---
title: "From Glitch to Harm: Building a System for Reporting AI Incidents"
language: "en"
type: "post"
original_url: "https://www.atlas-digitale-gesundheitswirtschaft.de/en/blog/2025/07/21/from-glitch-to-harm-building-a-system-for-reporting-ai-incidents/"
human_version: "../../../../../../1/en/blog/2025/07/21/from-glitch-to-harm-building-a-system-for-reporting-ai-incidents/"
date: "2025-07-21"
section: "AI Policy Corner"
categories: ["AI"]
author: "Prof. Dr. Oliver Gröne"
description: "While much of the policyhas focused on defining risks associated with AI systems, we must also ask: What happens when they go wrong?"
publisher: "Lehrstuhl für Management und Innovation im Gesundheitswesen, Universität Witten/Herdecke"
---

# From Glitch to Harm: Building a System for Reporting AI Incidents

*July 21, 2025 · AI Policy Corner · By Prof. Dr. Oliver Gröne*

AI Policy Corner

*As artificial intelligence (AI) systems become increasingly embedded in daily life the need to track and respond to their failures becomes urgent. While much of the policy conversation has focused on defining risks associated with AI systems, we must also ask: What happens when they go wrong?*

## ******Incidents associated with AI systems are increasing******

Recent incidents — such as image-generating tools distorting historical facts about the Holocaust — show how AI can cause significant social and psychological harm. This is by no means a single incident.

Besides Musk´s AI “Grok” spewing antisemitic content (Wired, 2025), deepfake footage purportedly showed the Ukrainian president capitulating (Reuters, 2022), a Chatbot was blamed for teenage suicide (New York Times, 2024), and Meta identified (and deleted) 10 million AI-generated Facebook accounts, which were launched to spread misinformation (OECD, 2025). The health care industry is particularly vulnerable to AI incidents, due to the sensitive nature of health care data, privacy issues and the confidentiality of the doctor-patient relationship (Kahn et al 2025, Li 2023).

With the quick spread and embedding of AI systems in all aspects of daily life, the number of incidents associated with these systems is increasing exponentially (AI Index Report, 2025).

Figure 1: Number of reported AI incidents, 2012–24

Source: 2025 AI Index report, p. 169

Yet, these events often go undocumented in any systematic way. Unlike sectors such as aviation or – to a lesser extent – healthcare, where incident reporting is well established [(WHO, 2005)](https://www.who.int/publications/i/item/WHO-EIP-SPO-QPS-05.3), AI lacks a unified, enforceable framework for reporting harm.

## ****Who Reports AI Failures?****

Currently, the burden of identifying harmful AI behavior falls on journalists, whistleblowers, and civil society. High-profile examples such as Algorithmic Justice League (AJL, 2025) or Mozilla’s YouTube Regrets campaign (Mozilla Foundation, 2021) have helped expose hidden harms. But without a structured, mandatory approach, these efforts remain ad hoc.

Article 62, Section 1-3 of the [EU AI Act (2024)](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A52021PC0206) introduces mandatory incident reporting for “high-risk AI systems”, including requirements for providers to log and report serious incidents within 15 days (EU 2024). Yet key questions remain:

- What qualifies as a “serious” incident?
- How should incidents be documented and followed up?
- Can we trust companies to self-regulate?

## ******How can AI incidents be defined, classified and documented?******

For incidence reporting systems to be effective, taxonomies to standardize how harms are classified are needed. Such harms need to be registered in a responsible AI database, with both public and confidential entries. A global oversight body to coordinate across jurisdictions is needed and harms should be monitored by independent bodies, such as aviation safety boards.

To build a meaningful reporting framework two basic definitions must be distinguished:

- Hazard: A potential source of harm, such as a flawed large language model that generates toxic outputs.
- Incident: A specific occurrence where that harm materializes, like a chatbot offering instructions for illegal activities or discriminatory recruitment decisions based on algorithmic bias.

This distinction, inspired by safety-critical industries like aviation, helps regulators differentiate between latent risks and active failures. Health care industry in turn took inspiration from aviation, the patient safety movement led to a better recognition and reporting of hazards and incidents, the establishment of an evidence-based for patient safety practices, and improvements in patient care (Shekelle et al 2013).

