Page 39 - The Annual AI Governance Report 2025 Steering the Future of AI

The Annual AI Governance Report 2025: Steering the Future of AI

as independent third-party evaluation often faces challenges due to a nascent reporting culture,
limited infrastructure, and insufficient legal and technical protections for researchers.

System/Model Card Disclosure and Safety Frameworks: System and model cards — a type
of structured documentation that details a model's capabilities and limitations, as well as its
training data and safety considerations — have emerged as a best practice for promoting
transparency and the responsible deployment of AI. Companies are increasingly publishing
such disclosures to inform users, regulators and external researchers about model risks and
mitigations. Alongside these disclosures, safety frameworks set out organisational policies for
risk assessment, emergency procedures, ongoing monitoring and human oversight. Yet, the
164
effectiveness of these disclosures and frameworks hinges on the quality, completeness and
accessibility of the information provided, as well as the capacity to translate abstract statements
into tangible results. In order to ensure consistency across this emerging practice, calls have
165
been made for standardised, well-defined metrics and unified approaches. 166

Limits of Self-Governance Approaches: Empirical research on AI governance shows that
voluntary, industry-led codes of conduct rarely lead to meaningful accountability. Without
external audits or sanctions, companies prioritise speed-to-market over risk mitigation, resulting
in a failure to curb bias, disinformation, and other issues. In May 2023, OpenAI's chief executive,
167
Sam Altman, told the US Senate that “it is essential to develop regulations that incentivize AI
safety,” even proposing a federal licensing regime for frontier models. However, at a follow-
168
up hearing in May 2025, he warned that requiring government approval before release would
be 'disastrous', marking a significant policy U-turn. Organisational studies of industry practice
169
describe this shift as indicative of 'minimum viable ethics', whereby corporate AI ethics teams
have limited authority, which is defined by product launch schedules and revenue targets.
170
This leaves voluntary governance unable to enforce rigorous standards of safety, transparency,
and accountability. Meta-analyses of 84 public- and private-sector AI ethics frameworks show
that high-level principles, when not backed by audits or legal sanctions, rarely produce durable
protections against bias, disinformation, and other externalities, underscoring the structural
limits of self-regulation in the AI sector. 171

6.4 Open Source and Open Weight AI: Trajectories, Debates, and
Global Practices

The debate around open source and open weight AI models has become central to current
discussions about access, accountability, and innovation in AI development. While “open
source” traditionally refers to models whose architecture, training data, and weights are publicly
available, “open weight” models typically allow access to pre-trained weights but not necessarily

164 See for instance: Introducing the Frontier Safety Framework. (2024, May 17). Google DeepMind.
165 Mukobi, G. (2024b, August 5). Reasons to doubt the impact of AI risk evaluations. arXiv.org.
166 Pistillo, M. (2025, January 27). Towards frontier safety policies plus. arXiv.org.
167 Maclure, J., & Morin-Martel, A. (2025). AI Ethics’ institutional turn. Digital Society, 4(1).
168 U.S. Senate Committee on The Judiciary Subcommittee on Privacy, Technology, & The Law. (2023). Written
testimony of Sam Altman, Chief Executive Officer of OpenAI, before the U.S. Senate Committee on the
Judiciary Subcommittee on Privacy, Technology, & the Law.
169 De Vynck, G., & Tiku, N. (2025, May 9). AI execs used to beg for regulation. Not anymore. The Washington
Post.
170 Ahlawat, A., Winecoff, A., & Mayer, J. (2024, September 11). Minimum viable ethics: from institutionalizing
industry AI governance to product impact. arXiv.org.
171 Mittelstadt, B. (2019). Principles alone cannot guarantee ethical AI. Nature Machine Intelligence, 1(11),
501–507.

30