Textbook of AI

The Frontier AI Safety Commitments are a set of voluntary, public undertakings by leading AI developers to identify, evaluate and mitigate severe risks from their most capable models, and to report transparently on their progress. The commitments emerged from the international policy process initiated by the UK Bletchley Park AI Safety Summit (November 2023) and crystallised at the Seoul AI Safety Summit (May 2024), where 16 leading developers signed the Seoul declaration's commitments.

The Seoul commitments (May 2024)

Each signatory committed to:

1. Identify, assess and manage risks across the model lifecycle, including pre-training, deployment and post-deployment.

2. Set out thresholds at which severe risks are considered intolerable absent mitigations, i.e., publish a Responsible Scaling Policy or equivalent.

3. Articulate how mitigations will be developed, including, where appropriate, pausing development if mitigations cannot keep pace.

4. Provide transparency on their approach, including external review where appropriate.

5. Be accountable to other actors, including governments and civil society.

Signatories included Anthropic, OpenAI, Google DeepMind, Microsoft, Meta, Amazon, Mistral, xAI, IBM, Inflection, Cohere, Naver, Samsung, Technology Innovation Institute (UAE), G42 (UAE), and Zhipu AI (China).

Subsequent developments

The Paris AI Action Summit (February 2025) broadened the agenda, with mixed results: not all attendees signed the final declaration, and the Trump administration's January 2025 executive order rescinded much of the Biden 2023 framework, complicating US engagement. The third summit (India, late 2025) refocused on practical evaluation infrastructure, with the network of national AI Safety Institutes (UK, US, Japan, Singapore, Korea, Canada, EU, France) playing a coordinating role.

Critique

• Voluntary, unenforceable; signatories can revise their own commitments.

• Self-evaluation, most evaluations are performed by the labs themselves, raising obvious conflicts.

• Slow, by design, the international process moves at the pace of diplomatic consensus; capabilities advance faster.

• Coverage gaps, the commitments cover frontier developers but not the long tail of fine-tuners, deployers, or adversarial users.

Defenders argue that the commitments are nonetheless a necessary scaffold on which enforceable rules (the EU AI Act's general-purpose-AI provisions, the UK AI Safety Bill in committee in 2026, the US AISI's testing authority) can be built.

Status

As of 2026, the Frontier AI Safety Commitments remain the most widely-adopted public framework for frontier-AI risk management. Each signatory's adherence is patchy and unevenly verified, but the practice of publishing system cards, RSPs, and pre-deployment evaluation results is now industry norm, attributable in large part to the Seoul process.

References

• UK Government (2024). Frontier AI Safety Commitments, AI Seoul Summit 2024.

• Bletchley Declaration (2023).

• International AI Safety Report (2024, 2025), chaired by Yoshua Bengio.