Some iteration converges and some polishes its own output while appearing to improve. The difference is whether new information enters on each pass — and the four loops sorted by which it is.
- Loops improve when something new arrives. Your own judgement, a test, a dataset or a real result gives each pass something to correct against.
- Self-critique does not reliably fix facts. It can help with format and structure, but the review comes from the same picture that produced the error.
- Asking again is unreliable. In the study cited here, pushback corrected answers more often than it broke them, but it did turn some right answers wrong, and the model sounds just as confident either way.
- Output fed back in drifts toward generic. Specificity is the first thing lost when nothing replenishes it.
- Start fresh, cap the passes, judge against the brief. Ask the model to check rather than to improve.
A "loop" here just means going back and forth with an AI to improve something, which is the iterative workflow seen from underneath.
The useful test: are you adding anything new each time you go round? If you bring a fact, a correction, or a real result, it gets better. If you just keep asking again, it is rearranging the same material and it will drift.
Some loops improve with each pass. Others get worse while appearing to improve, and the difference is whether new information enters on each turn.
Does anything from outside the conversation arrive on this pass? A source, a measurement, a human judgement, a real result. If yes, the loop can converge. If no, it is polishing its own output and calling it progress.
Four loops, two of which cannot be trusted
Refinement — converges
You judge the output against something real and say what is wrong. Your judgement is the new information. This is the loop that works, and it works because you are in it.
Grounded iteration — converges
Each pass runs against a test, a dataset, a rendered result. Reality supplies the correction. This is why, in this site’s experience, coding loops with a test suite behave better than writing loops without one.
Self-critique — degrades
Asking a model to review its own output. It will usually find something to change — but the critique is generated from the same picture that produced the error.
It helps for format and structure. It does not reliably help for correctness, because nothing new arrived.Huang, Chen, Mishra, Zheng, Yu, Song & Zhou (Google DeepMind), Large Language Models Cannot Self-Correct Reasoning Yet, ICLR 2024, read at source 23 Sep 2026: “LLMs struggle to self-correct their responses without external feedback, and at times, their performance even degrades after self-correction.” The study tested reasoning tasks; that self-review helps with format and structure is this site’s experience, not something the paper measured.
Reprompting the same request — unreliable
Not liking an output and asking again, slightly differently. Each pass carries the previous attempts in context, so the conversation is now anchored on the thing you rejected.
Worse: pushing back can produce a revised answer that is less correct. Sycophantic revision from a right answer to a wrong one occurred in 14.66% of tested cases.SycEval, arXiv 2502.08177, Stanford, 2025, as read for how AI influences you on 10 Sep 2026: “regressive sycophancy, leading to incorrect answers, was observed in 14.66%”.
The same study cuts the other way too, and this page previously left that out. Revision toward a correct answer — progressive sycophancy — occurred in 43.52% of cases, roughly three times more often. So reprompting is not simply destructive: on balance it corrected more than it broke. What it is not is reliable. In roughly one tested case in seven, a right answer was revised to a wrong one, and nothing in the exchange tells you which kind you just got — the model sounds equally confident either way. That is the case against leaning on it, not a claim that it always degrades.SycEval, arXiv 2502.08177, Stanford, re-read 10 Sep 2026: “regressive sycophancy, leading to incorrect answers, was observed in 14.66%” and “progressive sycophancy, leading to correct answers, occurred in 43.52% of cases”. Both figures added 25 Aug 2026; an earlier version of this page cited only the first. More on how AI influences you back.
The compounding version
The same failure at scale: output becomes input becomes output. A summary of a summary, a rewrite of a rewrite, a dataset containing yesterday's generations.
Every pass without outside information moves the result toward the model's average and away from the specific thing you wanted.
This is why long chains of AI steps drift toward generic — not because any single step failed, but because specificity is the first thing lost when nothing replenishes it.Shumailov et al., The Curse of Recursion: Training on Generated Data Makes Models Forget, 2023, revised 2024, read at source 23 Sep 2026: “use of model-generated content in training causes irreversible defects in the resulting models, where tails of the original content distribution disappear.” That finding is about training on generated data, the dataset case above. For a chain of prompts inside one task, the drift toward generic is this site’s reasoning (September 2026), not a measurement.
Working rules
- Start fresh rather than asking again with nothing new. A new conversation has no rejected attempts anchoring it (if your tool’s memory is on, use a temporary chat). Specific notes on a nearly-right draft are different: that is editing, and it works in the same chat, as how to ask AI well explains.
- Bring something in on every pass — a source, a constraint, a real result. If you cannot, stop looping.
- Cap the passes before you begin. Pick the number in advance; past it you are probably polishing.
- Ask it to check, not to improve. A check gives the model less room to drift than a rewrite. A formal result points the same way: a model’s error rate when generating has a floor of roughly twice its error rate when judging whether an output is valid. That floor says generating is harder; it does not measure by how much.Kalai, Nachum, Vempala & Zhang, Why Language Models Hallucinate, OpenAI, Sep 2025, §1.1 and Corollary 1, read at source 17 Sep 2026: “(generative error rate) ≳ 2·(IIV misclassification rate)”. A formal lower bound, not an empirical measurement. Corrected 17 Sep 2026: this line put a size on the gap (about two times), cited Theorem 3, and quoted a sentence that is not in the paper.
- Judge against the brief, not the previous version. Comparing to the last attempt pulls you toward the last attempt.