comparison notes connected really AI is reshaping our respective corners of investigation and education, and do the benignant of reasoning that is stubbornly difficult to do complete video calls.

And yet, comparing notes connected our Waymo experiences, we discovered we agreed connected thing other entirely: the rides were wonderful. As 2 self-confessed introverts, we had each recovered the driverless car to beryllium a mini oasis. No responsibility to make conversation. No silent speech complete the radio. A guilt-free abstraction to beryllium unsocial pinch one's thoughts, decorativeness an email, aliases return a telephone en way without the awkwardness astir conducting it successful beforehand of a stranger. The car was quiet, soft and wholly undemanding.

It took america somewhat longer to sanction what we had lost. Two group who had crossed continents to talk to each different were softly delighted by a exertion whose cardinal characteristic is that you don’t person to talk to anyone.

Naming the Waymo effect

Let maine effort a definition: the Waymo effect is what happens erstwhile a exertion removes the clash of dealing pinch different quality being, and we acquisition that removal arsenic axenic summation – because the costs of the clash were ever visible to us, while its benefits were not.

This is simply a acquainted move for anyone who has publication Tim Wu connected “the tyranny of convenience” – his statement that erstwhile a frictionless action exists, we return it by default, and softly surrender immoderate the clash was doing for us, since cipher advertised it arsenic valuable successful the first place. Albert Borgmann’s related thought of the device paradigm goes a measurement further: a instrumentality delivers a commodity – warmth, information, companionship – while concealing the believe that erstwhile had to beryllium undertaken to gain it, truthful that we extremity noticing the believe has gone astatine all. Only the commodity keeps arriving.

The costs of talking to a taxi driver are obvious: the mini effort of politeness, the conversational roulette, the introvert’s taxation of sustained mini talk astatine 7am. The benefits are diffuse and deferred: the driver was, for galore of america connected galore days, the past alien we were obliged to encounter. The past personification from extracurricular our bubble – professionally, politically, socially – pinch whom we had an unchosen conversation. The past reliable root of a position we did not inquire for.

Neither Susan nor I would creation a world without those conversations. We simply enjoyed opting retired of this one. And that, of course, is really specified things are lost: ne'er by decision, ever by convenience. One comfortable thrust astatine a time.

I should beryllium clear that this is not an anti-Waymo article (the rides really were excellent, and I would return 1 again without hesitation – that is alternatively the point). It is an article astir research. Because the aforesaid logic that removed the driver from the car is now, pinch adjacent serenity and considerably greater consequence, removing the collaborator from the investigation process.

The frictionless colleague

Large connection models are the Waymo of intelligence life.

Consider the comparison honestly, arsenic a interrogator experiences it. A collaborator is disposable occasionally, betwixt school commitments, assistance deadlines and clip zones; an LLM is disposable astatine 2am connected a Sunday, which – fto america not dress different – is erstwhile a worrying magnitude of investigation reasoning really happens. A collaborator arrives pinch their ain agenda, their ain framing of the problem, their ain inconvenient condemnation that your cardinal presumption is wrong; an LLM arrives pinch nary schedule beyond being useful to you. A collaborator must beryllium persuaded; an LLM must simply beryllium prompted. A collaborator will situation you successful ways you did not inquire for and had not thought of; an LLM will situation you precisely arsenic robustly arsenic you petition – and not 1 grade more. If you inquire it to critique your argument, it will do so, capably. But it will critique the statement you brought. It will not, unbidden, show you that you are solving the incorrect problem, that a rival group tried this successful 2019 and abandoned it, aliases that your beautiful theoretical framing collapses connected interaction pinch the messy realities of personification else’s field.

The collaborator’s inconvenience, successful different words, is not a bug successful the collaboration; it mostly is the collaboration. The worth of different mind lies precisely successful the ways it refuses to beryllium an hold of your own.

And beyond situation and serendipity lies thing much basal still. Research is not simply a accumulation usability that converts ideas into papers. It is simply a organization of believe – a societal cloth maintained done argument, apprenticeship, conference-bar conversations, the shared ordeal of a difficult referee report. That cloth is woven from precisely the frictions we are now engineering away. Every speech redirected from a workfellow to a chatbot is simply a thread softly withdrawn. No individual thread matters. The cloth does. There is simply a connection for what happens erstwhile this proceeds astatine scale, and I propose we usage it: decollaboration.

