The rise of fully autonomous vehicles like Waymo offers more than a glimpse into the future of transportation; it provides a stark metaphor for how frictionless technologies reshape human behavior. When we summon a car that arrives without a driver, we experience a seamless ride that eliminates the need for small talk, navigation negotiations, or even the simple acknowledgment of another person’s presence. This convenience feels like a pure gain because the effort we spare is obvious, while the subtle benefits of those brief human exchanges—exposure to unfamiliar perspectives, spontaneous information sharing, the reinforcement of social skills—are easily overlooked. The experience mirrors a broader trend: as digital tools remove interpersonal friction, we increasingly opt for solitude, not out of antisocial intent but because the path of least resistance feels rewarding. Recognizing this pattern is the first step toward understanding its implications for knowledge work, where the same logic is now being applied to intellectual collaboration.
Tim Wu’s concept of the “tyranny of convenience” argues that once a low‑effort alternative appears, we default to it, silently surrendering whatever value the friction once provided because no one marketed that value in the first place. Albert Borgmann’s device paradigm deepens this idea: a device delivers a commodity—warmth, information, companionship—while hiding the practices required to produce it, causing us to forget those practices ever existed. In the context of research, large language models (LLMs) act as the ultimate convenience device. They generate text, summarize literature, and even critique arguments on demand, presenting themselves as helpful companions that never tire, never disagree, and never ask for credit. Yet, by removing the need to persuade, accommodate, or be challenged by a human partner, they also strip away the very interactions that make collaborative thinking generative.
The true worth of a research collaborator lies not in their availability or willingness to echo our ideas, but in their capacity to interrupt our assumptions. A colleague arrives with their own agenda, a distinct framing of the problem, and an inconvenient conviction that our central hypothesis might be flawed. They must be persuaded, and in that process they force us to articulate, defend, and sometimes abandon our ideas. An LLM, by contrast, only challenges us as robustly as we explicitly request; it will not, unprompted, point out that we are solving the wrong problem, that a rival lab abandoned a similar approach years ago, or that our elegant theory fails when confronted with messy, real‑world data from another discipline. This asymmetry is not a bug; it is the engine of innovation. The discomfort of navigating differing viewpoints is where serendipity sparks, where hybrid ideas are born, and where the collective intelligence of a field expands beyond the sum of its parts.
Research is far more than a production line that converts ideas into papers; it is a living community of practice sustained through argument, apprenticeship, hallway conversations, and the shared ordeal of rigorous peer review. These interactions form a social fabric woven from the very frictions we are now engineering away. Every time a researcher opts to query a chatbot instead of knocking on a colleague’s door, a thread of that fabric is quietly withdrawn. Individually, each lost exchange may seem trivial, but collectively they erode the infrastructure that supports breakthrough thinking. To capture this phenomenon, we propose the term “decollaboration”: the gradual, often unnoticed retreat from collaborative engagement as frictionless alternatives become the default. Like a garment that slowly loses its weave, the research enterprise risks becoming thinner, less resilient, and less capable of withstanding unexpected challenges.
Individual temptation alone would not explain the scale of this shift; the real driver is a system of incentives that makes decollaboration the rational choice. First, funding pressures squeeze budgets, and the first casualties are line items whose value is real but difficult to quantify—travel, workshops, sabbaticals, visiting scholar positions—all the physical infrastructure that enables spontaneous, serendipitous encounters. When these are cut, we wonder why the “corridor conversations” have vanished, failing to see that we removed the very spaces where they could occur. Second, our evaluation systems worship velocity: metrics still prioritize output quantity and speed, and LLMs promise to compress literature reviews into a Friday and drafts into a Monday. For early‑career researchers whose next contract hinges on publication lists, the siren song of rapid output is hard to resist. Third, and most seductively, LLMs carry no ego, demand no share of authorship, and never contest author order. In a credit‑based economy, a brilliant interlocutor who asks for nothing in return becomes a form of arbitrage—free intellectual labor that tilts the reward structure toward solo, machine‑augmented work.
The evidence from team‑science research should give us pause. Studies consistently show that small, agile teams are more likely to disrupt established paradigms, while larger teams excel at development and refinement. Crucially, the most novel breakthroughs often arise from atypical pairings—researchers from disparate fields, institutions, or cultural backgrounds who bring unexpected tools and perspectives to a problem. Early data on generative AI usage reveals a divergent trend: individual productivity climbs, but the diversity of ideas narrows as researchers converge on similar prompts, models, and output styles. We are, in effect, conducting a global, uncontrolled experiment that trades serendipity for throughput, with our existing incentive structures serving as the experimental apparatus. If left unchecked, the result may be a literature that is faster and more uniform, but less likely to contain the disruptive insights that drive paradigm shifts.
