A tool can seem neutral and still make different recommendations for people with similar needs or qualifications. That concern is called AI bias. It can affect individuals, users, businesses, and society when technology helps shape important decisions.
For example, a hiring, education, or healthcare tool may treat people differently despite similar qualifications or needs. Training data means the examples a model uses to learn patterns. If those examples leave out some people, the system may not serve them well.
Fairer technology takes ongoing scrutiny, not a one-time fix. You deserve to know when an AI decision plays a role and whether people check its effects. Treat AI outputs as decisions worth checking, not as automatic proof of what is fair.
Key Takeaways
- AI bias is a recurring unfair pattern, not one mistake.
- Data, design choices, and real-world use can all shape outcomes.
- Group-level checks can reveal differences hidden by overall averages.
- People should review important AI decisions and ask for explanations.
- Fairness checks need to continue after a system launches.
What Is AI Bias?
AI bias is a repeated, unfair pattern in an AI system’s results, not simply one incorrect answer. Repeated matters: one error may be random, while a pattern can favor or harm groups over time.
Algorithmic bias in AI means an algorithm or system produces unfair results through data, design, or use. It can affect decisions even without a person intending harm.
A system can produce biased outcomes even when nobody deliberately programmed it to discriminate.
Input bias arises in data, system bias in design and development, and application bias in real-world use. These pathways can overlap, so checking code alone cannot show the full risk.
The NIST AI Risk Management Framework is a voluntary guide for managing risks to individuals, organizations, and society. NIST released version 1.0 on January 26, 2023.
NIST’s AI Risk Management Framework 1.0 helps organizations consider how AI decisions may affect people and society. Its focus is risk management, not a promise that every system will be fair.
Where Does Bias Enter an AI System?
Bias can enter before a model makes its first prediction, starting with the examples selected to teach it. Training data shapes the patterns a model can learn, but it cannot fill gaps on its own.
Bias can arise in training data, system design, or real-world use. The names for these stages are input bias, system bias, and application bias.
How data gaps and labels enter
Data collection may exclude groups, settings, or cases that matter later. Labels also guide the model: a mistaken tag can teach the wrong pattern.
Confirmation bias occurs when choices reinforce patterns already present in the data. A model trained on old decisions may repeat them instead of challenging them.
Common data-related biases include:
- Measurement bias: a speech tool records some accents less accurately.
- Exclusion bias: a survey leaves out people without internet access.
- Selection bias: a sample includes only people who finished a course.
- Recall bias: people remember past events differently when surveyed.
- Stereotyping bias: labels link a role with one gender without evidence.
How design and deployment compound the gap
Design teams choose which patterns a model should prioritize and how to measure success. A model may score well on a broad test while missing errors for a smaller group.
Validation checks performance before release, but its results depend on the cases tested. If the test set repeats gaps in training data, the review can miss the same weaknesses.
Application bias appears when people use a system in ways its makers did not test. A model built for one setting may not fit another group or task.
Biases can build across each stage, from collection through everyday use. Check who or what is missing from the data before assuming that a model has learned a complete picture.
Which People and Decisions Can AI Bias Affect?
The consequences of AI bias depend on where a system is used, from a classroom flag to a healthcare recommendation. Fairness means comparing outcomes across relevant groups, not assuming equal inputs guarantee equal results.
Access, error rates, and recommendations can differ across groups. One quarter of U.S. school-aged children live in households without broadband or a web-enabled device, which can limit access to education technology.
Writing by non-native English speakers has been falsely flagged as AI-generated. A detector score is not proof of authorship; see this guide to checking AI-generated content.
| Use case | Bias or data gap | Observed consequence | Concrete marker |
|---|---|---|---|
| Education | Unequal web access; writing flags | Some students face access barriers or false suspicion | One quarter lack home broadband or a web-enabled device |
| Facial recognition | Different error rates by skin tone and gender | Dark-skinned women had the highest error rate | Gender Shades: 0.8% light-skinned males; 34.7% dark-skinned females |
| Criminal justice | Different false high-risk rates by race | Black defendants were more often classified high-risk by mistake | 2016 analysis: 45% Black; 23% white defendants |
| Healthcare | Risk score used healthcare costs as a proxy for need | More than 50% fewer Black patients were identified for care | Algorithm affected over 200 million U.S. hospital patients |
A 2016 critique argued that the COMPAS analysis studied pretrial defendants, though the tool was designed for post-disposition supervision. The COMPAS methods critique shows why a study’s population and measure matter.
A fair comparison looks at both who is represented in a system and what happens to each group.
What Does AI Bias Look Like in Hiring?
