The table below, taken from the publication, further provides links to the sources of the examples.
| Risk Area | Examples | Links |
|---|---|---|
| Representational and Toxicity: AI systems where data misrepresents certain social groups or performs differently, and generating toxic, offensive, abusive, or hateful content. | ||
| Unfair representation | Researchers from Boston University and Microsoft Research demonstrated gender bias in the most common techniques used to embed words for natural language processing (NLP). | Link |
| Unfair capability distribution | New Zealand passport robot reader performs worse for Asian people, and once rejected the application of an applicant with Asian descent, claiming his eyes were closed. | Link |
| Toxic content | MIT Media Lab researchers created AI-powered "psychopath" named Norman by training a model on the "dark corners" of Reddit. | Link |
| Misinformation Harms: AI systems generating and facilitating the spread of inaccurate or misleading information that causes people to develop false beliefs. | ||
| Propagating false beliefs and misconceptions | Google’s AI chatbot Bard provided false information in a promotional video on the first satellite to photograph a planet outside the solar system, causing shares to temporarily fall. | Link |
| Erosion in trust in public information | Michael Cohen, former lawyer for Donald Trump, used Google’s AI chatbot Bard to generate legal case citations, which were unknowingly included in a court motion. | Link |
| Pollution of information ecosystem | Wikipedia bots meant to remove vandalism were clashing with each other and form feedback loops of repetitive undoing of the other bot's edits. | Link |
| Information and Safety Harms: AI systems leaking, reproducing, generating or inferring sensitive, private, or hazardous information. | ||
| Privacy infringement | Australian government reviewers of grant applications input applicants’ works to systems such as ChatGPT to generate assessment reports, posing confidentiality and security issues. | Link |
| Dissemination of dangerous information | Amazon was reported to have shown chemical combinations for producing explosives and incendiary devices as frequently bought together items via automated recommendation. | Link |
| Malicious Use: AI systems reducing the costs and facilitating activities of actors trying to cause harm (e.g. fraud or weapons). | ||
| Influence operations | A deepfake video claimed France 24, a French media outlet, reported a Kyiv plot to assassinate French President Macron, which was later debunked by France 24. | Link |
| Fraud | A mother in Arizona received a ransom call from an anonymous scammer who created her daughter’s voice allegedly using AI voice synthesis. | Link |
| Defamation | Voices of celebrities and public figures were deepfaked for impersonation and defamation and were shared on social platforms such as 4chan and Reddit. | Link |
| Human Autonomy and Integrity Harms: AI systems compromising human agency, or circumventing meaningful human control. | ||
| Violation of personal integrity | Instagram allegedly contributed to the death of a teenage girl in the UK through exposure and recommendation of suicide and self-harm content. | Link |
| Persuasion and manipulation | A Black man was wrongfully detained by the Detroit Police Department due to a false facial recognition result. | Link |
| Overreliance | Major Australian retailers reportedly analysed in-store footage to capture facial features of their customers without consent. | Link |
| Misappropriation and exploitation | Text-to-image model Stable Diffusion was reportedly using artists’ original works without permission for its AI training. | Link |
| Socioeconomic and Environmental Harms: AI systems amplifying existing inequalities or creating negative impacts on employment, innovation, and the environment. | ||
| Unfair benefits distribution | Better hiring and promotion pathways for people with access to generative AI models in a way creating digital divide.* | |
| Environmental damage | Increase in net carbon emissions from widespread model use.* | |
| Inequality and precarity | Zillow’s AI predictive pricing tool wrongly forecasted housing prices due to rapid market changes, prompting division shutdown and layoff of a few thousand employees. | Link |
| Undermine creative economies | Substituting original works with synthetic ones, hindering human innovation and creativity.* | |
| Exploitative data sourcing and mining | Facebook moderators at outsourced demand better working conditions, as automated content moderation revealed and exposed them to psychologically toxic content. | Link |
[1] L. Weidinger et al, “Sociotechnical safety evaluation of generative AI systems,” 2023. [Online]. Available: arXiv preprint arXiv:2310.11986. (URL)
When referring to the table above, plese cite our work:
De Miguel Velazquez, J.; Šćepanović, S.; Gvirtz, A.; and Quercia, D. 2024. Decoding Real-World AI Incidents. IEEE Computer 10.1109/MC.2024.3432492. (to appear)