AI safety and governance landscape (September 2026)
| Governance framework | Status | Key provisions |
|---|
| EU AI Act | Effective 2026 (phased) | Risk-based classification; high-risk AI systems require conformity assessment; banned uses (social scoring, etc.) |
| US Executive Order 14110 (AI Safety) | Effective 2023 (Biden); revised 2025-2026 | Reporting requirements for frontier AI training; safety testing; red-team exercises |
| White House AI Safety Institute | Established Feb 2024 | Voluntary pre-deployment safety testing; standards development |
| NIST AI Risk Management Framework (AI RMF) | Released Jan 2023 (v1.0); updated 2025 | Voluntary framework for AI risk management; widely adopted |
| US AI Bill of Rights (Blueprint) | Released Oct 2022 | Five principles: safe and effective systems, algorithmic discrimination protections, data privacy, notice and explanation, human alternatives |
| UK AI Safety Institute | Established Nov 2023 | Frontier AI safety testing; international coordination |
| China AI Safety Governance Framework | Released 2024 | Algorithm filing requirements; generative AI rules; content controls |
| California SB 1047 (vetoed) | Vetoed Sept 2024 | Would have required safety testing for frontier AI models |
| California SB 53 (enacted) | Effective 2025-2026 | Transparency requirements for large AI developers; safety incident reporting |
Source: White House, EU Commission, NIST, UK AI Safety Institute, California legislature, China CAC (2023-2026).
Industry 'responsible scaling' policies
Major AI labs have adopted Responsible Scaling Policies (RSPs) defining capability thresholds that trigger additional safety requirements:
| AI lab | Capability threshold | Trigger |
|---|
| Anthropic | ASL-2, ASL-3, ASL-4, ASL-5 | Increasingly capable and dangerous models; specific safety mitigations at each level |
| OpenAI | Preparedness Framework (Low/Medium/High/Critical) | Capability evaluations in cyber, CBRN, persuasion, autonomy |
| Google DeepMind | Frontier Safety Framework (v1, v2, v3) | Critical Capability Levels triggering additional safety measures |
| Meta | Frontier AI Framework | Voluntary safety commitments for Llama models |
Source: Anthropic Responsible Scaling Policy; OpenAI Preparedness Framework; Google DeepMind Frontier Safety Framework; Meta Frontier AI Framework (2024-2026).
Each framework is voluntary and lab-specific. Critics argue: (1) thresholds are too high to be triggered by current models; (2) labs set their own safety standards; (3) competitive pressure may lead to under-triggering. Proponents argue: industry-led standards are more practical than government mandates (Anthropic; OpenAI; DeepMind; Meta, 2026).
What 'slowing down' could mean in practice
Possible mechanisms for slowing AI development:
- Compute thresholds: training runs above 10^26 FLOPs trigger regulatory review (similar to Export Control limits).
- Mandatory safety testing: pre-deployment safety evaluations required for frontier models (currently voluntary at US AI Safety Institute).
- Capability-specific moratoria: pause specific capabilities (e.g., fully autonomous AI agents, AI-designed biological agents) while continuing general development.
- International coordination: G7/G20 agreement on frontier AI safety standards (analogous to nuclear non-proliferation).
- Industry self-restraint: voluntary industry agreement on training pauses above capability thresholds.
- Compute-based reporting: frontier AI training runs above compute thresholds must be reported to government (current US EO 14110 requirement).
- Safety incident disclosure: mandatory reporting of AI safety incidents, near-misses, and capability discoveries.
Each mechanism has trade-offs: aggressive slows may disadvantage US labs vs China and other competitors; voluntary approaches may be insufficient; government mandates raise First Amendment and innovation concerns (Bloomberg, September 2026).
What this means for AI's trajectory
Implications of the 'open to slowing' shift for the AI industry:
- OpenAI: could affect competitive position vs Anthropic, Google DeepMind, Meta; could reassure safety-concerned customers.
- AI investors: $1T+ in capex commitments may face lower returns if development slows; could indicate industry maturation.
- AI customers: delayed access to more capable models; trade-off vs safety and reliability.
- China competition: slowing US AI development could allow Chinese AI labs to close capability gap.
- AI safety community: validates case for caution; supports regulatory and self-governance efforts.
- AI policy: may accelerate US AI safety legislation (Congressional proposals pending).
- Public perception: shifts the AI safety debate from 'fringe concern' to 'mainstream position'.
FAQ
Is Sam Altman now pro-regulation?
Altman's comments were measured - he is 'open to slowing down' but did not propose specific regulations. He has historically supported US government engagement on AI safety (testified at Senate hearings, participated in White House voluntary commitments) while criticizing heavy-handed regulation. His current position is more aligned with the 'responsible scaling' middle path than with strong regulation. Whether this represents a genuine policy shift or a strategic response to pressure is unclear (Bloomberg, September 2026).
What did OpenAI's safety team departures mean for AI safety at OpenAI?
OpenAI's safety team departures in 2024 (Jan Leike, Ilya Sutskever, Daniel Kokotajlo, and others) raised concerns about OpenAI's commitment to safety. The departures were widely interpreted as safety priorities being deprioritized relative to product development. The 'Superalignment' team, originally dedicated to long-term safety, was disbanded. OpenAI's Preparedness Framework (2023) was updated in 2025 in response to criticisms. The current shift in Altman's rhetoric may reflect a renewed focus on safety, or a response to regulatory and public pressure (Bloomberg, 2026; public statements from former OpenAI safety team members).
Photo: Steve Jurvetson, CC BY, via Wikimedia Commons (https://upload.wikimedia.org/wikipedia/commons/2/2f/Sam_Altman_speaking_at_TED_%28cropped%29.jpg?utm_source=commons.wikimedia.org&utm_campaign=imageinfo&utm_content=original)
Fazlur Rahman
