THE LIVING ATLAS

Open question · Working / Judgement

How do we judge AI-assisted work when we cannot check every step?

Reviewers have limited time and expertise. Which parts do they inspect, which signals do they trust, and what makes them challenge an apparently finished answer?

Open question · no dedicated Atlas answer yet

Question history ↓

Keep this question in My Atlas to see future contributions and map changes. No email notifications.

What this question includes

Trust concerns the support for an answer. This question concerns how people evaluate and use that support in practice, including when to ask someone else to review it.

A possible investigation · not scheduled or tested

What do reviewers miss when an AI answer looks finished?

  1. 01Set the criteria
  2. 02Choose what to inspect
  3. 03Accept or challenge
  4. 04Learn from missed errors

What would count as progress?

  • Compare recognisable error types with a known answer where possible.
  • Record what reviewers actually detected, alongside time and unresolved uncertainty.
  • Distinguish confidence in an answer from successful detection of a consequential error.

These articles explain parts of the problem. They do not amount to a dedicated answer to this question.

Explainer · 23 October 2026

Better ideas, less variety? →

Examine how the task and evaluation instructions shape the resulting ideas. More variation is not automatically a better decision.

Four studies · 2024–2026. Includes two preprints; bounded writing and product tasks.

Explainer · 23 October 2026

Can AI discover something new? →

See how the choice of an independent check changes what can be claimed as a discovery.

Concrete research cases, with different kinds of evidence. No claim that AI can autonomously solve science as a whole.

How this question changed

Question versions describe its framing, published contributions and place in the map. Article versions keep their own evidence and history.

Map change · question version 1 · recorded 24 October 2026

Atlas now gives every current article a question home and opens seven further questions with a scope and related reading. Open means no dedicated Atlas answer yet. No new research finding or test is claimed, and existing article-based follows keep their scope.

See this change across Atlas →
All questions →What changed across Atlas →