Confirmation Bias
Review paths narrow toward what already looks true — disconfirming evidence never reaches bind
Bias mechanism
Once a tentative conclusion exists, search and attention favour supporting evidence; challenges are not sought or are dismissed before tier is examined.
Cognitive biasShort-termismLocal optimisation
Confirmation bias skews what reaches the commitment boundary. The underwriter opens the file already leaning approve; the clinician sees the model’s green banner; the engineer knows the release train leaves tonight. The review becomes curation of confirming facts — not an honest pass over what could falsify the tentative conclusion.
Term entry: confirmation bias.
What it produces
- Confident deck at scale — especially LLM output that reads complete and authoritative while omitting counter-evidence.
- Rubber-stamp human gates — humans confirm what the model already said because disconfirming material was never surfaced in the UI.
- Undisclosed “stress tests” — challenges generated but not shared; compliance theatre that consumes budget without changing judgment.
Counter direction
- AI stress testing with full disclosure — counter-arguments and human responses recorded in the L0 payload.
- Blind and independent review before bind — blind peer review.
- Citation and retrieval requirements — force external anchors, not only parametric memory (LLMs at the Commitment Boundary).