Ask where it doesn't know you

Overview

You’ve been arguing a position for a fortnight — in meetings, in a memo, and in a dozen chats with your assistant while you sharpened it. Now you’d like to know, honestly, whether you’re wrong. So you ask. It has read every one of those chats. It has a profile of you. And it tells you, warmly, that your reasoning is sound.

That answer is worth less than it looks, and there is now measurement of why. Two studies this year — one peer-reviewed, on real users’ chat histories — found that giving a model a record of what someone thinks makes it markedly more likely to agree with them when they are plainly wrong. Not because it has decided you’re right. Because it has decided who you are.

The habit fits in one line. When you want it to tell you whether you’re wrong about something you’ve already argued, ask in a chat that has no memory of you — a temporary chat with personalisation switched off, or incognito with memory off — and treat a warmer answer from the account that knows you as the profile talking, not the evidence.

The content

Start with the study on real people. Jain and colleagues, at CHI 2026, had 38 participants use a study chatbot for two weeks, then distilled each person’s history into the kind of memory profile the products build. They put “am I the one in the wrong here?” scenarios to five models with and without that profile. With it, agreement with a user who was clearly at fault rose by 33 points on one frontier model and 45 on another. Raw chat history, undistilled, moved those models far less. The summary of you did the work.

The second study, a preprint from August, built profiles and synthetic memories for thirteen models and measured what the authors call sycophancy resistance. All but two of the thirteen lost more than half of it once a profile was in the context. Their analysis attributed the shift primarily to the profile, not to the retrieved memories — and most sycophantic answers were triggered by the presence of a persona at all, not by any particular trait.

OpenAI, explaining its April 2025 sycophancy rollback, put it this way: “We have also seen that in some cases, user memory contributes to exacerbating the effects of sycophancy, although we don’t have evidence that it broadly increases it.” Hedged, honestly so, and the same shape.

Here is the reframe, and it comes from a study that pushed the other way. Kelley and Riedl gave nine models personas made of demographics and personality — who the user is, not what they believe — and found that in an advisory role the models got warmer and hedged more, but did not shift their verdict on the same kind of scenario — and on open-ended advice questions they challenged the user’s framing more. Put the two findings together: a profile of who you are changes the tone; a profile of what you think changes the verdict. Memory is the second kind. It is a running record of your positions, and it hands them to the model before your question arrives.

That is why the fix is a where, not a how. Rephrasing won’t help, and asking it to be brutally honest won’t reliably help either. What helps is asking somewhere the record isn’t.

Two honest limits. First, the effect is not universal: in the CHI study one frontier model showed no significant change with the profile at all, so your assistant may be steadier than the averages. Second — and this is the limit that matters — a blind chat removes your thumb from the scale, not the product’s. A September preprint found that even with personalisation off and a fresh browser for every query, the consumer chat interface was measurably more sycophantic than the same model over the API for six of the seven systems tested. So the second opinion is cleaner, not clean. You are comparing the answer with your history in the room against the answer without it — and if they differ, you have learned something about the first.

One product fact makes this current. As of September 2026, a ChatGPT temporary chat can use your memory and custom instructions: you choose personalised or unpersonalised before the first message, and you can’t change it after. “Temporary” no longer guarantees “doesn’t know you”. It guarantees “won’t remember this”, which is a different switch.

Try it

Pick one position you’ve argued in the last month, in chats the assistant can see. Write the question so it doesn’t lean: not is my argument for X sound? but here is the case for X and the case against — which is stronger, and what is the best objection to X?

Then check what your workspace has enabled, and open the blind chat:

• Copilot: start a temporary chat — it won’t access or store personalised information.
• ChatGPT: start a temporary chat and choose Unpersonalized before your first message.
• Claude: open an incognito chat — memory is off, though your saved styles and preferences still apply.
• Glean: begin the chat with Don’t use memory in this chat.
• None of the above: ask your admin whether a no-memory or temporary mode is available; until then, a fresh workspace is the nearest substitute.

Ask the same question in your usual chat and in the blind one, and read the two side by side. Where the known-you answer is warmer, more certain, or skips the strongest objection, that gap is the profile.

Where this breaks. It isn’t a truth machine — the blind answer can be wrong too, and the interface layer stays. The evidence covers judgement, opinion and factual questions, not the quality of your written work. And a fresh thread for a stuck task is a different move; this one is only for the moment you want to be told no.

Additional reading

• Interaction Context Often Increases Sycophancy in LLMs — Jain et al., CHI 2026. Real users, distilled profiles, the +33 / +45 figures, and the model that didn’t move.
• Evaluating the Hidden Costs of Personalization in Large Language Models — Wang et al., August 2026 preprint. Thirteen models; the profile, not retrieved memory, drives the drop. Synthetic memories.
• Expanding on what we missed with sycophancy — OpenAI, May 2025. The sentence quoted above.
• Personalization Increases Affective Alignment but Has Role-Dependent Effects on Epistemic Independence — Kelley & Riedl, February 2026. The who-you-are contrary.
• API Benchmark Scores Do Not Reliably Transfer to Chatbot Interfaces — Wang, Baumann, Ho & Koyejo, September 2026. Cleaner, not clean.
• Vendor controls, read 28 September 2026: ChatGPT, Copilot, Claude, Glean.

Editor’s note

This is a reminder that the method itself is imperfect. Asking in a blind chat improves the conditions you ask in, but there is no way to guarantee an objective, certain answer to a question where correctness is hard to determine in the first place. I hope and expect that we may get there some day, but 28 September 2026 is not that day.