Partnership on AI

Multi-stakeholder non-profit, funded by technology companies, with a board split between corporate partners and non-profit ones. It runs the ai-incident-database, which is the reason it is paged here.

The governance structure explains the database's most consequential design choice. PAI's own membership contains the companies whose systems appear in the incidents, so a single shared taxonomy — one authority deciding what counts as an AI harm and how severe it was — would have to be negotiated between parties with opposite interests in the answer. The AIID's namespaced, competing-taxonomy architecture routes around that, and the paper describing it says as much.

Sits in the same soft-governance layer as oecd-ai-principles: voluntary, industry-participatory, producing infrastructure and norms rather than binding rules, and unable to compel a report the way eu-ai-act Article 73 can.