2026 Transparency Report on Foundation Model Impacts

As foundation model providers increasingly share information privately, public transparency on impacts grows more essential. But progress is uneven.

Realizing the benefits of AI requires an ecosystem of trust and accountability, with transparency as its foundation. Yet voluntary commitments and recent legislation are increasingly focused on disclosing information behind closed doors. As foundation models shape our society, private disclosures alone cannot foster the trust and accountability that public transparency enables.

This report measures the progress of 13 organizations in publicly documenting the impacts of their foundation models, drawing on a review of more than 150 papers, articles, websites, and reports. It builds on our 2023 Guidance for Safe Foundation Model Deployment and our 2025 progress-tracking methodology.

Overall Findings

1. Several leading organizations are defining what information to share and how. However, the rest of the field often lags in adopting information-sharing practices, raising questions about how to ensure that society benefits from whole-of-field impact information.
2. PAI’s analysis reveals a scattered and incomplete landscape of public impact data. The public has access to more information on AI usage and potential labor market impacts than ever, but we don’t know the prevalence of harms, serious incidents, or broader societal impacts.

Progress Analysis: Where the field stands across four core practices

We evaluated how well “practice leaders” are adopting four information-sharing practices, and how the rest of the field is keeping pace.

Key Insights: What the data reveals

Providers that host their own services can track why and how people use foundation models without compromising privacy. This information can directly inform public policy.

Providers are signaling risks to entry-level workers and shifting skill requirements, but have not published research on risks related to worker abuse and labor sourcing.

Providers report environmental impacts differently, so standards are crucial.

Providers report AI-generated or uploaded Child Sexual Abuse Material (CSAM), but the prevalence of other foundation model misuses remains unknown.

Some security incidents are reported voluntarily — both privately and publicly, where appropriate. This accelerates threat detection, but broader harms to users and society are often overlooked.

Providers prioritize collecting feedback on security risks over broader safety concerns.

Scientific and policy guidance for open model impact reporting faces significant implementation challenges that reduce actual reporting.