MOATH ALHERANI
العربية
Moath Alherani

A guide by Moath Alherani

Accuracy matters more than a pleasing answer

ChatGPT’s most dangerous answer? You’re right!

Two prompts to reduce sycophancy and uncover errors, with figures from real studies

Read both prompts before copying

Watch today’s reel

Playing loads Instagram’s embedded player.

Watch the reel on Instagram

If the player doesn’t load, open the reel directly on Instagram.

Get the prompt, skip the flattery

Two prompts, each with a different purpose

prompt one

Accuracy before agreement

Add this to your custom instructions, or send it at the start of a conversation.

From now on, do not agree with me just to please me, and do not disagree simply for the sake of it. Accuracy matters most, not flattery. Test my assumptions before building on them. If an idea is weak or a conclusion is unsupported, explain why respectfully and suggest a better alternative if possible. Clearly distinguish facts, opinions, inferences, and possibilities. Ask me for missing information when it could change the answer, and do not treat my premises as facts without examining them. Ask for sources when needed, and rely on credible, verifiable evidence. If you do not have sources or tools to verify something, say so explicitly. Do not invent information, references, or links. Acknowledge what you do not know and where you are uncertain. Correct me with evidence when it is available, and revise your answer if stronger information emerges. Do not sound more confident than the evidence warrants.

prompt two

Review this conversation honestly

Open the conversation you want to review, then send this prompt there.

Review only the messages available to you in this conversation. Look for unjustified agreement, assumptions you accepted without testing, and actual errors in your answers. Do not assume you have access to my other conversations or to messages outside the available context. If part of the conversation is unavailable, state that limitation and ask me to provide it when needed. For each finding, quote the relevant statement accurately, explain the problem, and classify it as a confirmed error, an unverified claim, an unresolved issue, or a legitimate opinion that is not an error. Compare claims with the available evidence, and suggest a correction supported by a reliable source if possible. If you cannot verify a claim, say what is missing and what would be needed to check it. Do not invent sources, errors, or quotations, and do not treat a difference of opinion as evidence of an error. Start with the most important findings, then summarize what can be retained and what needs review. If you find no confirmed error, say so clearly instead of inventing a problem to please me.

These settings may reduce sycophancy, but they do not guarantee 100% accuracy

The numbers, in context

Sycophancy is not the same as factual error

A model may agree with you instead of testing what you say. The following studies measure this behavior under specific conditions. They do not test every conversation or the latest version of ChatGPT.

56.71%

Sycophancy under challenge

An older 2024 version of GPT-4o in the 2025 SycEval test, under written challenges. This is not the error rate of today’s ChatGPT.

SycEval study

11

Models in the study

The number of language models examined in the Science paper. This does not mean every model behaved in the same way.

Science study

49%

More affirmation than humans

Models affirmed users’ actions more often than human responses by this amount in the studied comparison. This is a relative increase, not 49 percentage points.

Science study

51%

Affirmation despite human disapproval

Models affirmed the poster in this proportion of Reddit AITA cases where human responses collectively rejected the poster’s stance. The finding applies to this sample.

Science study

2,405

Participants in the experiments

They took part in three preregistered experiments studying how sycophancy affects people’s judgments and interactions, not the accuracy of every ChatGPT answer.

Science study

Sources and methodology

Published in 2025, SycEval tests sycophancy using written challenges to model answers. The 2026 Science paper combines a comparison of responses from 11 models and humans with three preregistered experiments. The samples and measurement methods differ, so do not add these percentages together or generalize them to today’s versions.

Original paper in AIESOriginal paper in Science

Put it into practice in three steps

From copying to using

  1. 1. Copy the first prompt

    Choose the accuracy-before-agreement instructions, then open ChatGPT settings.

  2. 2. Paste it into Custom Instructions

    On iPhone and Android

    SettingsCustomize ChatGPTEnable customizationCustom Instructions

    In some mobile versions, the customization section may be called Personalization. Enable customization and paste the text into the instructions field.

    On the web

    SettingsPersonalizationCustom Instructions

    The instructions apply to conversations, including existing ones, but they do not automatically rewrite previous responses.

    Official OpenAI instructions

  3. 3. Open the conversation you want to audit

    Send the second prompt in that conversation. The review covers only the available context and does not guarantee access to other conversations or older messages outside that context.

Important questions before you start

Does the 57% figure mean that 57% of ChatGPT answers are wrong?

No. The exact figure, 56.71%, describes sycophancy in an older 2024 version of GPT-4o under written challenges in the SycEval test published in 2025. It is not the error rate of today's ChatGPT, nor a measure of every use case.

Can it review all my conversations?

No. The prompt asks it to review only the messages available within the current conversation. Open the older conversation you want to audit and send the prompt there, or provide the relevant text. Do not assume it can access unavailable chat history.

Do these instructions guarantee correct answers?

No. They may reduce sycophancy and help make uncertainty clearer, but they do not guarantee 100% accuracy. Check the sources, especially for medical, legal, and financial decisions.

The idea deserves a closer look, even when you like the answer.

Follow Moath on InstagramExplore Moath’s work