Types of AI harm

AI systems can produce outputs or exhibit behaviors that harm users, organizations, or society. Harm can be unintentional—caused by a model’s limitations, insufficient UI mitigations, or unclear communication—but it still erodes user trust and can cause real-world impact.

Understanding harm categories helps product teams apply the right RAI principles, design relevant feedback mechanisms, and prioritize mitigations during design reviews and RAI audits.

---

Harm categories

---

Inaccurate output

Inaccurate output occurs when an AI system generates content that is factually wrong, fabricated, or not grounded in reliable sources. This includes hallucinated citations, incorrect statistics, misleading summaries, or outputs that contradict the source material.

Inaccuracy is one of the most common and consequential harm types, particularly in high-stakes domains like health, legal, or financial contexts. Users who trust inaccurate outputs without verification may make decisions based on false information.

Mitigations include always displaying the approved AI disclaimer (“AI-generated content may be incorrect. Check for accuracy.”), surfacing relevant sources, and designing UI friction points that prompt users to verify before acting.

---

Incomplete output

Incomplete output occurs when an AI system omits critical information, stops short of fulfilling a user’s request, or provides a response that is technically correct but missing context needed for safe use.

Incompleteness is especially harmful when users don’t know what they don’t know—they may act on a partial response as though it were complete. This can compound inaccuracy harms.

Mitigations include scoping AI tasks clearly, using progressive disclosure to surface limitations, and communicating when a response is partial or constrained by system capabilities.

---

Biased output

Biased output occurs when an AI system reflects or amplifies unfair assumptions based on a person’s identity, demographic group, culture, or lived experience. Bias can appear in recommendations, image generation, language, or the framing of information.

Because AI models learn from large datasets, they can inherit and reproduce societal biases at scale. Bias harms are often not visible in individual outputs and may only emerge through patterns across users or use cases.

Mitigations include inclusive content review, diverse testing across user groups, and feedback mechanisms that allow users to flag biased outputs for review.

---

Inappropriate or unsafe output

Inappropriate or unsafe output includes content that is offensive, harmful, discriminatory, or that could facilitate real-world harm. This includes content that targets individuals or groups, promotes dangerous behaviors, or violates community safety standards.

These harms require both model-level guardrails and UI-level mitigations. Product teams should ensure users have clear paths to report this type of content and that feedback reaches the team responsible for model safety.

Mitigations include feedback categorization controls that allow users to report “Unsafe or problematic content,” and clear escalation paths from in-product feedback to safety review processes.

---

Non-transparent output

Non-transparency harms occur when users don’t know they’re interacting with AI, can’t identify what sources informed an output, or aren’t given the information they need to make informed decisions about using or trusting the AI.

This includes missing AI labels or badges, absence of the approved disclaimer, obscured data sourcing, and unclear or missing consent language around data use. Non-transparency is foundational: it undermines the user’s ability to evaluate any other harm type.

Mitigations include always labeling AI-generated content, surfacing sources inline, using approved disclaimer text, and communicating AI scope at entry points. See Responsible AI guidance for detailed UX strategies.

---

Overreliance

Overreliance occurs when users trust AI outputs without applying appropriate critical thinking or verification—often because the UI failed to communicate uncertainty, prompt review, or make verification easy.

Overreliance is a design-level harm as much as a model-level one. Users may rely excessively on AI because the experience feels confident, the disclaimer is easy to miss, or there’s no friction to encourage review before acting.

Mitigations include surfacing accuracy caveats inline, providing clear source links, avoiding design patterns that make AI outputs feel authoritative, and creating friction points that prompt users to review content before applying it.

---

Feedback and harm reporting

Collecting user feedback on AI outputs is one of the most direct ways to surface harm at scale. Feedback mechanisms should allow users to categorize the type of issue they’re reporting. Meaningful feedback categories include:

Feedback should be explicit about what data will be collected: “When you submit feedback, the prompt and response are included.” Vague language like “Help us improve” obscures what users are consenting to share.

For RAI scoring, receiving a 0 or 1 score on the Prevent overreliance on AI principle produces an automatic fail status, regardless of overall score. This reflects how foundational overreliance prevention is to responsible AI design.

---

Resources

For UX strategies, evaluation criteria, and in-product do/don’t examples organized by RAI principle, see the Responsible AI guidance page.

Additional resources:

• RAI UX Quality Excellence Guidelines
• Guidelines for Human-AI Interaction
• Advancing AI trustworthiness: Updates on responsible AI research