Page 58 - The Annual AI Governance Report 2025 Steering the Future of AI

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

Critical Technologies, IEEE). It is important to translate abstract values into verifiable, operational
properties to make ethical commitments enforceable.

A related difficulty, as pointed out by a participant, was the fluidity of concepts, such as defining
"autonomous systems," which even the EU's AI Act doesn't explicitly clarify. Context Chapter 1: Global

There is furthermore a tension, as described by Artemis Seaford (Head of AI Safety, ElevenLabs),
between the principle-based, "top-down" approach often seen in traditional institutions versus
the "bottom-up," problem-solving approach common in Silicon Valley tech startups. She
argued that the optimal solution involves meeting in the middle, likely at the regulatory layer.

1.6 Trust

Trust is seen as paramount to AI adoption. Without trust, even the most powerful systems risk
rejection by users and citizens. Panelists identified multiple dimensions of trust: explainability of
model decisions, robustness under stress, fairness across populations, and privacy safeguards.
Civil society voices emphasized that transparency is central.

Yet concerns were also raised that transparency has regressed. Model cards, once a standard
for documenting the limitations and risks of models, have become less informative in newer
releases, as pointed out by Udbhav Tiwari (VP Strategy and Global Affairs, Signal).

This tension between commercial pressures for secrecy and public demand for clarity was seen
as a fault line that governance must address. However, it is increasingly getting more difficult
for developers in a highly hyped industry to be transparent and honest about the limitations
and drawbacks of their AI models.

49