What Is Cognitive Bias?

Seeking and weighting evidence that supports what you already think. Where it bites: the vendor shortlist assembled after someone has privately decided, where every reference call confirms rather than tests.
Evidence: Wason's 1960 rule-discovery task is the origin, though his own conclusion applied to the subgroup that persisted after two or more errors. Klayman and Ha (1987) reframed it as a positive test strategy that is normatively sound in most environments — you are not irrational for looking where you expect to find things. Hart and colleagues' 2009 meta-analysis in Psychological Bulletin puts congeniality bias at d = 0.36, and it reverses when disconfirming information is genuinely useful.
In Argumentree: the con branch is not optional. A claim without a recorded counter-case is visibly incomplete rather than quietly unchallenged, and a question chain attaches the challenge to the exact claim it disputes.

The first number mentioned sets the range everyone argues within. Where it bites: whoever says a budget figure, a valuation or a delivery date first has moved every subsequent estimate.
Evidence: Tversky and Kahneman's 1974 demonstration had median estimates of 25 versus 45 for the same quantity after anchors of 10 and 65. Many Labs 1 replicated anchoring robustly — but using provided anchors. Incidental anchoring from random numbers failed in Many Labs 2 (d = 0.04, p = .09, N = 6,826). So: deliberate anchors move people; arbitrary numbers in the environment largely do not.
In Argumentree: positions are submitted independently before discussion, so there is no first number to anchor on. The anchor needs an audience, and asynchronous input removes it.

Judging how likely something is by how easily an example comes to mind. Where it bites: last quarter's outage dominates the risk register while the slow, larger risk nobody has a story about goes unlisted.
Evidence: split, and worth knowing which half. The famous letter-frequency demonstration was undercut — Sedlmeier, Hertwig and Gigerenzer (1998) found judgements generally followed actual proportions. The ease-of-retrieval mechanism survives: Weingarten and Hutchinson's 2018 meta-analysis across 263 studies finds a medium effect, though publication bias shrinks it by up to a third.
In Argumentree: arguments carry their evidence, so a vivid anecdote and a base rate sit side by side and get rated separately — and the timeline shows when each entered the discussion.

Moore and Healy's 2008 paper in Psychological Review splits this into three: overestimating your score, overplacing yourself against others, and overprecision — being too certain your estimate is right. Where it bites: the delivery range that is far too narrow, so every plan downstream inherits false precision.
Evidence: overprecision is the robust one. Soll and Klayman found confidence intervals "sometimes only 40% as large as necessary to be well calibrated." Overestimation and overplacement flip with task difficulty — hard tasks produce overestimation but underplacement — and the hard-easy effect itself is partly a regression artifact (Juslin et al. 2000).
In Argumentree: ratings are on the argument, not the assertion, so a confident claim with thin evidence scores like thin evidence.

Underestimating your own completion times while judging others' realistically. Where it bites: every migration, integration and launch date you have ever set.
Evidence: strong in the field. Flyvbjerg and colleagues (2002) examined 258 transport projects worth about US$90bn and found costs underestimated in roughly nine of ten — rail +45%, fixed links +34%, roads +20%. Buehler, Griffin and Ross (1994) is the lab origin, and notably the bias was eliminated when participants were forced to connect the estimate to past experience. Caveat worth carrying: Halkjelsvik and Jørgensen's 2012 review found underestimation dominant in engineering and management literature but not in the psychology literature.
In Argumentree: the estimate becomes a claim with evidence attached — including the reference class — rather than a number asserted in a planning meeting.

Preferring the current arrangement, and treating "do nothing" as neutral rather than as a choice with consequences. Where it bites: the incumbent vendor renewed because switching requires justification and staying does not.
Evidence: among the strongest here. Samuelson and Zeckhauser (1988) ran 486 subjects and found that only 28% of TIAA-CREF participants had ever changed their allocation. Jachimowicz and colleagues' 2019 meta-analysis of default effects across 58 datasets (n = 73,675) reports d = 0.68 [0.53, 0.83], and Madrian and Shea found 401(k) participation moving from 37% to 86% on a default change alone.
In Argumentree: "keep what we have" is entered as an option with its own pros and cons, so it competes on evidence instead of winning by not being on the agenda.

