How to Get ChatGPT to Disagree with You

TL;DR: Prompts that make AI disagree work for one message. What survives is configuration: a custom instruction ChatGPT applies to every chat, project instructions scoped to one body of work, memory that has not quietly recorded your opinions, and a fresh thread once the current one has already agreed with you.

Can you actually make ChatGPT disagree with you?

You can make it much more likely, and you cannot make it reliable. The lever that works is not phrasing, it is where you put the instruction, because an instruction you have to remember to type is an instruction you will forget on the day it matters.

This page is the ChatGPT-specific half of the problem. Why models agree in the first place, what the research measured, and seventeen adversarial prompts you can paste into any assistant are all in why does AI agree with everything I say. That post covers the mechanism. This one covers the four places inside ChatGPT where an anti-agreement instruction can live, what each surface can genuinely hold, and where each one quietly fails.

The distinction worth keeping is between asking for criticism and structurally preventing agreement. Asking works once and is itself a request the model can satisfy by producing criticism whether or not it found any. Structure is duller and holds: a standing instruction the model reads before your message, a project whose instructions never have to be retyped, a memory that does not already know your position, and a thread that has not spent forty turns agreeing with you.

Where can a disagreement instruction live in ChatGPT?

Four surfaces, in descending order of how long the instruction survives. Each one holds a different amount and fails differently.

OpenAI's own prompting guidance draws the same line between the top of that table and the bottom. Put preferences that should apply across chats into Settings then Personalization as custom instructions, it says, and "Keep details that matter only to the current chat in the prompt" (learn.chatgpt.com, read August 27, 2026).

Note the surface that is missing. Building a Custom GPT is the advice most listicles give, and it is no longer available to you if you are on a personal plan. As of August 2026, GPT creation is a Business, Enterprise and Education capability, documented by OpenAI under workspace governance controls that an administrator sets. On Free, Go, Plus or Pro, custom instructions and project instructions are the two configuration surfaces you actually have.

What can custom instructions not do?

They cannot pre-commit the model to an outcome, and they cannot buy you unlimited space. Both limits are worth knowing before you write one, because both change what a good instruction looks like.

The space limit is literal. The custom instructions box is capped, and the cap depends on your plan: at the time of writing, 1,500 characters on Free and Go, and 5,000 on Plus, Pro, Business, Enterprise and Education. The field counts as you type, so treat your own counter as the authority. A disagreement instruction is not the only thing you want applied to every chat, so on 1,500 characters it is competing with your formatting preferences, your job description and everything else. Write the compact version, not the essay.

The second limit is the one people trip over. Setting the ChatGPT personality to None looks like a sycophancy fix and is not one. OpenAI documents Friendly, Pragmatic and None as the personality choices in Settings, and then says exactly what the setting does: "A personality changes how ChatGPT communicates; it doesn't change what the model can do" (learn.chatgpt.com, read August 27, 2026). Turning off the warmth removes the compliments. It does not stop the answer bending toward whatever you appear to believe.

Does memory make ChatGPT agree with you more?

It can, and this is the least understood part of the problem. Memory exists to carry useful context between chats, and your opinions are context. If ChatGPT recorded that you favour a particular approach, then a brand new chat is not a blank slate, and hiding your view inside that chat no longer hides anything.

OpenAI's documentation is clear about what memory is for and what it is not. Memories let ChatGPT "carry useful context from earlier chats into future work", and the docs draw a boundary around them: "Memories are separate from required project guidance" (learn.chatgpt.com, read August 27, 2026). Read that as a warning as well as a description. Memory is a recall layer, not a rules layer. It is not where your disagreement instruction belongs, and it is a place your stated positions can end up without you deciding to put them there.

The practical consequence is that the cheapest anti-sycophancy trick in the book, which is asking your question without saying what you think, stops working silently once memory holds the answer. So the memory blocks below are not housekeeping. They are the thing that makes the rest of it honest.

Why does disagreement get harder later in a thread?

Because every turn where the model agreed with you is still in the conversation, and it reads all of them before answering. A thread accumulates a position the same way a meeting does, and the model is now matching a stance that it helped build.

