The societal stakes of artificial intelligence (AI)’s deployment continues to rise exponentially. Alongside the promised efficiency and innovation, the proliferation of AI systems has also generated a growing number of incidents where these systems malfunction, behave unexpectedly, or are implicated in harm.
An AI incident can be broadly defined as any situation where an AI system causes, contributes to, or is associated with an outcome that negatively impacts individuals, groups, systems, or the environment. This can include technical failures, ethical concerns, breaches of privacy, discriminatory outcomes, or unintended social consequences. Importantly, incidents may not always involve system errors – they can also arise from misuse, poor design choices, or the contexts in which the AI is deployed.
Tracking such incidents is vital to understanding the real-world implications of AI, fostering transparency, and informing governance. Two significant efforts in this arena are the OECD AI Incidents and Hazards Monitor (AIM) and the AIAAIC Repository (AIAAIC). Though still relatively under-the-radar outside specialist circles, these initiatives are critical tools in the global conversation around AI risk, accountability and trustworthiness.
AI incident tracking
The impetus for tracking AI incidents
The idea of tracking AI incidents stems from a growing recognition that conventional methods of auditing and regulating technology were falling short in the AI context. Unlike traditional software systems, AI applications often exhibit emergent, opaque behaviours due to their reliance on machine learning, vast datasets, and probabilistic inference. Consequently, ensuring that these systems remain safe, fair, and transparent requires novel approaches.
A string of high-profile failures, ranging from racist facial recognition systems to AI-based hiring tools that discriminated on gender grounds, further fuelled public concern. These incidents were not isolated bugs; they often revealed deeper structural issues about how AI systems were trained, deployed and governed. Media reporting and academic scrutiny began surfacing these cases more frequently, but without a systematic means to log, classify and learn from them, the opportunity for shared learning was lost. Enter the AI incident trackers.
The OECD AI Incidents and Hazards Monitor (AIM)
Origins and development
The OECD (Organisation for Economic Co-operation and Development) launched AIM as part of a broader commitment to responsible AI. It builds on the OECD Principles on AI (adopted in 2019) which were the first international standards agreed by governments for the responsible stewardship of trustworthy AI. The incident monitor initiative emerged as a logical step towards operationalising these principles, particularly those related to accountability, transparency and safety. Initially developed in collaboration with academic and civil society partners, including the Partnership on AI and the Center for the Advancement of Trustworthy AI, the AIM monitor was envisioned as a public utility: an open-access repository where researchers, regulators, developers and journalists could monitor and learn from incidents.
Features and scope
The monitor catalogues incidents where AI systems are reported to have caused, or contributed to, real-world harm or unintended consequences. These include issues related to algorithmic bias, lack of robustness, privacy breaches, safety failures and more. Each entry in the monitor typically includes a description of the incident, the type of harm caused, the context of the AI system’s deployment, and any known follow-up actions taken. Although the OECD emphasises quality over quantity, its monitor is not exhaustive. It focuses on well-documented cases from credible sources, which limits its comprehensiveness but enhances its trustworthiness.
The AIAAIC Repository (AIAAIC)
Mission
The AIAAIC is an independent resource, launched as a private initiative in 2019 to better understand to better understand the reputational risks of artificial intelligence, the AIAAIC Repository has now evolved into an open source, public interest initiative. Its foundational goal is to support a robust ecosystem of AI accountability by systematically documenting incidents where AI systems cause or contribute to harms. The AIAAIC is meant to serve as a collective memory for the AI field.
Structure and usage
The AIAAIC Repository entries include detailed metadata such as:
-
Incident dates
-
Locations
-
System developers
-
Types of AI technologies involved
-
Types of harms reported
The database supports search and filtering by these attributes, allowing researchers to identify patterns over time or across geographies. It also includes a taxonomy of incident types and causes that continues to evolve as more data is collected.
