Adversarial review: why one AI checking another beats asking twice

Ask an AI a question, get a confident answer, ask "are you sure?" — and watch it apologize and change its mind, whether or not it was wrong the first time. That little ritual is why "just ask it to double-check" doesn't work as a quality strategy, and it's the problem adversarial review actually solves.

The idea is old and very human: writers don't proofread their own drafts well, so someone else does. The AI version pits two roles against each other — one plans the answer, another one's whole job is to find fault with the plan — and only an approved plan gets executed.

one-chat.app — second opinion mode
everyday AIsecond opinion on
Plan flights and hotels for Japan in April — best value.
Draft plan — compare fares, book the cheapest week, pick hotels by rating…
 Reviewer pushed back — that's peak cherry-blossom week. Prices double. Check the shoulder dates and rail passes.
Revised — shifted one week, added rail pass, re-ranked hotels by location + breakfast…
 Approved after 2 rounds
Fly April 15–24 (one week after peak — same blossoms in the north, 38% cheaper). Here's the full plan with a rail pass that pays for itself by day 3.

Why models can't just check themselves

The obvious cheap fix — append "review your answer and fix any errors" — has been tested thoroughly, and the results are humbling. A widely-cited study titled Large Language Models Cannot Self-Correct Reasoning Yet found that without external feedback, models "struggle to self-correct their responses… and at times, their performance even degrades." The model that made the mistake usually can't see it, because the same blind spot that produced the error also runs the check. Worse, prompted to find problems, models often invent them — "fixing" correct answers into wrong ones.

What does work is changing who does the checking:

• A separate critic role. Self-Refine showed ~20% average gains when feedback comes from a genuinely separate feedback pass rather than the answer being second-guessed inline.
• Genuine disagreement. MIT researchers found that multiple model instances debating each other over several rounds measurably reduces hallucinations and improves mathematical reasoning — the models catch each other's errors precisely because they didn't make them.
• Trained critics beat human reviewers. OpenAI's CriticGPT work (LLM Critics Help Catch LLM Bugs) found model-written critiques of code were preferred over human critiques in 63% of cases on naturally occurring bugs.

The intellectual roots go back further: Irving, Christiano, and Amodei's 2018 AI safety via debate proposed that two AIs arguing opposite sides — with a judge picking the more truthful case — could surface things neither would volunteer alone. Adversarial review is the practical, everyday descendant of that idea.

How the loop actually works

1Chat's version reviews the plan, not the finished answer — because by the time a wrong answer is written, it reads exactly as fluently as a right one. The loop:

1. Plan. The model drafts how it will answer: what it knows, what it should look up, which connected apps it needs, what could go wrong.
2. Adversarial review. A reviewer — primed to be skeptical, not polite — attacks the plan: missing verification steps, stale assumptions, tools it should have used and didn't.
3. Revise. "Revisions requested" isn't a failure — it's the system working. The planner must address every point and resubmit. (This is the external feedback that the self-correction research says is required.)
4. Approve, then execute. Only the approved plan rides into the actual answer. Every round is visible in the chat's Deliberation trace, so you can read exactly what the reviewer objected to.

plan · round 1answer from memory, skip the search
review · revisions requested“prices change — verify against a live source”
plan · round 2search first, then answer with dates
review · approved
The answer you see was written from the approved plan — checked before it was born, not after.

Separately from planning, a verification-style pass over factual claims — the Chain-of-Verification pattern — catches errors that survive drafting, which is the same principle our deep research pipeline applies to sources: never let the first draft grade itself.

When it's worth the extra credits

Honest cost first: adversarial review adds a flat 4 credits to the turn on top of your model's normal rate, because the planning rounds are real model calls. That's the wrong trade for casual questions and quick iterations.

It's the right trade when:

• The answer feeds a decision. Anything you'd forward to a client, a boss, or a bank.
• The task is multi-step with tools. Plans that touch your connected apps benefit most — the reviewer regularly catches "you listed a tool but never planned to call it" class mistakes before they burn a turn.
• You'd otherwise ask twice anyway. Two manual attempts cost about the same and give you no visible critique; the review loop shows its reasoning.

Related reading: how routing styles pick your model in the Fusion Router announcement — Fusion is the sibling idea, three models answering independently instead of one model being cross-examined — and what AI memory actually is, since the reviewer also checks plans against what's known about you.

Turn it on per-chat from Controls in the harness bar, with a rigour dial for how hard the reviewer pushes. Watch the Deliberation trace on your next tricky question — the first time you see a plan get rejected for a reason you wouldn't have caught yourself, the 4 credits explain themselves.