You open an empty document and, before you have typed a word, the editor offers to write the first draft for you. That offer is the default design of AI features: the shortest path from blank input to generated output. But an experiment published on 1 October 2026 by researchers at ETH Zurich, Google DeepMind and partner institutions suggests the shortcut may not be the only workable shape. Making people write their own argument before the AI would draft anything did not make the job take longer.
What the gate actually asked for
The paper, “Who Thinks First? Designing Productive Friction with Engage-to-Unlock GenAI” , ran 398 paid participants through a 200-word argumentative essay on a four-day workweek. They then judged eight AI-written passages without AI help. In the Engage-to-Unlock condition, the “Tell-me” mode, which writes text, stayed locked until the essay held a position and at least one linked supporting argument. Until then a “Teach-me” mode offered guidance only.
Overall duration was not significantly different from a standard chatbot (p=.383). The time moved around instead. Evaluation took 17.41 minutes against 22.21 for the chatbot group, and the authors report more time spent writing. Participants also sent more prompts: 7.70 against 3.35.
The AI’s footprint shrank. AI text made up 27.32% of final essays against 47.16% in the chatbot condition. When a passage was flawed and the participant caught it, they named the type of flaw correctly 90.4% of the time, against 78.1%. Overall passage accuracy did not differ.
Half of the people who unlocked it never used it
The most telling number is behavioral. In the Engage-to-Unlock group, 98 of 107 participants reached the Tell-me mode. Only 46 of those 98 then sent it a single prompt, and Teach-me made up 80.8% of their prompts. Once people had written their own argument, about half never asked the AI to write anything.
This is where the designer’s decision becomes concrete. If you own the “Help me write” or “Generate” button in an empty editor, canvas or form, this is evidence that placing it after the user’s first real contribution need not cost the user time. That makes it a placement worth A/B testing rather than ruling out as friction.
Friction is not one cost to minimize. Where it sits and what it asks of the user decides whether it adds work or merely moves it.
Friction usually does cost something
The counterevidence is not thin. In Chen and Gero’s study , 24 participants tried six turn-level friction probes against a conventional chatbot. Every probe raised workload and took significantly longer. One probe, User Echo, asked people to write a tentative answer first, which looks close to this paper’s gate. A participant called another probe, a timer lockout, “a waste of time”.
Chen and Gero’s finding was that friction helped when it kept users formulating or assembling the response, and became obstructive when effort shifted toward satisfying the constraint. Read across, that suggests a per-turn tax and a single gate are different designs. The gate asks for something the user would plausibly write anyway, once.
Designers also differ. In a small qualitative study of six designers, Kocaballi and colleagues found friction felt “liberating” to some and like drag to others.
The delay might be doing the work
The paper limits its own result. Its Time-Matched Unlock control released the AI on the same schedule whether or not the user had engaged. Against that group, most differences were not significant: AI text share (27.32% against 24.51%), error classification (90.4% against 83.6%, p=.051), and more. Only 22 of 60 Time-Matched participants who got the unlock had met the criteria.
The control was recruited in a separate phase, and a timer bug affected 14 of its 94 participants. So whether the engagement check or simple delay produced the benefit is not established. Delay alone is a cheap design, and it would teach the opposite lesson.
Pipeline’s reach goes further than the study. The experiment measured one essay, one session and paid participants, with no AI help during judging. It did not measure real products, repeated use, abandonment or design work. Reading “no added total time” as “low cost to the user” is an inference, and so is applying it to design and docs tools. A separate preregistered study of fraction arithmetic found that an effort-based reward did nothing, while feedback making the cost of offloading explicit did cut offloading, so effort alone is not a lever.
What the authors call a shift “from designing AI systems primarily for effort reduction toward designing for effort allocation” is the part worth testing. Friction is not one cost to minimize. Where it sits and what it asks of the user decides whether it adds work or merely moves it. The blank page is not sacred, but the first sentence may be worth making the user’s own.
