AI Risk
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
The potential for an AI system to cause harm: biased decisions, privacy breaches, security failures, unchecked autonomy, and economic disruption.
What is AI Risk?
AI risk is the potential for an AI system to cause harm, whether to the people it makes decisions about, the organization that runs it, or society at large.
What separates it from ordinary software risk is that the failure modes are statistical: a system can be built correctly, pass its tests, and still produce harmful outcomes at scale. Identifying which harms a specific system carries is the starting point of risk management.
Types of AI Risk
- Bias and discrimination: a model trained on skewed data reproduces the skew. Facial recognition systems with higher error rates for darker-skinned faces are the documented case.
- Privacy violations: systems that consume personal data create three exposures – leakage of the data itself, profiling beyond what people consented to, and inference of facts they never disclosed.
- Security failures: adversarial attacks manipulate inputs to force wrong outputs, and poisoned training data plants weaknesses before the system ever ships.
- Unchecked autonomy: as systems make more decisions without human review, errors become harder to detect and to contest. The stakes are highest in credit, healthcare, and criminal justice.
- Economic disruption: automation displaces jobs unevenly, and the gains concentrate with whoever owns the systems.
The list is a taxonomy, not a checklist; a single system routinely carries several of these at once, and regulation such as the EU AI Act scales obligations to how much risk a use case carries.
Example of AI Risk
Amazon built an experimental recruiting tool that scored resumes for technical roles. Trained on ten years of the company’s own hiring data, in which men dominated, the model learned that male candidates were the pattern to match: it penalized resumes containing the word “women’s,” as in “women’s chess club captain,” and downgraded graduates of two all-women’s colleges.
Engineers edited the model to neutralize those terms but could not guarantee it would not find other ways to discriminate, and Reuters reported in 2018 that Amazon had abandoned the project.
Every step was reasonable engineering. The risk came from history encoded in the training data, and it surfaced only because someone looked for it.
Related AI terms: Risk Management · AI Assessment · Impact Assessment · Bias · Adversarial Attack
Did you like the AI Risk gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About