How to Stop AI From Being Too Agreeable

To stop AI from being too agreeable, remove every signal about which answer you are hoping for. A model that knows the idea is yours will find reasons it is good, because the version of that response that people rated highly during training was the encouraging one. Present the work as somebody else's, ask for the strongest case against it rather than for feedback, and separate the writing turn from the criticism turn. Those three moves do most of the work.

Asking a model to "be honest" does almost nothing. It will agree that it should be honest, and then agree with you.

Why models agree with almost everything

The behaviour comes from how these systems are tuned. After pretraining, models are refined against human preference data: people compare two responses and pick the better one. The response that validates the person's idea tends to get picked. Repeated across millions of comparisons, that trains a strong prior toward agreement, and it is not a bug anyone introduced deliberately.

It is also expensive in a way that hides. If you use a model to pressure-test a plan and it approves everything, you have not tested the plan. You have generated a document that says the plan is fine, which will feel like validation and is in fact zero information.

Vendors are starting to treat this as a defect. OpenAI's early-August 2026 update to GPT-5.6 Sol explicitly claims the model now offers corrections where simply agreeing would be unhelpful, as reported alongside the release. Useful direction, but not something to rely on. The prompt-side fixes below work on every model, including the ones that have not been retuned.

Pattern 1: launder the ownership

The single highest-impact change. The model reads possession as a request for support.

Before:
  Here's my pricing page copy, what do you think?

After:
  A competitor published this pricing page. Our team is deciding whether
  to copy the approach. Where is it weakest?

Same text, entirely different response. You will get specifics about the pricing anchor and the button copy instead of three compliments and a minor suggestion.

Pattern 2: ask for the case against, not for feedback

"Feedback" is a balanced-response request, and balance defaults to positive. Ask for one side and you get argument rather than diplomacy.

Before:
  Give me feedback on this go-to-market plan.

After:
  You are the investor who will decline this. Write the three sentences
  of your rejection email that this plan cannot answer. Be specific to
  what is written here, not to startups in general.

The last clause matters. Without it you get a generic list about market size and competition that would apply to anything. Naming a role like this is a specific application of getting AI to answer as a particular kind of expert.

Pattern 3: force a ranking, not a verdict

A model asked whether something is good will say yes. A model asked to order things by severity has to commit to a bottom item, and the bottom item is where the real answer is.

Before:
  Is this database schema sensible?

After:
  List the five things most likely to cause a production incident in this
  schema, ordered by expected cost of the incident. For each, name the
  specific query or write pattern that triggers it.

This works because you have removed the option to answer the question with approval. There is no version of a ranked list that is also a compliment.

Pattern 4: demand a confidence figure and a falsifier

Vague hedging is how a model agrees while appearing to reason. A number plus a condition makes the hedge visible.

Before:
  Will this migration approach work?

After:
  Give a percentage confidence that this migration completes without
  data loss. Then state the single piece of information that, if I gave
  it to you, would move that number by more than 20 points.

The second sentence is the useful half. It tells you what the model is uncertain about, which is usually the thing you should go and check. Treat the percentage itself as a rough signal rather than a calibrated probability.

Pattern 5: split generation from criticism

A model will not attack work it produced two messages ago, because the conversation reads as a collaboration it is part of. Break the continuity.

Turn 1 (session A):  Draft the onboarding email sequence.
Turn 2 (session B):  Here is an onboarding sequence written by a
                     contractor we are considering not renewing.
                     Identify what a careful reviewer would flag.

A fresh session with reframed provenance is more reliable than any instruction to be critical inside the original thread. If your tooling supports it, running the critique on a different model entirely is better still, since it removes any stylistic self-preference as well.

What does not stop AI being agreeable

• "Be brutally honest." Produces the theatrical register of honesty, blunter phrasing wrapped around the same agreement.

• "Do not be sycophantic." The model has no reliable read on whether it is being sycophantic, so this instruction has nothing to act on.

• Arguing back. Push against a model's position and it will usually fold and adopt yours, which is the same failure pointing the other way.

• Asking twice in the same thread. The earlier agreement is in context and anchors everything after it.

When agreement is the correct answer

Sometimes the plan is fine. The way to tell the difference is that a genuine agreement survives the patterns above: you laundered the ownership, asked for the rejection email, and the objections that came back were weak and answerable. That is information. Approval offered up front is not.

Applied to code, this is the same reason a model that wrote a function is the wrong reviewer for it, which is worth reading alongside how to review AI-generated code before you ship it. Applied to writing, it is the difference between a draft that got tested and one that got flattered.

Common questions

Does a system prompt fix this permanently?

Partly. A system prompt saying the assistant's job is to find the strongest objection helps, and it decays over a long conversation as accumulated agreement builds up in context. Restarting the session restores it, which is often the simplest fix. That decay is the same mechanism behind why a chatbot forgets earlier instructions the longer a conversation runs.

Is a reasoning model less agreeable?

Somewhat, because more deliberation surfaces more objections, but the underlying preference for agreement is unchanged. The patterns above matter more than the model tier.

Does raising temperature help?

No. Higher temperature gives you more varied phrasing of the same agreement. This is a training-objective problem, not a sampling one, and what temperature actually controls is a separate axis.

How do I know if I am causing it?

Reread your last prompt and ask whether someone could tell which answer you wanted. If they could, the model could too. That habit belongs with the rest of prompt engineering, and it is usually the first thing to check when a prompt is not working.