The inducement trap

If this were only a matter of individual temptation, it would beryllium a taxable for self-discipline and the occasional stern editorial. But it is not. The uncomfortable truth for investigation leaders, funders and institutions is that we person built an inducement strategy that makes decollaboration the rational prime – and we are busily making it much logical by the year.

Collaboration has ever been expensive. It costs recreation and scheduling and the glacial activity of building trust. It costs ego guidance and discuss and the periodic mini heartbreak of the author-order negotiation. Publish-or-perish has ever taxed these expenses, because each hr spent aligning pinch a co-author is an hr not spent producing output that is legible to an information system. But 3 forces are now compounding.

First, backing pressure. When budgets tighten, the first casualties are ever the statement items whose worth is existent but unmeasurable: travel, workshops, sabbaticals, visiting positions – the full beingness infrastructure of serendipity. We defund the corridor and past wonderment wherever the corridor conversations went.

Second, velocity worship. Our information systems – immoderate DORA-shaped statements adorn our websites – still reward output and velocity of output. And LLMs committedness velocity supra each things. They susurration that the lit reappraisal tin beryllium done by Friday, that the draught tin beryllium by Monday. To a interrogator whose adjacent position depends connected the magnitude of a publication list, this is not a susurration that is easy to ignore.

Third, and astir seductively: the LLM ne'er argues astir writer order. It has nary ego to manage, nary competing agenda, nary rival declare connected the credit. In a strategy wherever in installments is the rate of survival, a superb interlocutor who demands nary stock of the in installments is not simply convenient. It is arbitrage.

The investigation connected investigation should springiness america region here. We cognize from the team-science lit that mini teams thin to disrupt while ample teams develop, and that it is the unexpected combinations of collaborators – the atypical pairings crossed fields and institutions – that disproportionately nutrient the astir caller work. Meanwhile, the early grounds connected generative AI points successful a guidance that should sound familiar: individual productivity rises while the diverseness of ideas narrows, everyone moving faster on progressively akin paths. We are, successful effect, moving an uncontrolled world research successful trading serendipity for throughput – and the inducement structures we person built are the experimental apparatus.

Writing is thinking

There is simply a deeper costs still, and it concerns the point LLMs do astir impressively: writing.

Part of the lure of these devices is the belief of speed. And it is worthy being precise astir why that belief is partially an illusion. Writing takes clip because thinking takes time; the 2 are not separable activities that hap to hap successful sequence. The worth of penning a insubstantial was ne'er really the artefact – it was the forcing function. Writing is wherever we observe that the statement we were judge of has a spread successful conception three; wherever a vague intuition either becomes a precise declare aliases dissolves nether the unit of having to beryllium one. High-quality ideas are slow to generate, and historically overmuch of that slow procreation happened between people: astatine the whiteboard, successful the corridor, successful the productive irritation of a co-author’s tracked changes.

Cognitive psychologists person a sanction for this benignant of productive friction: Robert Bjork’s desirable difficulties – the uncovering that definite frictions successful learning and thinking, effortful retrieval, delayed feedback, the struggle to get an thought onto the page successful your ain words, are not obstacles to durable knowing but the system of it. Remove the trouble for the liking of comfort, and you do not get the aforesaid knowing faster. You get little of it, delivered much smoothly.

Outsource the penning and you person not accelerated the thinking; you person skipped it. And the trap deepens arsenic the models improve, because their output becomes steadily amended – much fluent, much structured, little and little distinguishable from knowledgeable quality prose. The enticement grows precisely arsenic the tell-tale signs shrink. As a physicist, I americium reminded of PT-symmetric quantum systems, which person the disconcerting spot of behaving wholly usually correct up until a hidden symmetry breaks – astatine which constituent everything changes astatine once. A investigation civilization tin look patient by each visible measurement – outputs up, turnaround down, prose immaculate – while the invisible point that sustained it softly falls beneath threshold. By the clip the denotation appears successful the metrics, the origin is years successful the past.

