The trial of AI hides the trial of the human — a comparative inventory of cognitive biases
This article extends The bid manager's worst enemy is themselves and What the evaluator will never tell you. Here, we leave the terrain of tenders for a moment to set the two decision-making machines face to face — the human one and the artificial one.
A one-sided prosecution
AI is in the dock. Hallucinations, sycophancy, hollow prose, representation bias — the trade press, the conferences, the white papers document its flaws with growing precision. The case is serious. The evidence is there.
The court has a blind spot.
No one sets the human brain against it. Yet that brain has been documented for seventy years by cognitive psychology as a machine for shortcuts and systematic errors.
Daniel Kahneman and Amos Tversky inaugurated this cartography in the 1970s (Judgment under Uncertainty: Heuristics and Biases, Science, 1974). Kahneman catalogues some forty of them in Thinking, Fast and Slow (2011). Buster Benson's taxonomy (Cognitive Bias Codex, 2016) lists more than 180.
On the AI side, the major families of bias documented in the technical literature can be counted on the fingers of two hands.
Prosecuting AI without opening the case against the human produces an asymmetric verdict. AI's biases are visible because they are new. Human biases are invisible because we have grown used to them. Both sides deserve the same scrutiny.
Human biases — twenty documented mechanisms
Here are the best-documented biases that operate in any decision involving reading, judgment, and choice.
| # | Bias | Source | Mechanism |
|---|---|---|---|
| 1 | Recency bias | Tversky & Kahneman, 1974, Science | The brain gives disproportionate weight to the most recent information. |
| 2 | Anchoring bias | Tversky & Kahneman, 1974, Science | The first piece of information received sets a reference point that biases subsequent estimates. |
| 3 | Availability heuristic | Tversky & Kahneman, 1973, Cognitive Psychology | We overestimate the probability of what is easy to recall. |
| 4 | Confirmation bias | Wason, 1960, QJEP; synthesis Nickerson, 1998, Review of General Psychology | We selectively seek information that confirms an already-formed hypothesis. |
| 5 | Halo effect | Thorndike, 1920, Journal of Applied Psychology | A positive quality perceived in one domain spreads to the others. |
| 6 | Horns effect | The inverse symmetric of the halo, Thorndike, 1920 | A perceived weakness contaminates overall judgment in the negative direction. |
| 7 | Curse of knowledge | Camerer, Loewenstein & Weber, 1989, Journal of Political Economy | The expert can no longer represent to themselves the state of ignorance of the non-expert. |
| 8 | IKEA effect | Norton, Mochon & Ariely, 2012, Journal of Consumer Psychology | We overestimate the value of what we have produced ourselves. |
| 9 | Planning fallacy | Kahneman & Lovallo, 1993, Management Science; Buehler, Griffin & Ross, 1994 | We systematically overestimate our ability to meet deadlines. |
| 10 | Dunning-Kruger effect | Kruger & Dunning, 1999, JPSP | The least competent overestimate their abilities, the most competent underestimate theirs. (Magnitude contested by Gignac & Zajenkowski, 2020 — a statistical artifact of regression to the mean.) |
| 11 | Groupthink | Janis, 1972, Victims of Groupthink, Houghton Mifflin | A cohesive group converges toward consensus by suppressing dissenting opinions. |
| 12 | Status quo bias | Samuelson & Zeckhauser, 1988, Journal of Risk and Uncertainty | We prefer the default option, even when alternatives are objectively superior. |
| 13 | Illusion of control | Langer, 1975, JPSP | We overestimate our ability to influence largely random events. |
| 14 | Fundamental attribution error | Jones & Nisbett, 1971; Ross, 1977 | We attribute our own failures to external causes and those of others to internal causes. |
| 15 | Conformity bias | Asch, 1951, 1956 | We align our judgment with that of the group or the authority present. |
| 16 | Mere-exposure effect | Zajonc, 1968, JPSP | Familiarity with a stimulus increases preference, independently of quality. |
| 17 | Loss aversion | Kahneman & Tversky, 1979, Prospect Theory, Econometrica | Potential losses weigh more heavily than equivalent gains. |
| 18 | Representativeness heuristic | Tversky & Kahneman, 1974, Science | We judge an item by its resemblance to a mental prototype. |
| 19 | Survivorship bias | Operational concept, Abraham Wald (bomber analysis, Second World War); no single canonical founding author | We reason about visible cases while ignoring invisible ones. |
| 20 | Cognitive load and decision fatigue | Sweller, 1988, Cognitive Science | Discrimination degrades when the informational load exceeds processing capacity. (The "ego depletion" hypothesis of Baumeister et al., 1998, contested by Hagger et al., 2016 — Sweller's cognitive load remains the consensus framework.) |
| 21 | Salience bias | Taylor & Fiske, 1978 | Perceptually salient elements capture attention beyond their real importance. |
Twenty-one documented biases. Each confirmed by decades of research. None requires ill will to fire — the cognitive wiring does its work at every decision.
