Ethics, Bias & Fairness

Why AI ethics matters

AI is a category of machine-based systems, not a synonym for machine learning. [nist-ai-glossary] Machine learning is one major family of methods. Rule-based systems, search, optimization, and hybrids can also produce unfair or harmful outcomes. When those systems affect loans, hiring, healthcare, or criminal-legal decisions, the harm is real even if the method is not “learning from examples.”

Responsible development is not optional. It is part of building and using the system.

The bias problem

Bias can enter through data, objectives, measurement, and deployment context — not only through a training set. [selbst2019]

Historical bias: Past decisions encoded in data can be replayed. If a hiring process historically favored one group, a model trained on those records can reproduce that pattern.

Representation bias: Benchmarks and training sets that under-represent a group often produce worse performance for that group. Buolamwini and Gebru (2018) evaluated three commercial gender-classification systems and found darker-skinned females were the most misclassified group, with error rates of up to 34.7%, while the maximum error rate for lighter-skinned males was 0.8%. [buolamwini2018]

Measurement bias: A convenient proxy (for example, arrest records as a stand-in for crime) can punish over-policed communities rather than measuring the intended outcome.

Aggregation bias: One model for everyone can hide that a system fails for a specific population.

Sources of bias in machine-learning pipelines

For ML systems specifically, bias can enter at every stage:

1. Problem definition: Who defines success? Whose values are encoded?
2. Data collection: What gets measured? Who is represented?
3. Feature engineering: Which attributes are used? Do proxies hide bias?
4. Model training: What objectives are optimized?
5. Evaluation: Which groups are tested? What metrics are reported?
6. Deployment: Who has access? How are decisions acted on?

Fairness is a sociotechnical problem, not only a model-weight problem. [selbst2019]

Fairness: multiple definitions

There is no single definition of “fair.” Different definitions can conflict:

Demographic parity: Equal positive rates across groups

Equalized odds: Equal true-positive and false-positive rates across groups

Calibration: A stated probability should mean the same thing across groups (if the system says 70%, the outcome should occur about 70% of the time for each group)

Individual fairness: Similar individuals should be treated similarly

Except in trivial cases, a classifier cannot satisfy common combinations of these criteria at once when base rates differ. [chouldechova2017] [kleinberg2016]

Transparency and explainability

People affected by AI decisions often need an explanation:

Local explanations: “Why was this application denied?” Global explanations: “How does this model generally work?” Model cards: Document purpose, limits, and disaggregated evaluation [mitchell2019]

Deep models can be hard to explain. That is a reason for documentation and evaluation, not a reason to skip them.

Privacy considerations

AI systems and personal data are often in tension:

• Some methods need data, which may be personal
• Models can memorize sensitive information
• Inference can reveal private attributes

Whether prompts are stored or used for training is provider- and settings-dependent. There is no universal rule that “nothing is stored” or that “everything is used for training.”

Do not paste passwords, government IDs, financial account numbers, health records, other people's private data, or secrets and API keys into an AI tool unless that exact product is approved for that data.

Technical mitigations (used where they actually apply):

• Differential privacy
• Federated learning
• Data minimization

Accountability

Who is responsible when an AI system causes harm?

Developers: Built the system Deployers: Put it into practice Operators: Made specific decisions Data providers: Supplied training data

Clear accountability chains are essential and often missing.

Consent and autonomy

People should have meaningful choice about AI affecting them:

• Informed about AI involvement
• Able to opt out where a real alternative exists
• Given a human path for high-stakes decisions
• Not manipulated by persuasive systems they cannot inspect

Societal impacts

Beyond individual harms, AI systems can affect labor, market concentration, civic information, surveillance, and energy use. Those are empirical questions. Do not attach an invented confidence percentage to them.

Governance snapshot — reviewed August 27, 2026

This is a dated snapshot, not legal advice, and not a complete map of world law.

United States federal policy. Executive Order 14110 (October 30, 2023) is not current policy. It was revoked on January 20, 2025. [eo-14110-rescinded-2025] On June 2, 2026, Executive Order 14409 directed federal work on AI-enabled cybersecurity and a voluntary framework for pre-release government access to designated covered frontier models. The order states that it does not authorize mandatory licensing, preclearance, or permitting for new AI models. [eo-14409-2026]

California SB 1047. The bill did not become law. Governor Gavin Newsom vetoed it on September 29, 2024. It should not be listed as a current California AI safety statute. [ca-sb1047-veto-2024]

For a longer governance snapshot, see AI Safety & Alignment.

Principles for responsible AI

Organizations often publish principles such as beneficial, fair, transparent, accountable, privacy-preserving, safe, and human-centered. Principles are not the same thing as evaluated practice.

Moving from principles to practice

• Include affected communities
• Document decisions and limitations, including model cards [mitchell2019]
• Test for disparate performance across groups
• Create feedback and appeal paths
• Do not treat an overall accuracy number as proof of fairness