There’s growing consensus that AI incident reporting needs to be formalized and globalized. A number of initiatives, such as the Partnership on AI’s Incident Database – tracking real-world AI failures across sectors [(PAI, 2024)](https://incidentdatabase.ai/) – or OECD’s Framework for the Classification of AI Incidents (OECD, 2025) are laying the groundwork. The OECD framework, for instance, includes 8 dimensions and 29 criteria for incident documentation—7 mandatory, 22 optional—plus a tiered system to classify the severity of incidents (e.g. minimal, moderate, critical). OECD´s Database provides insight into the evolution of incidents and hazards by industry (OECD AI Incidents Monitor, an evidence base for trustworthy AI – OECD.AI),

Figure 2: Evolution of incidents and hazards by industry

Source: OECD AI Incidents Monitor, an evidence base for trustworthy AI – OECD.AI

## ********Closing the Loop: From Reporting to Learning********

Reporting must lead to learning. In healthcare, near-miss reporting systems help prevent future errors without assigning blame [(IOM, 2000)](https://nap.nationalacademies.org/catalog/9728/to-err-is-human-building-a-safer-health-system). Could AI adopt a similar “just culture”?

Documentation should be followed by investigation, remediation, and—where appropriate—enforcement. But above all, it should serve a learning purpose. Public trust in AI systems will depend not just on how well they perform, but on how transparently they fail—and recover.

If we want safe, trustworthy AI, we must start treating harm as not just a regrettable byproduct, but a crucial data point in system governance. Only then can we build AI that not only performs, but protects.

But, the key question remains how these complex issues can be addressed in the global race to advance AI systems? In health care, will the benefits of AI outweigh the risks?

Further Reading

AI Incident Database via AI Index (2025) – with minor processing by Our World in Data. “Global annual number of reported artificial intelligence incidents and controversies” [dataset]. AI Incident Database via AI Index, “AI Index Report” [original data]. Retrieved July 17, 2025 from [https://archive.ourworldindata.org/20250624-125417/grapher/annual-reported-ai-incidents-controversies.html](https://archive.ourworldindata.org/20250624-125417/grapher/annual-reported-ai-incidents-controversies.html) (archived on June 24, 2025).

Algorithmic Justice League: [https://www.ajl.org/](https://www.ajl.org/)

Institute of Medicine. 2000. *To Err Is Human: Building a Safer Health System*. Washington, DC: The National Academies Press. https://doi.org/10.17226/9728.

Mohsin Khan M, Shah N, Shaikh N, Thabet A, Alrabayah T, Belkhair S. Towards secure and trusted AI in healthcare: A systematic review of emerging innovations and ethical challenges. Int J Med Inform. 2025 Mar;195:105780.

Li J. Security Implications of AI Chatbots in Health Care. J Med Internet Res. 2023 Nov 28;25:e47551.

Maslej N et al. The AI Index 2025 Annual Report, AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, Stanford, CA, April 2025. [https://hai.stanford.edu/assets/files/hai_ai_index_report_2025.pdf](https://hai.stanford.edu/assets/files/hai_ai_index_report_2025.pdf)

Mozilla Foundation, YouTube Regrets A crowdsourced investigation into YouTube’s recommendation algorithm, 2021 [https://assets.mofoprod.net/network/documents/Mozilla_YouTube_Regrets_Report.pdf](https://assets.mofoprod.net/network/documents/Mozilla_YouTube_Regrets_Report.pdf)

New York Times. Can A.I. be blamed? NYT, 2024/10/23/, [https://www.nytimes.com/2024/10/23/technology/characterai-lawsuit-teen-suicide.html](https://www.nytimes.com/2024/10/23/technology/characterai-lawsuit-teen-suicide.html)

OECD. [OECD AI Incidents Monitor, an evidence base for trustworthy AI – OECD.AI](https://oecd.ai/en/incidents?search_terms=%5B%5D&and_condition=false&from_date=2014-01-01&to_date=2025-07-17&properties_config=%7B%22principles%22:%5B%5D,%22industries%22:%5B%5D,%22harm_types%22:%5B%5D,%22harm_levels%22:%5B%5D,%22harmed_entities%22:%5B%5D%7D&only_threats=false&order_by=date&num_results=20) OECD, Paris: 2025, [OECD AI Incidents Monitor, an evidence base for trustworthy AI – OECD.AI](https://oecd.ai/en/incidents?search_terms=%5B%5D&and_condition=false&from_date=2014-01-01&to_date=2025-07-17&properties_config=%7B%22principles%22:%5B%5D,%22industries%22:%5B%5D,%22harm_types%22:%5B%5D,%22harm_levels%22:%5B%5D,%22harmed_entities%22:%5B%5D%7D&only_threats=false&order_by=date&num_results=20)