Comparing the AIM and AIAAIC
While both initiatives share a commitment to transparency and responsible AI, they differ in several key respects:
-
Institutional grounding: The OECD AIM has formal international backing and is integrated with broader governance frameworks. AIAAIC is grassroots and community-driven.
-
Scope and submission model: AIM uses a curated approach with controlled input channels; AIAAIC allows more open submissions from a broader set of contributors.
-
Focus: AIM leans towards policy-relevant cases that align with governance interests. AIAAIC includes a wider array of incidents, from technical failures to social impacts.
The initiatives are complementary, not competitive. Each supports the other by offering different lenses on the evolving landscape of AI risk.
Types of AI incidents tracked
Based on data from both the OECD and AIAAIC, AI incidents can be broadly categorised into the following groups:
1. Algorithmic bias and discrimination
Incidents in this category involve AI systems producing outputs that are biased against certain groups, often along lines of race, gender, age or socioeconomic status.
Example: A facial recognition system used by law enforcement in the US misidentified Black individuals at significantly higher rates than white individuals, leading to wrongful arrests.
2. Safety and robustness failures
These are cases where AI systems behave unpredictably or fail in critical tasks, particularly in high-stakes contexts like autonomous driving or medical diagnostics.
Example: A self-driving car failed to recognise a pedestrian due to sensor misinterpretation, resulting in a fatal accident.
3. Privacy and surveillance abuses
AI systems are often involved in large-scale data processing, raising significant privacy concerns.
Example: AI-powered recruitment software scraped data from social media without consent, profiling candidates in ways that violated data protection laws.
4. Misinformation and content manipulation
Generative AI and recommendation algorithms can be implicated in spreading false or misleading information.
Example: A recommendation engine on a social media platform amplified conspiracy theories during a political election, contributing to public unrest.
5. Economic harms and workforce impacts
Some AI systems lead to job displacement or unfair treatment of workers.
Example: An algorithm used to schedule gig workers’ shifts systematically disadvantaged those who could not respond immediately, disproportionately affecting carers and disabled workers.
6. Legal and regulatory breaches
These incidents involve AI systems that inadvertently or deliberately violate legal norms.
Example: An automated credit scoring system was found to contravene anti-discrimination laws by systematically offering lower credit limits to women.
7. Ethical concerns and social harm
This broad category covers AI uses that may be legal but raise ethical red flags.
Example: A university used predictive analytics to monitor student behaviour and flag potential dropouts, raising questions about autonomy, fairness and surveillance.
Challenges and criticisms
Despite their value, AI incident trackers face several challenges:
-
Underreporting: Many incidents go unreported due to proprietary constraints, lack of media coverage, or fear of reputational damage.
-
Verification: Ensuring that submitted incidents are accurate, well-documented and contextualised is a non-trivial task.
-
Bias and coverage gaps: Certain regions, sectors or populations may be underrepresented.
-
Interpretation: Differing legal, cultural and ethical norms make it hard to categorise and evaluate incidents in a universally accepted way.
Nonetheless, both the OECD AIM and AIAAIC have made strides in improving their methodologies, and continued community engagement is helping address these gaps.
Importance of tracking AI incidents
AI incident tracking is not just an academic or bureaucratic exercise; it should be a cornerstone of responsible AI governance. The trackers:
-
Enable learning from failure, a critical aspect of system safety.
-
Support public accountability by making opaque harms visible.
-
Inform policy and regulation with empirical evidence.
-
Drive technological improvement by highlighting areas for redress.
-
Enhance trust in AI by showing that harms are being monitored and addressed.
They also help shift the narrative from AI as a purely speculative risk to a present, real-world concern that requires ongoing, iterative oversight.
As AI systems become more capable and ubiquitous, the question is not whether incidents will occur, but how we respond when they do. The OECD AIM and the AIAAIC represent the infrastructure for this response, documenting where and how things go wrong so that future deployments can be safer, fairer and more trustworthy. Their value lies not just in what they track, but in the message they send: that AI accountability is not optional, and that learning from the past is the best way to safeguard the future.