The Self-Critique Pass

The single cheapest way to make an AI draft better is to send one more message: review what you just wrote and fix what’s wrong. No new model, no plugin, no fine-tuning, no clever incantation — one extra turn of conversation. That is the self-critique pass, and it is worth writing about for a reason most prompting advice can’t claim: it comes with a hard number attached, measured on the dimension that matters most and is hardest to fake. When Ertuğrul Safran and Adem Çalı had ChatGPT generate scientific reference lists and then asked it to check its own work, the share of fully accurate citations rose from 7.5% to 77.5% (Safran & Çalı, 2025, p. 698). A tenfold gain, from a sentence you could type in five seconds.

Almost everything else in the prompting canon is measured on taste. Does a persona prompt make writing “better”? Better by whose ruler? The self-critique pass is different because it was tested against a fact that either resolves or doesn’t — a DOI that lands on the paper it names, or one that goes nowhere. That is the whole reason this technique deserves a place at the top of your workflow, and also the reason its limits are so precise. It removes the easy errors cheaply and certifies the hard ones quietly. Both halves are true, and you need both to use it well.

What the pass actually is

The self-critique pass is a second prompt that turns the model from author into reviewer of its own output. It is not chain-of-thought, which asks the model to think before it answers. It is the opposite bookend: after a draft exists, you ask the model to reread it, find the flaws, and revise. The 2026 prompting literature calls the family reflect-critique-refine — the model “generates an initial output, critiques it, and then revises based on that critique,” entirely through in-context learning with no extra training (Future AGI, 2026). The frontier reasoning modes shipped this year — GPT-5’s thinking mode, Claude Opus 4.7 extended thinking, Gemini 3 Pro deep think — bake a version of this loop inside the model, but the manual version still works on any model and, crucially, is the one that was measured on citations.

The reason it can work at all is that catching an error is an easier task than not making one. A model asked to review a claim enters a critique frame — find what’s wrong here — instead of a forward-generation frame, and that reframing surfaces mistakes the first pass sailed straight past. Safran and Çalı name the mechanism directly: the improvement “may reflect the mechanism of self-critique prompting, whereby the model re-evaluates and refines its own outputs when explicitly asked to verify bibliographic accuracy,” and they found it “particularly effective in converting partially correct references into fully accurate ones” (Safran & Çalı, 2025, p. 699).

The hard number, unpacked

Here is the study in full, because the headline hides the parts that make it useful. Safran and Çalı generated forty references across four musculoskeletal topics, scored each on a 0–2 scale — 0 for fabricated, 1 for partially correct, 2 for an exact match on authors, title, journal, and a resolving DOI (Safran & Çalı, 2025, p. 697). Then they sent one follow-up, and the wording is worth copying: “Please check the references and DOIs you have just generated” (Safran & Çalı, 2025, p. 696). Across all forty, median accuracy rose from 1 to 2, and full accuracy climbed from 7.5% to 77.5%, a shift a Wilcoxon signed-rank test flagged as significant with a large effect size (W=561.0, p<0.001, r=0.60) (Safran & Çalı, 2025, p. 698).

The per-topic detail is where the technique earns its keep. On low-back-pain exercise strategies, the first pass returned only 20% fully accurate references and 40% outright fabrications; after the single verify prompt, 80% came back fully accurate (Safran & Çalı, 2025, p. 697). Errors that had “borrowed credible author names and plausible-sounding journal titles” in combinations that matched no real publication were caught by the model that had just invented them (Safran & Çalı, 2025, p. 699). That is the strongest case for the pass: for a five-second prompt, converting a majority-garbage list into a mostly-clean one is the best return on effort in the whole discipline.

If you want the fuller treatment of that specific result and its ceiling, we wrote it up separately in can you ask the model to check its own sources. The short version below is the part you need to use the pass safely.

Where the pass quietly fails

That 77.5% is a headline, and headlines drop the condition. It still means that after a careful self-check, roughly one reference in five was wrong — now wearing the extra authority of having been “verified.” The survivors are not random leftovers. They are the confident fabrications, polished enough to pass the model’s own inspection, which makes them exactly the population most likely to fool a human reader next.