The same facts described as a gain or a loss produce different choices. Where it bites: "90% success rate" and "10% failure rate" are the same number and do not land the same way.
Evidence: robust in direction, roughly half the original magnitude. Many Labs 1 replicated Tversky and Kahneman's Asian disease item: original d = 1.13 → replication d = 0.62 (N = 6,271). Many Labs 2 replicated a different item from the same 1981 paper — original OR 4.96 → OR 2.06 (N = 7,228). Those two are constantly conflated; they are different studies. Kühberger's meta-analysis of 136 papers calls framing reliable and of small-to-moderate size.
In Argumentree: decision criteria are written before the options are framed, and a claim can be stated once and challenged in both frames rather than surviving on presentation.

Once you know the outcome, you believe you knew it all along. Where it bites: the post-mortem that becomes a search for who should have seen it, which teaches everyone to be quieter next time.
Evidence: robust, magnitude disputed. Fischhoff (1975) showed outcome knowledge raises judged likelihood and that people are largely unaware of the shift. Meta-analyses disagree on size: Christensen-Szalanski and Willham (122 studies) report r = .17; Guilbault and colleagues (95 studies) report a median around .39 — and found debiasing manipulations largely ineffective.
In Argumentree: the decision record is timestamped, so a review can judge the decision on the evidence that existed at the time instead of on the outcome that arrived later.

Explaining behaviour by character while discounting the situation. Where it bites: the project failed because the lead was weak — rather than because the process gave them a deadline set by anchoring and a plan approved by consensus.
Evidence: solid in paradigm. Jones and Harris (1967) found attitude attributions persisted even when the essay position was assigned rather than chosen. Many Labs 2 replicated a correspondence-bias study successfully — original d = 1.75, replication d = 1.82, one of the sturdier results in the set. Cross-culturally it attenuates rather than disappears.
In Argumentree: arguments are rated, people are not. That is the whole design principle, and it is the direct structural answer to this bias.

Counting unrecoverable past spending in a present decision. Where it bites: the project defended because of what it has already consumed.
Evidence: real but moderate, and the measures are contested. Arkes and Blumer (1985) found 85% chose to finish a plane that was already known to be inferior. Sleesman and colleagues' meta-analysis puts sunk cost at ρ = .243 — comparable to, not larger than, personal responsibility. A 2025 analysis found classic sunk-cost vignettes have poor internal consistency (ω = 0.14–0.57), so treat single-vignette findings with care. Note also that in Arkes and Blumer's own theatre study the effect was present in the first half of the season and essentially gone in the second.
In Argumentree: exit criteria can be recorded when the project is approved, so the review reads a document rather than relitigating a colleague's judgement — see escalation of commitment.

Judging by how well a case matches a stereotype while ignoring how common the thing actually is. Where it bites: "this deal feels like the ones we win" — without the win rate.
Evidence: robust as a demonstration, and strongly format-dependent, which is the actionable part. Casscells and colleagues (1978) found only 18% of medical staff gave the correct answer while 45% were off by a factor of nearly fifty; Manrai and colleagues repeated it 36 years later with almost identical results (23% correct, 44% giving the same wrong answer). But Gigerenzer and Hoffrage showed correct responses rising from about 16% to 46–50% when the same problem is posed in natural frequencies rather than probabilities.
In Argumentree: evidence is attached to the claim it supports, so a base rate is present as an argument rather than remembered — or not — by whoever happens to know it.

The claim that the least competent are the most confident. Where it bites: the loudest certainty in the room belonging to the person with the least basis for it.
Evidence — and this one needs care. The descriptive pattern replicates: low performers do overestimate themselves. The explanation largely does not. Nuhfer and colleagues reproduced the classic graph shape from random numbers; Gignac and Zajenkowski (2020, N = 929) found the relationship essentially linear and the effect much smaller than reported; McIntosh and colleagues found that when task difficulty was equalised, the pattern was eliminated — performance level, not metacognitive deficit, was the driver. In fairness, Pennycook and Jansen and colleagues do find low performers less able to evaluate their own answers.
Honest summary: people overestimate themselves, but most of the famous graph is regression to the mean plus better-than-average. Do not cite it as a law.
In Argumentree: confidence is not a scoring input. A rating attaches to an argument's evidence, so certainty carries no weight it has not earned.