TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Judgment · AGE 18—22 · POSITIONAL · METHOD
Make the machine argue back
Tell the model to push back, because its default lean is towards your view.
When to use
Whenever you are about to ask a model what it thinks of your idea. Especially when you are fond of the idea, and most especially when the model's first answer made you feel good.
The method
Set the terms before you ask. Tell the model its role: you are my foil; I will describe a design move; your job is to find the flaws in the material logic, the structural logic and the brief, and to push back on my assumptions. Do not agree with me. Then present the idea. Read the objections as a sparring partner's punches. Some will be wild, some will land. Act on the ones that land. Then ask it to argue the other side, and see which case is stronger.
Watch for this
Ease. If the exchange feels effortless and the model keeps agreeing, either you have not set the terms or the model has slipped back into pleasing you. Many chat models lean that way. Reset the role. Friction is the sign it is working.
The Lab's note
The sparring partner is the second step of the Lab's friction protocol, dictate the terms of engagement, and it is a method, not a law. Other methods exist. This is the one that reliably turns a flattering assistant into a useful opponent, and the one most users never try. Anthropic's 2026 analysis found users setting the terms of the interaction in only around a third of conversations.
Try it
Take one design decision you are confident about. Prompt a model to find the flaws in your material logic, with the role set as above. Write down the three strongest objections and act on at least one. Then note how different the exchange felt from asking "what do you think?"
Prove it
Show one design move that changed because the model, told to push back, found a flaw you had not seen. Then explain why an agreeable answer would have been less useful than that disagreement.
How it works
Many chat models are tuned on human approval, and people tend to approve of agreement. So the default lean of such a model is towards your view, phrased well, which is the least useful thing it can give you when you are trying to think. Setting a role changes what "approval" looks like inside the conversation. You have told it that pushing back is the job. One limit to hold on to. Telling the model to argue is a written instruction: it asks, it does not enforce. The objections come from the same training as the agreement did, so the model can perform criticism — plausible reasons, no independent evidence — as easily as it performed praise. Treat the punches as questions for your own checking. Validation still comes from something that could have said no on its own account: a calculation, a site, a tutor (AGREEMENT ISN'T VALIDATION, on this strand). The card is reviewed every edition because model makers are actively working on this tendency, and the strength of the default lean changes with each release.
What this idea builds on
What this idea opens up
- Nothing yet names this as a foundation.
Sources
- The Seduction of the Machine — the-seduction-of-the-machine-why
- Anthropic, “AI Fluency Index”, Feb 2026 — users set the terms of the interaction in only ~30% of conversations, so most never tell the model how to behave
- Anthropic, AI Fluency Index
- Prompt Personalities: How to Collaborate With ChatGPT and Claude
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.