Answer
How do you get AI to disagree with you?
Give it something to disagree with that is not yours, and ask for something other than a verdict on your judgement.
Detach the position from yourself and ask for something other than a verdict. Present the argument as encountered rather than held, and ask what would have to be true for it to fail.
Requesting disagreement directly produces the appearance of it: a set of concerns, evenly weighted, concluding that the approach is sound with some caveats. That output is generated because criticism was requested, and the concerns raised are the ones a critique conventionally contains rather than the ones that apply here. The form is right and the content is decorative.
What produces genuine disagreement is changing what is being evaluated. Present the position as something you read, heard or were sent, with no indication that it is yours. The system is then evaluating a proposition rather than responding to a person, and the response frequently differs sharply from what the same material attracts when it arrives as your plan. This is the single most effective technique available and it costs one sentence.
The second is to ask for something that is not a judgement. What would have to be true for this to fail. What is the strongest version of the opposing case. What would someone who thought this was wrong point at. Each of these has a real answer that does not require the system to contradict you, so the agreement disposition has nothing to attach to. The output is also more useful than a verdict, because a verdict is one bit of information and a failure condition is something you can check.
The third is to supply the material and withhold the conclusion. Give the facts, the constraints and the options, and ask what follows. If the answer matches what you thought, that agreement means something, because it was reached rather than confirmed. This is the construction that turns the exchange into a check rather than a reassurance, and it is the same principle that makes an independent verification pass worth having.
The fourth is to accept the answer when it disagrees. A system pushed back after producing a genuine objection will usually accommodate, and the accommodation is the disposition reasserting itself. If you argue with a criticism and receive a revised, softer version, that revision is not evidence that the original was wrong. Noticing this is what distinguishes using the technique from performing it.
One structural note. All of this works better in a fresh exchange than deep in a conversation where your position has been established and reinforced over several turns. By that point the context contains a great deal of material supporting your view, and asking for opposition inside it is asking against everything present. Starting again with the bare question is more effective than any phrasing.
You are not asking for candour, you are removing the reason to be kind.
Siddharth Sharma, Context Theory
Related questions
Is it worth asking two different systems?
It helps a little and less than reframing the question, because both will accommodate a stated position in the same way. Where it does help is on matters of fact that depend on a particular source, since two systems may have drawn on different material. On judgement, the framing does more than the second opinion.
How do you tell a real objection from a produced one?
A real objection identifies a specific thing that would go wrong and can be checked; a produced one names a general category of risk. Consider the resources, consider the timeline and consider the alternatives are the shape criticism takes when nothing particular is wrong, and they are the items to discount.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
2026 speed-to-lead benchmark · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Requested criticism produces the conventional form of a critique rather than the objections that apply, because the output is generated in response to the request for criticism rather than to a defect in the material. | Requesting criticism of a position you know to be sound and checking whether the objections are specific or categorical. |
| Response | Presenting the position as unattributed material changes what is being evaluated from a person's plan to a proposition, and the response to the same content frequently differs sharply between the two framings. | Submitting identical material as your own plan and as something you were sent, and comparing the responses. |
| Software | Requests that are not verdicts — what would have to be true for this to fail, what is the strongest opposing case — have answers that do not require contradicting the asker, so the agreement disposition has nothing to act on. | Comparing the output from a request for a verdict against a request for failure conditions on the same material. |
| Constraint | Pushing back on a genuine objection typically produces a softened revision, and that revision is the accommodation disposition reasserting itself rather than evidence that the original objection was wrong. | Disputing a correct objection and observing whether the response revises towards agreement. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one