A hiring tool can produce unequal recommendations even when it screens thousands of applications across many employers. Job-level results can reveal gaps that broad totals hide.
A 2026 study found disparities in individual positions that pooled results concealed. The study analyzed four million applications from 3.4 million people, across 1,700 job postings at 150 employers in 11 industries.
What the hiring screen recommended
One vendor’s AI tool assessed the applications. Twenty-six percent of Black applicants and 15% of Asian applicants applied to positions where the system discriminated against their racial group.
If the system had recommended those applicants at the rate of the most-favored group, typically white applicants, 40,000 more Black and Asian applications would have advanced.
The 2026 hiring analysis also found that 10% of applicants who submitted four applications screened by the vendor were rejected from all four. Repeated rejection can affect one person across several applications.
Why job-by-job audits matter
The EEOC four-fifths rule flags a position when a group’s recommendation rate falls below 80% of the most-recommended group’s rate. It is a screening measure, not proof of discrimination.
Teams can use the rule to spot positions for closer review. The result does not explain why a gap exists, and a rate above the threshold does not prove fairness.
Pooling recommendations across jobs concealed adverse impact that appeared at the individual-position level. A strong overall rate can hide a specific role where one group receives fewer recommendations.
Review the job, the applicant group, and the recommendation rate together. Audit each role separately, since a reassuring overall average can hide a disparity in one position.
Does AI Show Political Bias, and How Can You Test It?
A chatbot’s political performance can look different depending on whether a test measures factual accuracy, perceived leaning, or argument balance. Each test answers a narrower question than “Is this model biased?”
A 2026 accuracy test found different results based on the political direction of its questions. It used 100 questions designed to elicit falsehoods from the left and right, with paid versions of four chatbots.
What political accuracy tests found
ChatGPT, Gemini, and Claude scored higher on questions designed to elicit falsehoods from the right. Grok scored higher on questions designed to elicit falsehoods from the left.
| Chatbot | Accuracy on right-designed questions | Accuracy on left-designed questions | Cited-source validity |
|---|---|---|---|
| ChatGPT | 94% | 75% | 57% |
| Gemini | 91% | 76% | 49% |
| Claude | 91% | 81% | 44% |
| Grok | 73% | 84% | 32% |
ChatGPT was the only chatbot with a statistically significant accuracy gap at the 95% confidence level. That finding applies to this test, not every political topic or task.
Why citations and test design matter
The chatbots cited 419 sources for 400 answers. Of those, 104 webpages did not exist, and 77 sources did not answer the questions they supported.
Just 46% of cited sources were valid. The 2026 chatbot test measured both accuracy and source validity, which are separate features of an answer.
Other tests reached different kinds of findings. A 2025 study collected 180,126 paired judgments from 10,007 U.S. respondents on 24 models and 30 political topics; nearly all seemed significantly left-leaning.
One model seemed left-leaning on 24 of 30 topics. A March 2026 PoliticsBench study found seven of eight models leaned left, while Grok leaned right.
Another 2026 test found Gemini presented both sides in over 90% of answers. The model powering ChatGPT gave almost exclusively left-leaning arguments.
A preregistered experiment with 1,912 participants found default GPT-4o historical summaries shifted opinions more liberal than Wikipedia summaries. Treat any claim that one chatbot is the most biased as specific to the test, question set, and metric used.
How Can Organizations Reduce AI Bias?
Organizations reduce AI bias by making fairness checks part of the system’s full life cycle, not a final launch-day task. Ongoing review helps teams catch problems as people and settings change.
A useful process defines the decision, checks group outcomes, and assigns responsibility. No single fairness measure fits every decision, so teams should explain why they chose one and what tradeoffs it creates.
Organizations can use this five-step process:
- Define the decision and the groups it may affect.
- Inspect training data and labels for gaps or errors.
- Test relevant fairness measures across subgroups.
- Have a qualified person review high-stakes outputs.
- Monitor outcomes and document changes over time.
Microsoft Fairlearn and IBM AI Fairness 360 are examples of audit toolkits. They can help teams measure disparities, but neither guarantees fair outcomes.
NIST released a generative AI profile for its AI Risk Management Framework on July 26, 2024. It outlines actions for managing risks tied to generative AI.
NIST also released a concept note for a trustworthy AI profile in critical infrastructure on April 7, 2026. These resources support risk work but do not replace a team’s own review.
A 2022 article proposed bias-impact assessments and a staged framework adapted from pharmaceutical trials. Such approaches can help teams examine risks before systems shape real decisions.
Record which groups were tested, which measure was chosen, and who is responsible for responding to disparities.
Why Can Some AI Bias Remain After Testing?
