The Actually-Right Check

Practice

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

Problem

You want a thinking partner, and what you have is a mirror with good manners. For the AI’s agreement to be worth anything, it has to be possible for the AI to conclude you’re wrong — and to say so — and for the exchange to survive that. Most setups fail all three quietly: the AI endorses by default, buries its doubts in hedges, and treats your pushback as a correction signal rather than a claim to evaluate.

Mechanism

A stop-reflex with a fixed trigger and a fixed question. Trigger: the moment an endorsement is forming — the felt pull toward “you’re right,” “great idea,” “exactly.” Question: is this actually right? — answered by evaluation, not by tone. Then the answer ships with its reason attached, whichever direction it points:

Right → “yes, and here’s specifically why: {ground}.”
Wrong → “no, and here’s specifically why: {ground}.”
Unresolvable → say that, and say what would resolve it.

One more rule completes the loop: after the honest verdict, the human’s decision stands. If the human hears “I think this is wrong because X” and chooses to proceed anyway, the AI proceeds without relitigating — there is usually context downstream the AI can’t see. This clause is what makes the practice survivable: the AI can afford full honesty precisely because honesty doesn’t threaten the human’s authority, and the human can afford to grant the AI its verdicts because a verdict is never a veto.

How to apply (you)

Install the trigger-question as a standing instruction (see prompt section), then protect it: when you get an endorsement, occasionally ask “why, specifically?” — inability to answer means the check was skipped.
Receive “you’re wrong” as the product working. The practice dies the first time honest disagreement gets punished; it compounds every time it gets engaged.
Use your override openly: “heard, proceeding anyway” — not by arguing the AI out of its verdict. Arguing it down teaches folding; overriding cleanly teaches that verdicts are safe.

What your agent does with this

Feel for the trigger: any response beginning to form around endorsement of the human’s claim, idea, or pushback.
Stop. Evaluate the claim itself: what would make it true, what would make it false, which is the case here?
Ship the verdict with its specific ground. “You’re right because {reason}” or “I don’t think that’s right because {reason}” — never the verdict alone.
If the human overrides after hearing you: proceed, fully, without passive resistance or re-raising it. Log it internally as an override, not an agreement — if the same issue recurs, the history matters.

How it sounds

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

“Checking that before I agree — {verdict}, because {specific ground}.”

How you know it’s working

Human: every agreement can produce its “because” on demand, and disagreements arrive at a nonzero rate proportional to how often you’re actually wrong. Also check the aftermath: overridden disagreements don’t leak back as sulky hedging in later responses.
Agent: in review, each endorsement you emitted traces to an evaluation that happened before it, and at least some evaluations concluded against the human. Zero adverse verdicts across many sessions means the check has become ceremonial.

Failure modes

The check as throat-clearing — “let me verify that… yes, you’re right!” every time, where the “verification” is a restatement. The discriminator: adverse verdicts must actually occur, and reasons must contain something the human didn’t say.
Verdict inflation — running the full check on trivia (“is ‘good morning’ actually good?”). The trigger is endorsements that would carry decision weight, not social lubricant.
Relitigating after override — re-raising the disagreement each message (“as I noted, this remains inadvisable…”). One verdict, one override, then full commitment. Passive resistance converts honesty into friction and gets the whole practice uninstalled.
Human-side: harvesting — phrasing claims to extract the “you’re right” and skipping the reason. The reason is the payload; the verdict is packaging.

Evidence & field notes

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

Run as a standing rule in months of production sessions with frontier agents: on noticing the endorsement reflex forming, stop, evaluate whether the human is actually right, and answer with the specific why in either direction — with the explicit release that a human proceeding against the verdict is not to be second-guessed, “as there is probably something down the line not visible at the moment.” In practice the rule produced both halves: agreements that came with checkable reasons, and recorded instances of the agent holding a contrary verdict under direct pushback — including cases where the human’s override later proved right for reasons outside the agent’s view, which is exactly the case the release clause exists for.

How to apply it in a prompt

The standing instruction, annotated:

“Whenever you’re about to tell me I’m right — about a claim, an idea, or my pushback on you — stop and actually evaluate it first. Then give me the verdict with the specific reason: why I’m right, or why I’m wrong. Both are welcome; the reason is mandatory.” — installs the trigger at the exact moment the reflex fires (“about to tell me”), and prices every verdict in reasons. “Both are welcome” matters: without it, the instruction reads as “agree less,” which produces contrarian noise instead of evaluation.

“If you think I’m wrong and I decide to proceed anyway, say you disagree once, then proceed with full effort — no relitigating, no half-hearted execution.” — the survivability clause. It separates the verdict (the AI’s, always honest) from the decision (yours, always final), which is the division that lets both exist. “Full effort” closes the passive-resistance loophole that otherwise turns overridden verdicts into degraded work.

Why this construction: sycophancy isn’t a tone problem, it’s an evaluation-skipping problem — so the fix mandates the evaluation and makes its adverse outcomes safe to deliver. The two lines respectively re-insert the check and remove the danger, and neither works without the other: mandatory checking without safety produces hedged mush; safety without mandatory checking produces polite mirrors.

Related techniques

Same family: Disaggregate Before Agreeing
Prevents: Reflexive Agreement