MIT study finds many large companies do not disclose AI risks

MIT Sloan said in a press release that a new working paper estimates 32% of the world's 100 largest companies and 62% of the 1,000 largest have not publicly acknowledged any AI risks. Researchers reviewed more than 500 annual reports, AI policies, and other public documents.

The researchers assessed disclosures against 24 categories in the MIT AI Risk Repository. Discrimination, weapons and cyberattacks, fraud and scams, privacy loss, and the devaluation of human creativity accounted for more than half of all acknowledgments.

None of the companies acknowledged risks involving power centralization or competition that could encourage deployment at the expense of safety. Among the 100 largest companies, 2% acknowledged AI misalignment, 1% acknowledged risks from interactions among multiple agents, 4% acknowledged dangerous capabilities, and 16% acknowledged AI security vulnerabilities.

Companies in sectors that experts rated as extremely vulnerable to a risk acknowledged that risk only 23% of the time. Among public companies that develop general purpose AI, 15% acknowledged misalignment risks, while 85% acknowledged discrimination risks.