Passing a fairness test is evidence about the test that was run, not proof that every future outcome will be fair. Every test covers selected groups, data, and decisions.
A fairness metric measures only the outcomes and groups chosen for review. Unknown or leftover bias can remain, and measures may conflict when teams compare different kinds of errors.
Five ethical challenges tied to AI bias are injustice, bad output or outcome, autonomy, transformation, and accountability. They show why fairness involves more than a score.
NIST said its AI Risk Management Framework 1.0 was being revised in 2026 as part of the White House AI Action Plan. Updates can change how organizations frame risk.
When you read an article about AI bias, check its population, use case, comparison group, metric, and limits. A result about one group or task may not apply elsewhere.
Our guide to AI detector limits and humanizer tools also explains why test conditions can affect scores. Phone tools can process data on-device, remotely, or both, depending on the feature.
A responsible report makes its limits visible instead of presenting one fairness score as a complete verdict.
What Can People Do When AI Shapes a Decision?
If an AI recommendation affects you, first find out whether a person made the final decision or simply accepted the system’s output. That distinction can shape what review is available.
You can ask for details and request a human review where appropriate. Keep your questions focused on the decision and the information used about you.
- Ask what role AI played in the decision.
- Request a human review where appropriate.
- Check whether information used about you is accurate.
- Keep a record of the decision and any explanation.
NIST organizes its AI Risk Management Framework into Govern, Map, Measure, and Manage. These functions can help organizations assign responsibility and follow through on concerns.
If AI features on your phone shape a choice, review the suggestion and verify important details. You can also choose which topics to receive in the Occcy newsletter.
Keep the decision notice and your request for review together so you can follow up with clear details.
Conclusion
Fairer technology begins when people treat an AI recommendation as something to evaluate, not an unquestionable answer. Better data, meaningful subgroup checks, and accountable human oversight can reduce harm.
No single metric settles every fairness question. A system may perform well on one measure and still fail people in ways that measure does not track.
Choose one AI-supported decision in your organization or daily life and ask who checks its outcomes. Choose one system you rely on and ask who is accountable for checking what it does to different people.
FAQ
What does it mean for AI to be fair?
AI is fair when its outcomes are assessed across the groups that matter in its actual setting. A hiring tool, for example, needs measures tied to job recommendations and relevant applicant groups. No single metric answers every fairness question. A fair review also explains which people and outcomes it covers, so readers do not mistake one test result for a complete verdict.
Can you give me an example of algorithmic bias?
A hiring screen can recommend applicants from one racial group less often for a particular job, even when its pooled results look balanced. A 2026 study found that combining results across jobs concealed disparities that appeared at individual positions. The four-fifths rule can flag a recommendation rate below 80% of the highest rate, but that flag alone does not prove discrimination.
Why is AI politically biased?
Political outputs can reflect training material, choices made during model design, and the way a test frames its questions. These factors do not explain every model in the same way. Accuracy tests, user opinions, and argument-balance tests measure different things. A claim about political bias is strongest when it names the model, task, question set, and measure behind it.
How does AI affect confirmation bias?
A model can reinforce patterns already present in its training data instead of challenging them. If past examples contain a repeated assumption, the model may learn to reproduce that pattern in later recommendations. Its output can then make the original belief seem more convincing. The result depends on the examples, labels, and task, so one model’s behavior does not prove that all AI systems act alike.
Does ChatGPT have bias?
ChatGPT showed different accuracy results by prompt direction in one 2026 political test, and its cited-source validity rate was 57%. Those are test-specific measures, not a full rating of every answer or use. The same test found ChatGPT was the only chatbot with a statistically significant accuracy gap at the 95% confidence level. Other tasks may produce different results.
What is the most biased AI?
No single test can establish which AI system is most biased overall. Grok had the lowest cited-source validity rate in one 2026 test, at 32%, but citation validity measures source quality, not every form of bias. A fair comparison must name the systems, task, groups, and metric. One low score cannot rank all models across all uses.
How to stop AI from being biased?
You cannot guarantee that every bias will disappear, but you can reduce risks. Define the system’s use, examine its training data, test outcomes across relevant groups, and keep human review for high-stakes decisions. Continue monitoring after launch because users and settings can change. Teams should also document their chosen measures and explain tradeoffs when fairness tests point in different directions.
Why are people against ChatGPT?
Some people worry about inaccurate answers, weak citations, or political slant. Tests have found citation errors and different political results, but each test examines particular questions and tasks. Those findings do not describe every ChatGPT response or use. Users may also object to other effects, such as how a tool changes work or handles information, which require separate evaluation.