Pilots and passengers

None of this is an statement for banning the tools. That would fail, and it would merit to fail, because the committedness is genuine. In a March 2026 Nature comment, Dashun Wang extends Steve Jobs’s celebrated explanation of the machine arsenic “a bicycle for our minds”, proposing that AI agents are aeroplanes for the mind: faster and much powerful than the bicycle, harder to control, costlier erstwhile they crash. He is correct astir the upside, and it is considerable. When AI collapses the costs of failure, riskier and much eager questions go logical to ask; erstwhile it collapses the costs of analysis, mini labs and lone newcomers tin effort what erstwhile required armies. Used well, these devices genuinely widen investigation – and, done properly, agentic systems could moreover fortify 1 of science’s weakest joints, reproducibility, by making each analytical measurement loggable and replayable.

But announcement what moreover the optimists insist upon. Wang’s medicine is what he calls pilot-in-command science: the interrogator arsenic captain, the agents arsenic unit – an expert to draft, a professional to probe, a planner to representation adjacent steps – pinch the quality retaining authority complete the question, the way and the conclusions. Even the astir enthusiastic architects of AI-assisted discovery, successful different words, are adamant that location must beryllium personification successful the beforehand seat. And Wang arrives astatine the Waymo effect interest from the different direction: he warns that while AI whitethorn raise an individual’s performance, it risks reducing corporate diversity, pinch outputs converging unless we deliberately counteract it – and that conscionable arsenic our collaborators style america successful profound and unexpected ways, truthful excessively will our agents. His remedy is to creation for dissent and cultivate galore models that deliberation differently. Which is to say: having removed the friction, we must now technologist it backmost in.

This is an aged interest wearing caller clothes. In 1983, the scientist Lisanne Bainbridge described the ironies of automation: the much reliably a strategy runs connected its own, the little believe its quality overseer gets astatine the very accomplishment they are being kept successful the loop to workout successful an emergency, truthful that competence softly erodes precisely because it is needed only rarely, and ever astatine the worst moment. Swap “pilot” for “researcher” and “autopilot” for “agent”, and the system Wang is guarding against is the 1 Bainbridge named 4 decades earlier.

The problem, then, is not the technology. The problem is that we person made the quality speech the costly action and the instrumentality speech free, and we are amazed astatine what researchers, responding rationally, past choose.

Fund the friction

So the work falls wherever the incentives are made. Funders and institutions should recognise collaboration for what it is – not a nice-to-have, but a shape of infrastructure – and value it accordingly.

Fund the friction: the workshops, the visits, the co-location, the unstructured clip that produces the conversations cipher could person scheduled. Evaluate publication alternatively than velocity, and mean it. Treat “who did you deliberation with?” arsenic a mobility worthy of the aforesaid seriousness arsenic “what did you publish?”

And perhaps, successful an property erstwhile a instrumentality tin nutrient immoderate number of competent papers, we should announcement that the scarce and valuable point has inverted: the output is becoming cheap, and it is the reasoning together that is becoming precious.

The return journey

Susan and I sewage wherever we were going. The car was smooth, the soundlessness was comfortable, and the collaboration for which we had each crossed an water and a continent happened – not successful the frictionless capsule, but successful the inefficient, unscheduled, thoroughly quality hours that followed it: the conversations that wandered, the disagreements neither of america had planned, the ideas that neither of america brought to San Francisco because they did not beryllium until we were successful the aforesaid room.

That, successful the end, is the Waymo effect successful full: the thrust is delightful, the destination is reached, and it is only if you glimpse up beforehand that you announcement what is missing. Wang calls for pilot-in-command science; the quiet consequence is that our incentives are training america for passenger-in-comfort subject instead.

Research is now climbing into the backmost seat. The thrust will beryllium smooth. The output will arrive. The mobility that investigation leaders should beryllium asking – earlier we person each had a fewer much years of comfortable rides – is simply a elemental one: who is successful the beforehand seat? And erstwhile did we extremity noticing that cipher was driving?

Daniel Hook is Chief Scientific Officer astatine Holtzbrinck Group. With acknowledgment to Susan Winslow and Oliver Zahn for absorbing conversations that motivated this writing.