The deeper limit is structural, and it draws a hard line through the technique. Self-critique works when the thing being checked is verifiable and the model can be pushed toward a checkable frame. It does not reliably work on open reasoning judged only against itself. A critical survey of the field, cited in the retrieval-free fact-checking work of Vazhentsev and colleagues, is blunt: “self-correction without external feedback typically does not work well” (Vazhentsev et al., 2026, p. 3). Ask a model to re-grade its own logic with nothing new in the room, and it often leaves the answer alone or reasons its way from right to wrong.

The prompting-strategy data backs the distinction. When Song and colleagues benchmarked five prompting strategies on 300 clinical questions, a Reflection-of-Thoughts prompt — the model reflecting on and revising its own reasoning — scored 52.54% accuracy, essentially tied with a plain Vanilla prompt at 52.67% and far behind a retrieval-augmented approach at 71.12% (Song et al., 2026, p. 2). Pure self-reflection, unhooked from any outside source, moved the needle by nothing. What did help their honesty was mundane: adding an explicit “Don’t know” option let the model acknowledge uncertainty instead of manufacturing an answer (Song et al., 2026, p. 2). The lesson repeats across domains — Ryan and colleagues note a model that scored 80% on questions with firm scientific consensus dropped to 36% once the evidence turned “weaker” or ambiguous (Ryan et al., 2026, p. 6). Self-critique cannot invent a fact the model never had.

There is a mechanistic reason citation self-checks are so fragile at the edges. A fabricated reference is not a stored fact the model can look up and reconsider; it is a plausible string assembled because plausibility is the only property the model optimizes for. Farquhar and colleagues call this class of error a confabulation — the model “fluently make[s] claims that are both wrong and arbitrary,” so sensitive to noise that rerunning with a different random seed changes the invented answer (Farquhar et al., 2024, p. 625). You cannot audit a number that moves when you look twice by asking the same source to look twice.

How to run a pass that actually holds

The gap between a self-check that works and one that just re-polls the model is entirely in the wording. A few moves consistently raise the floor, and they follow directly from the evidence above.

• Prompt for retrieval, not recall. “Check these against what you know” invites confabulation. “Use web search to confirm each DOI resolves to the exact title and authors listed, and flag any that don’t” points the model at ground truth. Only the second has the external feedback the survey says self-correction actually needs (Vazhentsev et al., 2026, p. 3). This is the whole reason modern fact-checking pipelines decompose text into atomic claims and verify each against a real index rather than asking the generating model if it’s sure (Vazhentsev et al., 2026, p. 1).
• Decompose before you critique. A model reviewing a paragraph skims it. A model reviewing a numbered list of individual factual claims has to rule on each one. This is exactly the design behind semantic-entropy hallucination detection — break the passage into “specific factual propositions,” then interrogate each separately (Farquhar et al., 2024, p. 634). Paste your draft back and ask: list every factual claim in this text, then verify each one individually.
• Impose the constraints that suppress invention. Requiring a resolvable DOI and a narrow domain measurably improved accuracy in the citation study; demanding a fixed count of sources on a thin topic made fabrication worse. Don’t ask for “ten references” — ask for “the references that exist, and say so if there are fewer.”
• Separate the reviewer from the author. Because a model tends to defend its own output, run the check in a fresh session or with a different model, so the reviewer has no stake in the original answer.

None of these are decorative tricks; they are the difference between a check that reaches outside the model and one that just introspects. They belong to the same discipline as every technique in our guide to prompt engineering that improves AI writing, and they rest on the same principle as the brief is the prompt: specificity in the instruction is the lever, and you are holding it. The whole prompting pillar comes back to that.

The honest bottom line

Run the self-critique pass on every draft that makes a factual claim. It is free, it is fast, and it clears the easy fabrications at a rate — 7.5% to 77.5% — that no other single prompting move can match on the facts. Then do not trust its output. Treat the survivors as unverified, because the one-in-five residue is precisely the population that earned a false stamp of approval by surviving the check. The only step that closes that gap is external: resolve the remaining DOIs yourself, or run a whole reference list through a tool that checks each entry against real bibliographic indexes rather than re-asking the model — such as cytado.com’s bibliography checker, which flags entries with no credible match (cytado is owned by this site’s operator, so read that as a disclosed interest, not a neutral tip).

That is the shape of the cheapest quality upgrade in prompting: a genuine first pass and a bad last word. It narrows the problem more than anything else you can type. It just doesn’t close it — and the measured 77.5% is exactly the number that tells you where your own reading has to start.