The EU is considering adding a requirement to the forthcoming AI Act for AI developers to undertake an impact assessment on new AI projects to assess their potential impact on health, safety, fundamental rights and the environment. However, the state of the art in anticipating impacts from AI is still in its infancy [1] so its use in high stakes regulatory assessment requires accelerating the evidence base available for impact assessment under the AI Act, especially maintaining what could be regarded as reasonable foreseeable risks that AI providers must take responsibility for mitigating. This project will explore how existing semantic models for AI risk assessment [2] could be use to capture impact assessment from publicly available AI incident reports [3] in a way that would be useful for the AI Act complaince and oversight.
[1] https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3829759