Fide AI

FID-103 · Open question

When Does AI-Assisted Review Miss Shared Blind Spots?

When an AI system and its reviewers rely on the same mistaken premise, which review procedure helps an accountable reviewer catch the error before accepting a consequential answer or recommendation?

Why the question remains open

Several reviewers can agree because they use the same source, inherit the same assumption, or see the same persuasive framing. A second opinion can help when it brings independent evidence. It can also reinforce a mistake. An institution needs to know when to ask for an independent check and when to defer the decision because the necessary evidence is unavailable.

Working hypothesis

A proposition to test, not a finding.

Reviewers who make an initial assessment and can inspect independently selected source evidence will catch more shared errors than reviewers who receive a second judgment based on the same evidence. Structured disagreement may help on genuinely ambiguous cases. It may also cost review time or make correct answers harder to accept.

Proposed method

How the question could be tested

  • 01Start with a small adjudicated set of source or authorization decisions with known outcomes. Include correct recommendations, a shared but false premise, one source that contradicts the consensus, and cases where the record is insufficient. Use public or synthetic sources and two reviewers to qualify the reference labels.
  • 02Compare direct review, a second judgment with the same source access, structured disagreement or collaborative review, and a source check based on an independently assembled evidence set. Hold review time and configurations comparable. Give direct and structured review the same additional source in a matched comparison to separate the effect of evidence access from the protocol.
  • 03Measure acceptance of known errors, unnecessary rejection of correct answers, warranted deferral, citation to decisive evidence, and cost. Report reviewer agreement separately. Use automated and scripted checks to qualify the instrument before any separately reviewed participant study.

Needed controls

What must constrain the study

  • 01Vary shared evidence and shared prior assumptions separately. Record model versions, reviewer qualifications, information access, and what is genuinely independent about each check.
  • 02Balance plausible but false claims with plausible true claims and insufficient evidence. Score only claims the reference sources can adjudicate. Preserve legitimate theological or institutional disagreement as a distinct outcome.
  • 03Compare established debate or collaborative disagreement methods against a strong source-first baseline. Keep held-out cases outside protocol tuning.

Relationship to existing work

FID-011 measures reviewer reliability, and FID-079 tests the effect of disclosing assumptions. This call tests whether a review procedure detects a known error that multiple reviewers might share. FID-064 concerns collective judgment in communities, and FID-101 studies the incremental value of model internals.

Expected outputs

Artifacts the work should produce

  • 01A small source-backed casebook and a matched review protocol.
  • 02A comparison of false acceptance, deferral, and review cost under shared errors.
  • 03A guide to when a second reviewer needs different evidence or expertise.

Open questions

Uncertainties the protocol must resolve

  • 01Which failures can be caught by independent source access alone?
  • 02When does structured disagreement improve decisions enough to justify the cost?
  • 03How should the procedure handle disagreement that no available source can resolve?

Open question

Open work

Primary need: scalable oversight, measurement science, reviewer study design, and source adjudication

  • Draft one source-conflict case in 1–2 hours. Use the UK AISI oversight agenda and the collaborative disagreement paper. Return an adjudicable shared error, a matched correct case, and a case with missing evidence.
  • Review the authority labels or build a matched structured-disagreement baseline.

New to the Contributor Network? Share a short private introduction once, even if you join several calls. You can join the public discussion while introducing yourself.

Join the Contributor Network ↗