Cognitive Bias

Speed vs accuracy – fast, intuitive processing (often “System 1”) trades rigour for speed; slow, analytical processing (“System 2”) corrects when engaged.

Common families (with examples)

• Confirmation bias – we search for and overweight evidence that supports our prior.

  Countermoves: pre-register disconfirming tests; assign a “red team”; force a “what would change our mind?” line.

• Loss aversion – losses loom larger than equivalent gains (often ~2×).

  Countermoves: show both frames (gain and loss); use long-horizon metrics; make reversible trials the default.

• Gambler’s fallacy – believing independent events self-correct (“after five tails, heads is due”).

  Countermoves: state independence/base rates explicitly; display binomial ranges; ban “due for” language in reviews.

• Availability cascade – repeated, vivid claims feel truer and spread via social proof, regardless of evidence.

  Countermoves: require source quality and independent confirmation; add a cool-off for viral topics; track what’s known vs said.

• Framing effect – choices swing when the same facts are presented differently (e.g., 90% survival vs 10% mortality).

  Countermoves: standardise wording; show both frames and absolute numbers with denominators; pre-set decision rules.

• Bandwagon effects (herding) – people adopt beliefs or actions because others have, not because of private evidence.

  Countermoves: collect independent estimates before discussion; use silent votes/Delphi; expose dissent and its evidence.

• Dunning–Kruger effect – novices overestimate ability; experts may under-estimate variance and assume shared context.

  Countermoves: calibration training (forecasting with feedback), checklists and peer review; compare to reference-class outcomes.

Context matters – stress, ambiguity, and incentives amplify bias; check the environment, not just the person.

Decision reviews for strategy, hiring, investment, pricing, and vendor selection.

Research & experiments – survey design, A/B test interpretation, sampling and power.

Risk & forecasting – reference-class baselines, ranges not points, premortems.

Communication – neutral wording, show both frames, order-effects control.

Ops & safety – checklists, standard work, second-checker on high-risk steps.

How to avoid

1. Define the decision & base rate – write the objective, options, criteria and a reference-class outcome.

2. Get independent estimates first – collect forecasts or scores before any discussion (prevents anchoring/herding).

3. Premortem – “It failed in 12 months — why?” Turn the top causes into mitigations or tests.

4. Consider the opposite – assign a red team and a “what would change our mind?” line.

5. Frame both ways – show gain/loss frames and absolute numbers with denominators; keep wording standard.

6. Check sampling – size, selection, survivorship, time window; prefer cohorts and hold-outs.

7. Pre-commit rules – thresholds, stop/scale criteria, decision windows; avoid ad-hoc pivots.

8. Run reversible tests – small trials beat argument; use expected value, not stories.

9. Calibrate – record probability ranges, track forecast accuracy, and retrain on misses.

10. Align incentives – metrics and pay should reward truth-seeking, not outcome theatre.

Bias-specific guardrails

• Confirmation bias: pre-register disconfirming tests; adversarial collaboration; ban HARKing.
• Loss aversion: use EV and break-even math; emphasise reversibility; evaluate over a longer horizon.
• Gambler’s fallacy: state independence/base rates; show binomial ranges or control charts.
• Availability cascade: require source quality and independent corroboration; add cool-off before big decisions.
• Framing effect: display both frames side-by-side; randomise presentation order.
• Bandwagon effects: silent votes/Delphi; capture minority reports and reasons.
• Dunning–Kruger: use checklists and supervision for novices; peer review; calibration training with feedback.

Checklist theatre – rituals without teeth; tie each step to a real go/no-go or redesign.

Bias labelling as argument – calling someone “biased” doesn’t resolve evidence; show the data and process defect.

Over-correction – slowing everything to a crawl; reserve heavy controls for material decisions.

One-and-done – biases reappear; keep a cadence of calibration and post-mortems.