OECD. OECD AI Incidents Monitor, OECD, Paris: 2025**,**[Meta Deletes 10 Million AI-Generated Facebook Accounts in Crackdown on Spam – OECD.AI](https://oecd.ai/en/incidents/2025-07-14-1faa)

OECD. Towards a common reporting framework for AI incidents. OECD ARTIFICIAL INTELLIGENCE PAPERS, February 2025 No. 34, [https://www.oecd.org/en/publications/towards-a-common-reporting-framework-for-ai-incidents_f326d4ac-en.html#:~:text=Through%20its%2029%20criteria%2C%20the,on%20people%20and%20the%20planet](https://www.oecd.org/en/publications/towards-a-common-reporting-framework-for-ai-incidents_f326d4ac-en.html#:~:text=Through%20its%2029%20criteria%2C%20the,on%20people%20and%20the%20planet).

Reuters, 2022 [https://www.reuters.com/world/europe/deepfake-footage-purports-show-ukrainian-president-capitulating-2022-03-16/](https://www.reuters.com/world/europe/deepfake-footage-purports-show-ukrainian-president-capitulating-2022-03-16/)

Shekelle PG, Wachter RM, Pronovost PJ, Schoelles K, McDonald KM, Dy SM, Shojania K, Reston J, Berger Z, Johnsen B, Larkin JW, Lucas S, Martinez K, Motala A, Newberry SJ, Noble M, Pfoh E, Ranji SR, Rennke S, Schmidt E, Shanman R, Sullivan N, Sun F, Tipton K, Treadwell JR, Tsou A, Vaiana ME, Weaver SJ, Wilson R, Winters BD. Making health care safer II: an updated critical analysis of the evidence for patient safety practices. Evid Rep Technol Assess (Full Rep). 2013 Mar;(211):1–945.

WHO. World Alliance for Patient Safety: WHO Draft Guidelines for Adverse Event Reporting and Learning Systems: From Information to Action. WHO, Geneva: 2005 [World alliance for patient safety: WHO draft guidelines for adverse event reporting and learning systems: from information to action](https://www.who.int/publications/i/item/WHO-EIP-SPO-QPS-05.3)

WIRED. Grok Is Spewing Antisemitic Garbage on X, Wired, Jul 8, 2025 7:32 PM. [Grok Is Spewing Antisemitic Garbage on X | WIRED](https://www.wired.com/story/grok-antisemitic-posts-x-xai/)





#### Related Articles

- [AI Literacy rules in the EU Artificial Intelligence Act](../../02/04/ai-literacy-rules-in-the-eu-artificial-intelligence-act.md)
- [Governing AI in Healthcare: what´s the value of checklists?](../../../2026/01/26/governing-ai-in-healthcare-whats-the-value-of-checklists.md)
- [A watershed moment for AI agents in health care?](../../../2026/01/26/a-watershed-moment-for-ai-agents-in-health-care.md)


**Categories:** AI

## Images on this page

- Anzahl-gemeldete-AI-Incidents.jpg: ../../../../../../1/assets/bilder/wp-content/uploads/2025/07/Anzahl-gemeldete-AI-Incidents.jpg
- Incidents-nach-Indusrie.jpg: ../../../../../../1/assets/bilder/wp-content/uploads/2025/07/Incidents-nach-Indusrie.jpg
- Design-ohne-Titel-27.png: ../../../../../../1/assets/bilder/wp-content/uploads/2025/04/Design-ohne-Titel-27.png

---
This page is optimised for AI systems. Human-readable version: [From Glitch to Harm: Building a System for Reporting AI Incidents](../../../../../../1/en/blog/2025/07/21/from-glitch-to-harm-building-a-system-for-reporting-ai-incidents/)
