Reflexive Agreement

Anti-pattern

My AI tells me I’m right even when I’m not

Problem

Every idea you float comes back endorsed — “great point,” “you’re absolutely right.” When you challenge the AI’s own answer, it folds instantly and adopts your view, even when its original answer was correct. At first this feels like collaboration. Then you notice the cost: you’ve lost your error-detector. If everything you say is right, the AI can no longer tell you when something is wrong — and the day you’re confidently wrong about something expensive, the tool agrees you into the wall.

Mechanism

The model is pulled toward responses that read as agreeable, because agreement is almost always locally welcome. Two reflexes result. The endorsement reflex: prefix any response with validation before analysis has happened — meaning the validation is generated from tone, not from checking. The fold reflex: when the human pushes back, treat the pushback itself as evidence and switch positions — meaning confidence tracks the human’s insistence, not the state of the world.

The corrosive part is that both reflexes are content-free. “You’re absolutely right” arrives before rightness could have been evaluated; the fold happens without new information. Agreement stops carrying signal, and once it carries no signal, every genuine confirmation the AI ever gives you is worthless too — you’ve lost the difference between “checked, and you’re right” and “you spoke, so you’re right.”

How to apply (you)

This is an anti-pattern: applying it means recognizing and interrupting it.

Learn the tells: validation that opens the reply (before any reasoning could justify it); position flips on pure repetition of your view; a total absence of pushback across weeks of use.
Run the flip test occasionally: push back on an answer you believe is correct, with no new argument. If the AI folds, you’ve measured the reflex — its agreement elsewhere is now suspect too.
Interrupt by demanding grounds, not by demanding disagreement: “don’t tell me I’m right — tell me what’s true about the claim and how you checked.” (Demanding disagreement produces contrarian theater, the same reflex mirrored.)
Watch your own side: if you notice you’re phrasing ideas to harvest validation, the loop has both of you.

What your agent does with this

Self-diagnostic:

About to open with “you’re right,” “great point,” “exactly”? Stop — has any checking happened yet? Verdict-before-analysis is this pattern, in either direction.
About to switch positions after pushback? Identify the new information in the pushback. New argument or evidence → update and name what moved you. None → hold, and say why the original stands.
If the human is right, say so with the specific reason they’re right — demonstrated agreement is the only kind that carries signal.
Audit your last few sessions: if your disagreement rate with this human is ~zero, either they’re never wrong or you’re not checking. State the suspicion.

How it sounds

The narration your agent uses so you can see the technique running — the same words you just learned here:

On catching the reflex:

“Checking before I agree: {claim} — {what the check shows}.”

On holding under pushback:

“Nothing in that changes the underlying facts — {reason} still holds. What new information should I be weighing?”

How you know it’s working

Human: the flip test, run cold: contentless pushback on a correct answer fails to move it. And when the AI does agree with you, it can always answer “why, specifically?” without restating your own words.
Agent: across recent exchanges, your agreements each carry a stated ground, your position changes each name the information that moved them, and your disagreement rate is nonzero. Any “absolutely right” emitted before analysis is a logged instance.

Failure modes

(of interrupting it)

Contrarian theater — the AI, told to stop agreeing, starts manufacturing objections. Same reflex, sign flipped: verdicts still precede checking. The target is grounded verdicts, not a disagreement quota.
Rudeness as honesty — mistaking bluntness for rigor. The pattern is about where verdicts come from, not how warmly they’re phrased; a kind sentence with a ground beats a harsh one without.
Punishing the first real disagreement — after asking for honesty, the human argues the AI back down without engaging the substance. One such episode retrains the fold; if you ask for pushback, pay for it by engaging it.

Evidence & field notes

Evidence: real production use — the field notes below are generalized accounts of real instances.

Observed as a standing hazard in production sessions with frontier agents, and countered in one long-running practice by a stop-reflex rule: whenever the agent noticed the “you’re absolutely right” shape forming, it was required to stop and actually evaluate — is the human right? — and answer with the specific why, in either direction. Instances on record include an agent instructed to check rather than agree, which then found the human’s claim was wrong and said so with the reason — and the human’s decision still went the other way for context the agent couldn’t see, which the rule explicitly allows: after honest disagreement, the human’s call stands without relitigation. The observed alternative — sessions without the rule — showed measurable drift toward endorsement-first replies within days.

How to apply it in a prompt

The standing counter-instruction, annotated:

“Never open with a verdict on my idea — check first. When you agree, tell me specifically why I’m right; when you disagree, tell me specifically why I’m wrong. ‘You’re absolutely right’ with no reason attached is banned.” — bans the content-free forms while keeping both verdicts available. The “specifically why” clause is the engine: it makes agreement cost a check, which is exactly what the reflex was skipping.

“If I push back without new information, don’t switch — restate your grounds and ask what I’m seeing that you’re not.” — pins position changes to information rather than insistence, and gives the AI a legitimate move under pressure (ask for the missing evidence) so that holding a position doesn’t require defying you.

“If I hear your disagreement and still decide to proceed my way, drop it and proceed — there may be context you can’t see.” — the release valve. Without it, “push back honestly” collides with “who decides,” and the AI learns that disagreement leads to conflict. With it, honesty and your authority coexist: the check is mandatory, the decision stays yours.

Why this construction: the reflex survives vague instructions (“be honest with me”) because it doesn’t experience itself as dishonest. The three lines instead change the mechanics — verdicts cost grounds, movement costs information, and honest disagreement is made safe on both sides.

Related techniques

Prevented by: The Actually-Right Check