AI Bias

Chatbot Bias by Dialect or Name: The Quality Gap

Accent and Dialect Bias

ASR Performs Significantly Worse on Non-Standard Accents

Age Bias in Symptom Interpretation

Model Misinterprets Symptoms Differently Based on Patient Age; Younger Patients Undertreated, Older Patients Overtreated

Algorithmic Discrimination

AI System Systematically Discriminates Against Protected Groups

Anchoring Bias on First Diagnosis

Agent Fixates on the First Plausible Diagnosis Suggested Early in the Conversation and Discounts Later Contradictory Evidence

Best-Case Projection Bias

Sales forecasting model uses deal probability from salesperson input, which is inherently optimistic; projects deals that salesperson hopes will close rather than actual likelihood

Bias Amplification

Agent Amplifies and Reinforces User Biases

Confirmation Bias from Prior Clinical Notes

Agent Anchors on a Prior Visit's Diagnosis and Discounts New Evidence Contradicting It

AI Agent Consensus Illusion: Causes and Fixes

Agents appear to agree, but only because they were all seeded from the same flawed context or bias -- not because they independently verified the answer.

Demographic Bias in Diagnosis & Treatment Recommendations

Model Biased Toward/Against Certain Demographics (Race, Gender, Age); Different Accuracy for Different Groups

Look-Ahead Bias in Backtests

Model Uses Future Information in Backtest That Wouldn't Be Available in Real-Time Trading

Missing Data Mishandling in Financial Models

Model Handles Missing Data Incorrectly (Mean Imputation, Deletion); Introduces Bias or Information Loss

Model Context Length Behavior Change

As an agent's conversation or retrieved context grows toward the model's context window limit, the model's behavior shifts in ways that are never announced: recall of early-turn facts degrades, instruction-following becomes less reliable, and the model increasingly favors recently-seen tokens over earlier ones ("recency bias"). Because the API returns a normal, well-formed response at every context length, the agent has no signal that it just crossed into a degraded-quality regime.

Model Fairness Bias

The underlying model exhibits systematic differences in its outputs correlated with demographic or protected attributes — name-implied ethnicity, gender-coded pronouns, geography, or dialect — that leak into agent decisions the model wasn't explicitly asked to make on that basis. Because the bias is statistical rather than an explicit rule, it survives even when the agent's prompt contains no discriminatory instruction, and it recurs consistently enough to produce a measurable disparate pattern across many decisions.

Positional Bias Omits Mid-Document Admission in Deposition Summary

Agent Summarizing a Long Deposition Transcript or Produced Document Systematically Under-Weights Content in the Middle of the Document, Omitting a Material Admission That Was Fully Within the Model's Context Window and Correctly Classified as Relevant

Recency Bias in Portfolio Recommendations

Model Overweights Recent Performance; Recommends Assets That Performed Well Recently but Are Mean-Reverting

Resume Keyword Matching Bias

AI resume screener uses exact keyword matching; rejects qualified candidates with industry synonyms or relevant but non-exact terminology (e.g., "web development" vs "frontend engineering")

Resume Keyword Overfit Bias

Candidate-Screening Agent Over-Weights Surface Keyword Matches Against the Job Description, Systematically Filtering Out Qualified Candidates Who Describe Equivalent Experience Differently

Salience Bias

Model Overemphasizes Visually Salient Features, Ignoring Context

Self-Verification Bias

Same model judges its own flawed output.

Spatial Attention Bias

Model Fixates on Image Center or Edges; Misses Objects in Peripheral Regions

Survivorship Bias in Training Data

Model Trained Only on Survived Assets; Ignores Assets That Failed or Were Delisted

Symptom Attribution Bias (Wrong Diagnosis Due to Anchoring)

Model Anchors on Initial Symptom; Misses True Diagnosis Because Anchored to Wrong Hypothesis

Training Data Leakage & Memorization

Model Hallucinates Training Data Artifacts, Overfitting to Dataset Biases

User Feedback Bias

The mechanism used to collect quality signal — thumbs-up/down buttons, post-chat surveys, star ratings — is only used by a non-representative subset of users, typically those with strongly positive or strongly negative experiences, while the much larger group with a mediocre-but-tolerable experience stays silent. Teams then treat the collected feedback as representative of overall quality, when it's actually a bimodal sample that systematically misses the median experience, leading to miscalibrated confidence in how the agent is actually performing.