Full Breakdown
AI Companies Sign New Safety Pledges, but Metrics Remain Vague
By Drooid · · How we work
New AI Safety Commitments at the White House
Earlier this week, leading AI firms renewed voluntary safety pledges at the White House, echoing a similar set signed in 2023. The 2023 commitments were initially signed by Amazon, Anthropic, Google, Inflection, Meta, Microsoft and OpenAI and later expanded. The 2026 round was signed by Anthropic, Google, Meta, Nvidia, OpenAI and xAI.
How the 2023 and 2026 Pledges Differ
The 2023 document covered eight broad safety themes, including internal and external risk testing, sharing safety information, protecting model weights, encouraging vulnerability reporting (e.g., bug bounties), watermarking AI-generated media, publishing transparency reports, investing in research on societal harms, and developing frontier AI for climate, health and cybersecurity challenges. By contrast, the 2026 commitments focus on four governance-oriented actions: internal monitoring of frontier models, an internal oversight team, partnership with an external auditor or evaluator, and an independent board committee to address issues raised by internal or external teams. The newer pledge also mentions regular meetings to “establish standards and best practices,” but provides no guidance on how those standards will be judged.
Measurement Gaps Highlighted by Experts
A paper from the National Institute of Standards and Technology (NIST) notes that voluntary standards are difficult to evaluate because the goals are stated only in broad terms. Julia Lane, writing for NIST, argues that without a “theory of change” linking specific activities to measurable outcomes, it is impossible to determine whether the pledges actually improve safety, public trust, or beneficial adoption.
Proposed Econometric Evaluation Methods
Lane proposes using established econometric techniques—such as matching, difference-in-differences, and synthetic-control methods—to compare companies that adopt the standards with comparable firms that do not. These approaches aim to isolate the effect of the standards themselves, rather than merely observing differences that may stem from pre-existing company characteristics.
Implications for Future Governance
If the AI industry continues to rely on self-regulation, the ability to demonstrate tangible results will become a key factor in maintaining public confidence and encouraging broader adoption of AI technologies. Independent evaluation frameworks, as suggested by NIST, could provide the necessary accountability mechanism for future AI safety commitments.