The 2025 Foundation Model Transparency Index

Transparency in AI is getting worse, not better: The average transparency score for foundation model developers dropped significantly from 58 out of 100 in 2024 to 40 in 2025, reversing the previous year's progress and indicating that companies are becoming less open about their practices.

Companies hide critical information about training and impact: Foundation model developers are most secretive about their training data, training compute requirements, and how their models are actually used and affect people after deployment—precisely the information the public needs to understand AI risks.

There's a massive gap between leaders and laggards: IBM achieved a score of 95 while xAI and Midjourney scored only 14, showing that transparency is technically feasible but not uniformly adopted, suggesting it's a choice rather than a constraint.

Major AI companies are playing it safe in the middle: The five members of the Frontier Model Forum (the largest AI developers) scored in the middle range, likely avoiding the reputational damage of low scores while also avoiding the effort required to be transparency leaders—indicating they lack sufficient incentive to be truly open.

Current transparency practices won't meet emerging policy requirements: As governments worldwide begin mandating transparency disclosures, the current state of industry practices suggests many companies are unprepared and may require more aggressive policy interventions to comply with public accountability standards.