Every generative tool is built on the assumption that reducing friction is the goal. Fewer clicks, faster response, lower effort, instant availability. It is the dominant instinct in interface design and it is usually right.
“Who Thinks First? Designing Productive Friction with Engage-to-Unlock GenAI”, posted 1 October 2026, tests whether it is right here. The authors are Xiaotian Su, Laura Rimell, Jiazheng Li, Amal Rannen-Triki, Ulrich Paquet, Lisa Anne Hendricks, Rida Qadri, Daphne Ippolito and Piotr Mirowski — a largely DeepMind-affiliated group, which is notable given the finding.
The mechanism
Engage-to-Unlock is a design pattern that unlocks generative capabilities after users meaningfully engage with the task. You cannot have the model until you have done some of the work.
The stated purpose is precise: to prevent “cognitive offloading before users develop their own ideas.”
That is a sharper claim than “AI makes people lazy.” Cognitive offloading is a real and usually beneficial phenomenon — writing things down, using a calculator, keeping a calendar. The concern is specifically about sequence: offloading after you have formed a view is leverage, offloading before is substitution. Ask a model for ideas before you have had any and you will adopt its framing, because you have nothing to compare it to. This is anchoring, and it is extremely difficult to resist once it has happened.
The study design, which is the best part
398 participants, four conditions:
| Condition | AI availability |
|---|---|
| Human-Only | None |
| Standard Chatbot | Immediate |
| Engage-to-Unlock | After meaningful engagement |
| Time-Matched Unlock | After the same elapsed time, regardless of engagement |
Participants completed writing tasks and evaluated passages for accuracy.
The Time-Matched Unlock condition is why this study is worth taking seriously. Without it, any benefit from Engage-to-Unlock has an obvious deflationary explanation: you just made people wait, and waiting gave them time to think. That is a much less interesting finding and it does not require anyone to engage with anything.
By matching the delay but not requiring engagement, the design separates “time passed” from “you did something.” Any difference between those two arms is attributable to the engagement itself.
This is the control that most “AI makes you worse/better” studies lack, and its presence is what moves this from an opinion to a result.
What they found
Engage-to-Unlock participants:
- spent more time writing and less time evaluating
- without increasing overall task duration
- and showed the highest accuracy-per-time evaluation efficiency across conditions
Read the first two together, because the combination is the finding. The time did not come from somewhere else — total duration was flat. What changed was where the time went: from assessing the model’s output to producing their own.
And the shift improved evaluation rather than degrading it. Less time spent evaluating, more accuracy per unit of that time. Which makes sense if you take the anchoring mechanism seriously: someone who has already formed a position evaluates a passage against something. Someone who hasn’t is reading it cold, with nothing to notice a discrepancy with, and so has to spend longer and does worse.
Where this applies to creative tools
Generative features are being added to every creative application, and the default implementation is a button that is always available. This result suggests the default is wrong — not that the feature is wrong.
Concretely, patterns that follow from it:
- Sketch before you generate. An image tool that wants a rough composition, a mask, or a few lines before the generate button activates. The user’s own spatial decision then survives into the output instead of being replaced by it.
- Write the brief yourself. Rather than a prompt box, a short structured statement of intent — what, for whom, why — which then unlocks generation. The brief is useful work regardless.
- Block out before you ask. In 3D, animation or audio: a greybox, a timing pass, a rough arrangement first.
- Drafts before critique. For any AI-feedback feature, require a draft of some substance before the critique tool opens.
The connection to existing practice is worth drawing. Every creative discipline already has a version of this and did not need a study to find it. Thumbnails before rendering. Blocking before polish. Writing a shitty first draft before editing. Life drawing before reference. The whole pedagogy of art and design is built on make something bad yourself before you look at how it’s done — and the reason is exactly the anchoring one.
What’s new is that generative tools removed the structural enforcement. Previously you had to block out your animation because there was no alternative. Now there is, and the discipline has to be designed back in.
The thing to be careful about
“Meaningfully engage” is doing a lot of work in that sentence, and it is where an implementation of this would live or die. A gate that is too easy becomes a ritual you click through. A gate that is too hard is a tool people stop using, and a creative application that lectures its users about effort will be abandoned for one that doesn’t.
Note also that the task here was writing and accuracy evaluation, which has a clear correct answer. Whether the same pattern helps for open-ended creative work — where there is no accuracy to measure and divergent ideas from a model may be genuinely valuable early — is not something this study can tell us. The mechanism is plausible there; the evidence is for the measurable case.