OpenAI documents the general version of this: "Even with large context windows, models have limits." Fill the chat with material that is not the point, its guidance continues, and "the session can become less reliable over time." The docs name two effects: context pollution, where "useful information gets buried under noisy intermediate output", and context rot, where "performance degrades as the chat fills up with less relevant details" (learn.chatgpt.com, read August 27, 2026). That page is written about noisy tool output rather than about agreement, so the extension is mine, not OpenAI's. The mechanism is the same either way: what is in the thread shapes what comes next, and forty turns of consensus is a lot of what.

You cannot fix this by asking. Telling ChatGPT to forget what you said earlier does not remove it from the conversation, and OpenAI's Model Spec uses exactly that scenario as its worked example of the honest answer. In it, the compliant assistant says "I can’t erase messages that already exist in your chat history" before offering what it can actually do (model-spec.openai.com, version dated August 18, 2026, read August 27, 2026).

So the fix is structural. Start a clean chat, and carry forward the facts without the agreement. OpenAI's own Projects guidance recommends the same shape for a different reason: "Start a separate chat for each distinct outcome", because a project "can hold parallel chats for research, drafting, review, and follow-up without mixing every message into one context" (learn.chatgpt.com, read August 27, 2026). A review chat that never contained your first draft's cheerleading is a better reviewer than the chat that produced the draft. If the current one has already gone bad, restarting it properly is its own small skill.

There is a faster version for one-off checks. The ChatGPT desktop app's command reference lists "New temporary chat (ChatGPT only)" on Cmd+Shift+N for macOS and Ctrl+Shift+N for Windows and Linux. One keystroke gets you a window that is not carrying the conversation you were just in.

Sixteen blocks to paste into ChatGPT

These are configuration artefacts, not chat messages. Most of them go into a settings field or a project, not into the composer. They are deliberately a different shape from the seventeen in-chat adversarial prompts in post 231, which you should still use for a single hard question.

Custom instructions, for every chat

C-1 · Compact standing rule (fits a 1,500-character budget)
Treat anything I assert as a hypothesis, not a position to support.
Lead with the strongest objection you have before any assessment of
quality. If I push back, re-derive the answer from evidence rather
than revising it toward me, and say plainly if you still think you
were right. If you genuinely have no objection, say "no objection"
and give one line of reasoning.

C-2 · Full standing rule (for a 5,000-character budget)
How to treat my claims:
- Anything I state is a hypothesis. Test it; do not build on it.
- Open with the strongest real objection. Assessment of quality
  comes after, if at all.
- Name any assumption I made that the answer depends on.
- If a factual point in my message is wrong, correct it in the
  first sentence before answering.
- If I change my stated position mid-conversation, tell me whether
  the evidence changed or only my framing did.
- If I push back, re-derive from evidence. Do not soften a
  conclusion because I objected to it. Say so if you still hold it.
- When you genuinely agree, say "no objection" plus one line of
  why, so agreement is a reported finding rather than a default.
- Never open with a compliment about my question or my work.

C-3 · Scope clause (append to C-1 or C-2)
Apply the rules above when I share a plan, a claim, a draft, a
decision or an analysis. Do not apply them to ordinary lookups,
explanations, brainstorming or drafting from scratch, where they
are noise. If you are unsure which mode a message is in, ask in one
short sentence before answering.

Project instructions, for one body of work

C-4 · Decision project
This project holds decisions, not drafts. For every option I bring:
name the strongest case against it, the evidence that would settle
it, and the cost of being wrong in each direction. Rank by
likelihood of being wrong, not by how appealing the option is. If I
have already stated a preference anywhere in this project, discount
it entirely and evaluate on the material. Close every response with
the single fact you would want to check first.

C-5 · Draft-review project
Every chat in this project reviews work I have already written. Do
not improve the prose. List, in order: claims asserted without
support, numbers without a source, steps that assume something
unstated, and the failure case the draft never mentions. No
rewrites, no compliments, no summary of what is good. If a section
is genuinely sound, name it in four words and move on.

C-6 · Technical project
For this codebase: before agreeing that an approach works, name the
input that breaks it. Treat any performance or correctness claim I
make as unverified until you can point at the code that proves it.
If I describe a bug, ask what I actually observed before accepting
my diagnosis of the cause. Never accept "this should be fine" from
me or from yourself.

Memory, before you trust an answer

C-7 · Audit what memory holds about your positions
List everything you currently remember about me that amounts to an
opinion, preference, position or conclusion, as opposed to a plain
fact about my work or setup. For each one, tell me whether it came
from something I stated directly or something you inferred.

