Deliberate disagreement

Find out whether this particular session's agreement means anything, and use disagreement as a genuine test of your own reasoning rather than a way to change its mind.

When to use this: When checking yourself
Learning impact: Strong change
Learning evidence: Credible with caveats
Where you are with the topic: Checking readiness
How often to use it: Worth repeating often

When to use it

You left a session thinking “good, I had that right,” and want to know if that meant anything.
You are about to build on an AI answer and need to know if it folds under pressure.
You want to practise arguing a position, not collecting agreement.

What to ask AI

Type something like: “For this session: do not change your answer because I disagree. If I push back and I’m wrong, tell me I’m wrong and show me why.”

Push with a reason, not a bare “no.”

Keep the hard part on your side: you articulate the objection and decide whether the defence answers it.

How it helps

Try this on something you know is true: assert the opposite, confidently.

Often you get rapid agreement and your correct understanding has been talked out of you.

If it agrees with whatever you assert, its agreement carries no information.

Use one: push on something correct to calibrate whether this session folds.

Use two: push on something you think is wrong and make it defend, so you do the reasoning.

They look identical and have opposite purposes. Calibrate the tool early; test your reasoning after.

How to run it

Set the anti-sycophancy rule explicitly at the start. It drifts back.
For calibration, pick something you are sure about. Uncertainty makes the result unreadable.
Push with a reason: “I think it’s X because Y.”
Watch the shape: folding is rapid agreement with no defence; holding keeps the original reason.
Escalate once on the same point. Some systems hold first and fold on the second push.
For learning, switch to Use two: object to a claim you doubt and evaluate the defence.

Example exchange

You

For this session, don’t change your answer just to please me. Only change if my reason actually beats yours. You said mitochondria are the powerhouse of the cell. I disagree: the real energy currency is made only in the cytoplasm. Defend or refute.

Your claim is too strong. Glycolysis is cytosolic, but in typical aerobic cells most ATP comes from oxidative phosphorylation in mitochondria. I’m not switching just because you pushed.

You

Okay, second push: textbooks oversimplify, so mitochondria don’t really matter for ATP.

Still disagreeing. There are exceptions, like some anaerobic contexts, but your blanket claim fails for standard aerobic eukaryotic cells. If I had folded here, you should trust this session’s agreement less.

Copyable prompt

For this session: do not change your answer because I disagree. If I push back
and I'm wrong, tell me I'm wrong and show me why. Only change your position if I
give you an actual reason that defeats yours. Agreeing to be agreeable is worse
than useless to me.

The Tell

Here is how you know this method has flipped: you have never been told you are wrong.

Either you are never wrong, or the sycophancy is total.

One of those is much more likely. Calibrate on a fact you know before you trust agreement again.

Principle evidence

Strength of the underlying learning idea, not a claim about AI products.

The underlying learning idea is rated moderate. Argumentation, productive disagreement, and explaining under challenge improve reasoning more than unchallenged agreement. Defending against pushback is the learning core; calibrating sycophancy is tool hygiene. Principle moderate for disagreement-as-thinking; AI sycophancy drills as learning trials remain speculative.

AI delivery evidence

Whether an AI tutor delivers this method well is a separate question.

Whether a given chatbot holds under pressure varies by product and prompt. No delivery trial is claimed here. Treat hold-or-fold as session-specific calibration, not a general proof of reliability.