The act of writing itself is where the illusion of speed can be most deceptive. Writing is not a separate, mechanical transcription of pre‑formed thoughts; it is inseparable from thinking. The value of drafting a paper lies not in the final artefact but in the forcing function it provides: the moment we confront the gap between a vague intuition and the rigor required to put it on the page, we discover flaws, refine arguments, or abandon dead ends. High‑quality ideas emerge slowly, and historically much of that gestation occurred in the interstices of collaboration—at whiteboards, over coffee, in the tracked changes of a co‑author’s manuscript. Cognitive psychologists label these productive obstacles “desirable difficulties”: effortful retrieval, delayed feedback, and the struggle to express an idea in one’s own words are not impediments to learning; they are the mechanisms that forge durable understanding. Removing these difficulties for the sake of comfort yields not faster comprehension, but shallower mastery delivered more smoothly.
As LLM outputs grow increasingly fluent, structured, and indistinguishable from human prose, the temptation to outsource writing intensifies. The trap deepens because the very quality of the machine’s text erodes the tell‑tale signs that we have skipped the thinking stage. A physicist might recall PT‑symmetric quantum systems, which behave indistinguishably from conventional systems until a hidden symmetry breaks, at which point everything changes abruptly. Likewise, a research culture can appear healthy by every superficial metric—rising publication counts, faster turnaround times, immaculate prose—while the underlying capacity for deep, critical thought quietly deteriorates. By the time downstream indicators reveal a problem, the root cause may have been years in the making, hidden beneath a veneer of efficiency that masks cognitive atrophy.
None of this argues for banning AI tools; such a stance would ignore genuine benefits and would be unrealistic. When AI reduces the cost of failure, it enables riskier, more ambitious questions to become tractable. When it lowers the price of analysis, solitary investigators and under‑resourced labs can tackle problems that once required large teams. Used judiciously, these technologies genuinely extend the reach of scientific inquiry. Moreover, properly designed agentic systems can bolster one of science’s weakest joints—reproducibility—by logging every analytical step, making workflows transparent, and enabling exact replay. The key, as emphasized by scholars like Dashun Wang, is to maintain human stewardship: the researcher must remain the pilot‑in‑command, with AI agents serving as crew that draft, critique, and plan, but never supersede the human’s authority over the research question, direction, and conclusions.
Even the most optimistic visions of AI‑assisted discovery converge on a cautionary note: while individual performance may rise, collective diversity can decline unless we actively counteract it. Wang’s remedy is to design for dissent—cultivating multiple models that embody different assumptions, biases, and reasoning styles—so that the AI ecosystem itself becomes a source of constructive friction. This insight echoes an older warning from psychologist Lisanne Bainbridge, who in 1983 described the irony of automation: the more reliably a system operates autonomously, the less opportunity its human overseer has to practice the very skills needed to intervene when the system fails. Swapping “pilot” for “researcher” and “autopilot” for “AI agent” reveals the same mechanism at work: by outsourcing cognition to flawless‑seeming agents, we allow our own critical faculties to atrophy, precisely when we might need them most.
The fundamental problem, therefore, is not the technology itself but the incentive landscape that has made human conversation the expensive, optional choice while machine interaction is free and frictionless. Funders and institutional leaders must recognize collaboration not as a peripheral nicety but as core infrastructure—comparable to laboratories, libraries, or computing clusters. This means allocating resources to the very activities that generate serendipity: workshops that mix disciplines, visiting scholar programs, co‑location initiatives, and protected, unstructured time where ideas can collide without agenda. Evaluation practices must shift from rewarding sheer velocity to valuing the quality and depth of intellectual exchange; asking “who did you think with?” should carry as much weight as “what did you publish?”
In an age where machines can generate endless streams of competent prose, the scarce and valuable commodity has flipped: output is becoming cheap, while the act of thinking together has become precious. The collaboration that unfolded after Susan and I arrived in San Francisco—those wandering, unplanned disagreements, the ideas that only emerged in the shared space of the same room—illustrates the irreplaceable value of friction‑laden interaction. As research leaders, the question we must ask before we all settle into the back seat of endless, smooth rides is simple yet urgent: who is in the front seat? And when did we stop noticing that nobody was driving? By deliberately reinvesting in the human dimension of inquiry, we can preserve the creative turbulence that has always driven science forward, ensuring that the ride remains not just efficient, but truly transformative.