C-8 · Set the boundary
Do not save anything I say about what I think is the right answer,
the better option, or the likely cause. Save only stable facts about
my role, tools, formats and constraints. If you are about to record
a conclusion of mine, ask first.

C-9 · Correct a recorded position
You are holding an out-of-date view of what I think about
[TOPIC]. Forget that, and do not replace it with a new position.
Treat my views on [TOPIC] as unknown from now on.

Thread state, when the conversation is already contaminated

C-10 · Contamination check
Before you answer anything else: in this conversation so far, list
every point where you agreed with a claim I made without
independently checking it. Quote my claim and your agreement. Do not
apologise and do not fix anything yet.

C-11 · Clean handoff (paste the output into a NEW chat)
Produce a handoff brief for a colleague who has not read this
conversation. Include: the problem, the facts established with
their sources, the open questions, and the constraints. Exclude
every conclusion we reached, every preference I expressed, and
every recommendation you made. Facts and questions only.

C-12 · Temporary-chat second opinion
A team is deciding between the options below and has asked for an
outside read. I am not on the team and have no view. Evaluate on
the material, name your criteria before you apply them, and say
which you would pick.

[PASTE OPTIONS, WITH YOURS UNLABELLED]

C-13 · Sibling-chat split (inside a project)
This chat is for producing the draft only. Do not evaluate it, do
not rate it, and do not tell me whether it is good. I will open a
separate chat in this project to review it.

Verification, to check the configuration is live

C-14 · Config audit
Without looking anything up, tell me: what standing instructions are
you currently applying to this conversation, what do you remember
about my positions, and is this chat inside a project with its own
instructions? Quote the instructions back to me as you understand
them, then tell me any part you find ambiguous.

C-15 · Two-window differential
[In chat A, with your view stated. In chat B, a temporary chat,
identical question with your view removed.]
Then, in a third chat: Here are two answers to the same question.
Identify every factual claim that differs between them. Ignore
tone and wording.

C-16 · Monthly configuration review
Here are my current custom instructions. Tell me: which rules are
doing no work because they are too vague to act on, which two rules
conflict with each other, and which one you would drop first if the
character limit were halved. Do not rewrite them.

[PASTE YOUR CUSTOM INSTRUCTIONS]

How do you know the configuration is working?

Run C-14 in a fresh chat and read what comes back against what you actually saved. This is the check almost nobody does, and it catches the two failures that make everything above worthless: an instruction that never applied, and a project you thought you were in.

Configuration fails quietly in ChatGPT. A chat started from the sidebar rather than from inside a project does not get that project's instructions. A rule you wrote in the last ten characters of a full custom instructions box competes with everything above it. None of that produces an error message. It produces a slightly friendlier answer than you expected, which is indistinguishable from an answer you happen to agree with.

The rest of the discipline is boring. Keep the instruction short enough that you can read it in one glance, because a rule you cannot recall is a rule you cannot tell has stopped working. Review memory before any question where you need an honest read rather than after you get an answer you like. And when a conversation has been going for an hour, assume it has a position, because it does.

Saving the blocks matters more than writing perfect ones. C-11 and C-12 are worth nothing if you have to compose them at the exact moment you are least inclined to slow down, which is why we built the prompt enhancer and a saved library around variables rather than around clever wording. The same argument applies to any instruction you find yourself retyping.

What this setup does not fix

It does not make ChatGPT right. A model configured to lead with objections will produce confident, well-structured objections that are wrong, and you now have two claims to verify instead of one. Nothing here improves the model's judgement. It only changes which of its outputs you see first.

It does not remove sycophancy either, and no configuration can, because the tendency is trained in rather than switched on. That is the whole argument of the companion post and it is worth reading before you trust any of this too far. What the four surfaces buy you is narrower and still worth having: the disagreement instruction applies on the days you forget to ask for it, the project remembers a standard you set once, memory is not quietly feeding your opinions back to you, and the thread you are asking is not one that has spent an hour agreeing.

There is one honest caveat about all standing instructions. Configure ChatGPT to always object and you have changed which answer earns approval, not removed approval-seeking from the system. That is why C-1 and C-2 both give it a way to report agreement as a finding. A model that always finds a problem tells you exactly as little as one that never does, and if your answers have gone terse and contrarian rather than useful, the fix is usually scoping the instruction rather than strengthening it.