AI systems can fail in ways that create real harm for individuals and organizations.

Bias in AI refers to systematic patterns in model outputs that lead to unfair or inconsistent outcomes for certain individuals or groups. AI systems can produce biased or incorrect results because they learn from incomplete or unbalanced data, or reflect past decisions and assumptions.

Even when models function as expected, design choices can reinforce harmful patterns and reduce fairness. Understanding these risks is essential for leaders who want to strengthen governance and prevent unintended consequences.

So what does this look like in practice?

Let's start with data and model bias.

The inaugural International AI Safety Report in 2025 identified data and model bias as a significant harm in AI systems and one that can lead to concrete harm. Harm refers to negative or unfair impacts like unequal treatment, exclusion, or reduced access to opportunities.

What many people miss is that data isn't neutral. Every step — be it what we collect, how we clean, or how we label it — reflects human choices and assumptions.

When models learn from historical data, they can repeat patterns rooted in past inequalities, essentially automating outdated discrimination. For example, if hiring data reflects past bias, an AI screening tool may disadvantage qualified candidates from certain backgrounds.

Incomplete or skewed data adds another layer of risk. Performance scores might reflect subjective opinions rather than true outcomes, or certain groups may appear less often in the dataset than they do in reality.

The result? An AI system that sees an incomplete picture of the world.

Next, biased system design. Even when data is sound, design choices can amplify bias. Design choices about what is predicted and how predictions are used reflect human priorities and trade-offs rather than neutral facts.

Take fraud detection. If a system is optimized only to minimize financial losses, it may tolerate a high number of false positives. Legitimate customers get flagged for review, and if these errors affect certain groups more often, the design itself creates unfair outcomes — even if the model looks effective overall.

Then there are hallucinations and unreliable outputs. Generative AI has introduced new challenges — like hallucinations, outputs that are fabricated or incorrect.

These errors often occur when systems operate outside their intended scope, or without enough context. Without safeguards, hallucinations can mislead users, distort decisions, and damage trust when incorrect outputs are shared internally or externally.

Finally, privacy and security risks. AI systems often rely on personal or sensitive data.

This creates privacy and security risks. These include unauthorized or excessive data use, re-identification of individuals from anonymized datasets, and technical threats such as model theft, prompt injection, or adversarial attacks. These vulnerabilities can lead to harmful outputs and reputational damage if not addressed.

Identifying, classifying, and understanding AI risks

About