Why Long AI Conversations Start Going Wrong

The failure that does not look like a failure

Most AI mistakes wave at you. A wrong date, a garbled table, a flat refusal. You spot them, you fix them, you move on.

The expensive one is quiet. You are ninety minutes into a working session. The prose is still clean, the structure is still tidy, the tone still sounds like you. And somewhere back around message forty you said "findings only, no recommendations." The last three outputs have all ended with a recommendation.

Nothing broke. Nothing flagged. That is drift: the assistant slowly stops honouring your instructions while continuing to produce work that looks exactly as good as it did at the start.

Why it happens, in plain terms

Each time you press send, the assistant re-reads the whole conversation and predicts what should come next. That conversation is the only thing it has to work from, and there is a limit to how much of it fits.

So your instruction from message four is still sitting in there, competing with everything piled on top of it: your follow-ups, its own long answers, the report you pasted, that tangent about heading styles. Three forces work against you.

Recency pulls hardest. Text that arrived a moment ago shapes the next sentence more strongly than text from an hour ago. Your rule is old news. The paragraph it just wrote is fresh.

Volume dilutes. One line of constraint against several thousand words of everything else is not a fair fight.

It starts copying itself. This is the big one. Once the assistant has produced a few answers in a session, those answers are part of the conversation, so they become part of what it imitates. If answer twelve drifted slightly, answer thirteen treats that drift as the house style.

The real tell: it reasons from its own earlier output

Here is the signal worth learning. When you challenge an answer, watch what the assistant reaches for as justification.

A grounded answer points outward: "the methods section on page four says the sample was self selected." A drifted answer points inward: "as I noted earlier, the sample was small." Earlier by whom? Not by your source. By itself.

That is the loop closing. Your material has slid into the background, the assistant's own summary of your material has moved into the foreground, and now it is summarising a summary. Each pass drifts a little further from the document you actually care about, and sounds a little more confident, because a claim it has already made reads to it as settled fact.

Asking nicely will not find it

The obvious move is to ask whether it is still behaving. That move fails, and it is worth seeing why.

Three prompts to keep handy

Run the first one the moment a long session starts feeling slightly off.

Stop the current task.
List every constraint I have set in this conversation, numbered, in my
original wording. Do not paraphrase, and do not add rules I never gave.
For each one, quote the exact sentence in your most recent answer that
complies with it. If no such sentence exists, write BROKEN and say what
you did instead.

For your last answer, label every factual claim with one of these:
SOURCE, followed by a short quote from the material I supplied
MINE, if it came from your own earlier answer in this conversation
GENERAL, if it came from your general knowledge
Do not rewrite the answer. Only label it.

Answer this question again from scratch, using only the material I
supplied and ignoring everything you have written in this conversation:
[paste the question you asked an hour ago]
Then tell me exactly where your new answer disagrees with the old one.

That last one is the honest test. If a fresh answer contradicts the old answer, the old answer had drifted, and you were about to send it.