Historical bias arises when the world itself is unfair, and data faithfully reflects that unfairness. If women were historically underrepresented in tech leadership, a dataset of past leaders will be mostly men — and an AI trained on it will learn that “leader = male.” Societal bias is embedded in language, culture, and institutions. Word embeddings trained on internet text learn that “doctor” is closer to “man” and “nurse” is closer to “woman.” These biases exist before any algorithm is built. The AI doesn’t create them — it inherits and amplifies them. This is the hardest type of bias to address because the “ground truth” data is itself biased.

Representation bias: the training data doesn’t represent the population the model will serve. ImageNet was predominantly Western images; dermatology datasets were predominantly light skin. Selection bias: the data collection process systematically excludes certain groups. Surveys that only reach internet users miss offline populations. Measurement bias: the way data is collected introduces systematic errors. Arrest data reflects policing patterns, not actual crime rates — neighborhoods with more police have more arrests. Label bias: human annotators bring their own biases to labeling. Studies show annotators rate identical text differently depending on the perceived race of the author. Temporal bias: data from one time period may not apply to another. COVID-19 changed consumer behavior, invalidating pre-pandemic models.

Even with perfectly balanced data, algorithms can introduce bias: Objective function bias — optimizing for overall accuracy means the model performs best on the majority group and worst on minorities. A 95% accurate model might be 99% accurate for Group A (90% of data) and 60% accurate for Group B (10% of data). Feature selection bias — using zip code as a feature is a proxy for race in many US cities. The model doesn’t need explicit race data to discriminate. Aggregation bias — a single model for all populations may fail for subgroups with different patterns. HbA1c thresholds for diabetes differ across ethnicities, but a single threshold disadvantages some groups. Evaluation bias — testing on a non-representative benchmark gives a false sense of performance.

AI systems don’t just reflect bias — they amplify it through feedback loops. Predictive policing: the model predicts high crime in neighborhoods with more historical arrests → police are sent there → more arrests occur → the model sees more “crime” → predicts even more crime. The prediction becomes self-fulfilling. Recommendation systems: users click on content the algorithm shows them → the algorithm learns those preferences → shows more of the same → creates filter bubbles and echo chambers. Credit scoring: people in underserved areas get denied credit → they can’t build credit history → future models see no credit history → deny credit again. Breaking feedback loops requires: monitoring outcomes over time, injecting randomness (exploration), and regularly auditing for disparate impact.

In 2018, Joy Buolamwini and Timnit Gebru published the Gender Shades study, testing commercial facial recognition systems from Microsoft, IBM, and Face++ on a balanced dataset of faces across skin tones and genders. Results: light-skinned males: 0.8% error rate. Dark-skinned females: 34.7% error rate. That’s a 43x difference in error rate. The cause: training datasets were overwhelmingly light-skinned and male. The systems worked well for the population they were trained on and failed catastrophically for everyone else. After the study, Microsoft and IBM improved their systems significantly. The study demonstrated that auditing AI for bias works — companies fix problems when they’re exposed.

In the early 2010s, Amazon built an AI system to automate resume screening. The system was trained on resumes submitted over a 10-year period — during which most successful hires were men (reflecting the tech industry’s gender imbalance). The AI learned to penalize resumes containing the word “women’s” (as in “women’s chess club captain”) and preferred candidates from all-male colleges. Amazon tried to fix the bias but couldn’t guarantee neutral recommendations. They scrapped the tool in 2018. The broader problem: a 2024 University of Washington study found that LLMs (GPT-4, Claude, Gemini) preferred white-associated names 85% of the time when evaluating identical resumes. Today, 83% of companies use AI for resume screening, and 50% use AI for initial rejections — many candidates never see a human reviewer.

LLMs are trained on internet text, which reflects every bias in human society. Stereotyping: LLMs associate professions with genders (“nurse” → female, “engineer” → male), races with traits, and religions with sentiments. Representation: English-centric training data means LLMs perform worse in non-English languages and underrepresent non-Western perspectives. Toxicity: internet text contains hate speech, which LLMs can reproduce. Sycophancy: RLHF training can make models agree with users rather than provide accurate information, reinforcing existing beliefs. Cultural bias: models reflect predominantly Western, English-speaking, educated perspectives. The challenge is unique: LLMs are general-purpose, so bias can manifest in unpredictable ways across millions of use cases.

Disaggregated evaluation: break down model performance by demographic group. Report accuracy, precision, recall, and false positive/negative rates for each group separately. Disparate impact testing: check if the model’s decisions disproportionately affect one group. The “four-fifths rule” (US EEOC): if the selection rate for a protected group is less than 80% of the rate for the most-selected group, there’s evidence of adverse impact. Counterfactual testing: change the protected attribute (e.g., swap names, genders) and check if the prediction changes. Bias benchmarks: standardized datasets designed to test for specific biases (WinoBias for gender, BBQ for social biases). Red teaming: adversarial testing by diverse teams to find bias in edge cases.