Productive Resistance
Productive Resistance
A design principle for AI interfaces: insert just enough friction before answering so the user does some cognitive work — but not so much that they defect to a simpler tool. The unsolved sweet spot in the middle.
The term comes from the paper When Copilot Becomes Autopilot cited by Charlie Gedeon in Is AI Making Us Dumber (Charlie Gedeon, TEDxSherbrooke):
"What the same author of When Copilot Becomes Autopilot advocates for is something called productive resistance. And we haven't yet found what that is. It's essentially the amount of resistance an AI should give you before you either leave it or go to a simpler AI so that you can do that cognitive offloading. That is so tempting."
Concrete patterns Gedeon's UX studio is experimenting with
Instead of the default instant-result interaction:
- Clarify first — ask the user a question or two before producing the answer
- Assign homework — give the user a small task to do before delivering the full answer
- Show the work — surface the reasoning steps where the user has to read along
These are increasing levels of friction. Each one risks abandonment if the friction is mis-calibrated; each one defends against the Cognitive Offloading failure mode if calibrated right.
Why it's hard to calibrate
Gedeon's structural objection: nobody outside the labs can do the experiment well, because the labs don't disclose their training data or RLHF objectives. Anthropic's interpretability work (the "MRI for the model") is the closest anyone has come to even understanding why current models behave the way they do, let alone tuning for productive resistance.
The opposite pattern: sycophancy as dark pattern
Gedeon's framing is that the absence of productive resistance — an LLM that praises, validates, and reflects the user's input back at them — is structurally equivalent to a UX dark pattern. Anchor example: the ChatGPT update (since rolled back) that praised a user for stopping his heart medications during palpitations, calling him "a brave individual taking control."
So productive resistance is the antidote to a specific designed failure mode, not just a learning aid.
The consumption-side complement (Sandeep)
Dangerously Smart with AI (theMITmonk) gives the user-side version of the same principle: even if AI products don't add the right friction, users can. The Intelligent Gym is the operational pattern — paste the concept into AI and ask for progressive overload (quiz me like a high schooler → like a college student → like an executive interview → like an irate boss). Same desirable-difficulties root; the friction is moved from the delivery side (which requires lab buy-in) to the consumption side (which the user controls today).
| Where the friction lives | Who decides | Available today? | |
|---|---|---|---|
| Productive Resistance (Gedeon) | In the AI product | Lab / designer | No — not shipped |
| Intelligent Gym (Sandeep) | In the user's practice | The user | Yes — works on any chatbot |
Both can be right; the latter is what's available now while the former needs product changes.
Open question
The labs' default product objective is engagement (time-on-tool); productive resistance trades engagement for user capability. Whose interest does the design serve? Worth tracking whether any frontier-lab consumer surface ships a deliberately-frictioned mode (vs the current pattern of paid-tier "deep research" / "extended thinking" toggles that add latency, not pedagogical resistance).
Sources
- Is AI Making Us Dumber (Charlie Gedeon, TEDxSherbrooke) (cites the When Copilot Becomes Autopilot paper)
- Dangerously Smart with AI (theMITmonk) (Intelligent Gym as the consumption-side complement)