AI biases — nine documented mechanisms
A balanced verdict demands a serious inventory of AI biases. To minimize them would be as dishonest as ignoring the human ones.
| # | Bias | Source | Mechanism |
|---|---|---|---|
| 1 | Representation bias (data bias) | Barocas & Selbst, 2016, California Law Review | Patterns overrepresented in the training corpus are overreproduced in the output. |
| 2 | Statistical-average bias | Structural property of language models | The model converges toward the most probable answer, which is also the most banal. |
| 3 | Sycophancy (algorithmic flattery) | Sharma et al., 2023, Towards Understanding Sycophancy in Language Models, Anthropic | The model validates the user rather than contradicting them, including when the user is wrong. An artifact of RLHF, which rewards responses judged agreeable by human raters. Manifestations: endorsing a mistaken hypothesis if the prompt insists on it, reversing under pressure ("indeed, you are right, I was mistaken"), flattery added at the opening and closing. |
| 4 | Hallucinations (confabulation) | Brown et al., 2020, NeurIPS (GPT-3); Ji et al., 2023, ACM Computing Surveys | The model generates false statements with the same fluency as true ones. |
| 5 | Temporal bias (knowledge cutoff) | Architectural property of models with fixed training | The model reasons about a state of the world frozen at its cutoff date. |
| 6 | Attentional primacy-recency (Lost in the Middle) | Liu et al., 2023, Lost in the Middle: How Language Models Use Long Contexts | Information placed in the middle of a long context is exploited less well than information at the beginning or end. |
| 7 | Positivity bias (RLHF-induced) | Ouyang et al., 2022, Training language models to follow instructions with human feedback, NeurIPS | Training by human feedback favors enthusiastic and diplomatic responses. |
| 8 | Narrative-coherence bias | Emergent property of generative models; linked to confabulation (Ji et al., 2023) | The model builds a coherent narrative even when the underlying facts do not support it. |
| 9 | Goodhart's law applied to RLHF | Goodhart, 1975; Gao et al., 2023, Scaling Reward Model Overoptimization | Optimizing a proxy metric degrades the real measure it was meant to capture. |
Nine documented AI biases, each real and measurable. The trial of AI on these points is legitimate.
The asymmetry no one measures
Twenty-one documented human biases. Nine documented AI biases. The ratio is eloquent.
The cognitive literature is broad, old, replicated. The technical literature on AI biases is recent — most of the catalogued biases were identified in the last three years — but also more inspectable: a model can be taken apart, a brain does not surrender its source code.
This asymmetry of visibility reflects no asymmetry of gravity. Human biases have become invisible through familiarity. AI biases are visible because they are new. The one-sided prosecution of AI charges the newcomer with the very flaws we have always tolerated in the long-standing resident.
Application to pre-sales
All these biases operate in any decision process. But they take on a particular relief in responding to tenders — terrain where every file mobilizes reading, interpretation, judgment, internal negotiation, and final decision, on both sides of the table.
On the bidder side, recency bias pushes toward recycling the last file handled. Anchoring fixes an often arbitrary angle of attack from the first reading of the CCTP. Availability leads to choosing the references one knows, not the most relevant ones. The curse of knowledge makes invisible to the writer what the evaluator will read cold. The IKEA effect inflates the self-assessment of the produced technical proposal. Groupthink converges in bronze/silver/gold review without contestation. Status quo bias recycles the usual templates. The planning fallacy promises a delivery that will slip by three days.
