2018 - 2025 AI Failure Mode and Effects Analysis
Over ten months, BRITE Institute applied failure mode and effects analysis to more than 700 documented cases in which people were harmed by AI systems between 2018 and 2025.
What is this research about?
AI failures are usually reported one incident at a time, as isolated malfunctions or cautionary headlines. That makes it hard to see the patterns underneath. BRITE Institute set out to change this by applying failure mode and effects analysis (FMEA), a discipline long used in aviation, manufacturing, and medical device safety, to real-world AI harm. Over ten months, our team examined more than 700 individual cases from 2018 to 2025 in which people suffered harm involving an AI system. For each case we identified how the system failed, why the failure occurred, and what effects followed. From this analysis we developed a failure mode taxonomy for AI that classifies failures by [their origin, mechanism, and consequence], so that harms which look unrelated on the surface can be understood, compared, and addressed systematically.
Why is this research important?
Mature safety disciplines do not start with principles. They start with a shared understanding of how things go wrong. Engineers in safety-critical industries design around known failure modes because those modes have been catalogued, studied, and agreed upon. AI has no equivalent foundation. Incidents are recorded inconsistently, described in different terms by different parties, and rarely analyzed beyond the individual case. As a result, organizations repeat failures others have already experienced, risk assessments rely on guesswork rather than evidence, and regulators lack a structured basis for deciding where controls are needed. A grounded taxonomy, built from what has actually happened to real people, is a prerequisite for moving AI safety from good intentions to engineering practice.
What are the conclusions?
Our analysis of more than 700 cases produced a taxonomy of distinct AI failure modes, organized into categories [e.g., data, model, deployment context, human–AI interaction, and organizational governance]. Among our key findings was that while technical failures were more numerous than human / governance failures, governance failures tended to result in more severe outcomes. This demonstrates that improving model accuracy alone will not adequately control these risks. Governance, post-deployment surveillance, human oversight, subgroup monitoring, and incident reporting are central controls.
Our research also found that while atients were the most frequently affected by AI failures in healthcare, the exposure of instiiutitions was also great, appearing in 52% of cases.
Finally, performance drift of AI was shown to be a significant issue. An AI that had historically performed well may suddenly create a significant risk due to software changes or enviromental changes.