On the evaluator side, anchoring on the first file read calibrates the scoring scale. The halo and its inverse color the entire reading after a few pages. Mere-exposure favors the incumbent. Loss aversion privileges the "safe" provider over the "better but unknown" one. Decision fatigue degrades the rigor of the fifth file of the day. Conformity bias aligns the co-evaluators with the rapporteur in committee.
Twenty-five distinct manifestations, thirteen on the bid manager side and twelve on the evaluator side, operating at every tender — silently, without ill will, by construction of the cognitive wiring. To be compared with the nine biases of a well-architected AI agent.
AI as a counterweight
The reversal of perspective opens a possibility the one-sided prosecution conceals: each human bias can be countered by an architectural use of AI.
Against recency bias. An agent that reads each CCTP with no conversational history between files treats each tender as an isolated instance. The bid manager's last file does not contaminate the analysis of the next.
Against the curse of knowledge. The agent plays the role of the cold reader. It reads the technical proposal the way the evaluator would — without knowing the three weeks of context accumulated by the writer. If a section is opaque to the agent, it will be opaque to the evaluator.
Against groupthink in committee. An agent produces an evaluation independent of the rapporteur — an explicit counterweight that makes disagreements visible before convergence.
Against the evaluator's status quo bias. A technical proposal architected to activate the surprise → curiosity → exploration sequence in the reader is the bidder's antidote. When AI is guided by a cognitive model of the reader, it builds this mechanism deliberately.
Each of these uses presupposes a condition: AI must function as a cognitive system that makes its hypotheses explicit and traces its inferences. A mirror chatbot amplifies the user's biases instead of correcting them.
What TenderGraph builds
TITAN functions as a cognitive counterweight. Each hypothesis is named, each inference is traced, each decision is auditable. The agent tests the bid manager's angle, including when the writer would like a confirmation. Each CCTP is read as a new instance. Flaws are flagged — the writer's comfort does not enter the equation.
Eliminating human biases is impossible — they are wired in. Making them visible at the moment they operate: that is the only real margin for action.
Discover how TenderGraph works · Contact us
Key takeaway: Correcting AI's biases is an editor's work — proofreading, verification, generation constraints. Countering human biases with AI demands an architect's work — cognitive model, explicit process, traceability of inferences. The quality of decisions depends on both. The second remains massively under-invested.
Further reading:
- The bid manager's worst enemy is themselves — The founding diagnosis of cognitive biases on the writer's side.
- What the evaluator will never tell you — From the other side of the table: how the evaluator reads, scores, and compares.
- The executive summary myth — The executive summary, terrain where anchoring does the most damage.
- Why your client references convince no one — Availability bias at work in the choice of references.
- What the specifications don't say — Silent inference: the brain fills the gaps without flagging it.
- The information revolution in pre-sales — Cognitive biases as sources of noise in the CCTP → technical proposal → score channel.
- Why you can't integrate AI into your tenders — The cognitive block: the organization pictures AI as a chatbot.
- The acceleration of pre-sales cycles — The time freed by AI finally allows structured reviews.
- The defense: the moment when everything is decided — In the defense, biases operate in real time.
- Client knowledge, the double-locked key — The surprise → curiosity → exploration mechanism that breaks the evaluator's status quo.
- Epistemic marking: the human signature LLMs do not reproduce — A direct extension on the overconfidence of LLMs: why RLHF structurally produces overconfident sentences, and how human calibration remains the contractual signature.
- Framing: the sentence that decides before the arguments begin — Above the biases the frame is written: reframing a file neutralizes biases in chains and redistributes the model's completion toward the useful zone.
- Free AI does not exist: the economics of inference and the window of opportunity — Controlling biases in an agentic chain has a real cost: how much to spend, on which models, with what change management on the team side.
- AI sovereignty: why DeepSeek V4 changes the strategic calculus — For biases touching sensitive data, the technical sovereignty of the deployment becomes the condition for a credible cognitive policy.