Bottom Line
Google DeepMind, Microsoft, and xAI have officially joined the CAISI (Center for AI Standards and Innovation) frontier model safety testing program, agreeing to provide early access to new models before public release for national security assessment.
While this appears “voluntary,” in the context of the US government’s recent AI policy shift, it marks that AI model regulation is moving from “industry self-discipline” to “substantive government review”.
What Happened
CAISI Program Expansion
CAISI is a US government-led frontier AI model safety assessment organization. Previously, OpenAI and Anthropic had joined through negotiation, and now Google DeepMind, Microsoft, and xAI have officially joined.
| Company | Status | Models Covered |
|---|---|---|
| OpenAI | Already joined (after renegotiation) | GPT series |
| Anthropic | Already joined (after renegotiation) | Claude series |
| Google DeepMind | Newly joined | Gemini series |
| Microsoft | Newly joined | Copilot/Phi series |
| xAI | Newly joined | Grok series |
Policy Reversal Timeline
| Time | Event | Direction |
|---|---|---|
| January 2025 | Trump signs executive order, revoking Biden-era AI regulations | Hands-off |
| 2025 | CAISI established, initially with only 1-2 companies | Probing |
| April-May 2026 | Google/Microsoft/xAI join CAISI | Tightening |
| May 2026 | White House discusses AI model pre-release review executive order | Stronger regulation |
Within 16 months, policy shifted from “comprehensive deregulation” to “multiple companies voluntarily accepting government testing,” to discussing “mandatory pre-release review” — the speed and magnitude of this shift exceeded market expectations.
Key Interpretations
1. The Real Drivers Behind “Voluntary”
Although CAISI is positioned as a “voluntary program,” participating companies face real policy pressure:
- Executive order threat: The White House is discussing a pre-release review executive order — not participating in the “voluntary” program could mean mandatory review
- National security concerns: AI models’ cybersecurity, biosecurity, and disinformation generation capabilities have become core government concerns
- Industry consensus: Leading companies recognize that establishing a credible safety assessment framework is necessary for the industry’s sustainable development
2. Scope of CAISI Testing
Based on public information, CAISI’s assessment covers at least:
- Cybersecurity: Whether models could be used for automated cyber attacks
- Biosecurity: Whether models could be used to design biological weapons or dangerous pathogens
- Disinformation: Assessment of models’ ability to generate high-quality fake content
- Alignment and safety: Behavioral prediction of models in edge cases
3. Impact on Competitive Landscape
CAISI expansion has subtle impacts on AI industry competition:
| Dimension | Impact |
|---|---|
| Entry barriers | New entrants may need to bear equivalent safety testing costs |
| Release cadence | Government testing may extend the cycle from development to release |
| Chinese models | CAISI primarily covers US companies; Chinese models may face different regulatory frameworks |
| Open source models | Safety assessment of open-weight models remains an unresolved challenge |
Actionable Recommendations
| Role | Recommendation |
|---|---|
| AI Startups | Monitor CAISI evaluation standards and processes, prepare safety compliance materials in advance |
| Enterprise Users | Evaluate whether vendors participate in CAISI testing as a reference factor for selecting AI services |
| Developers | Follow CAISI’s published evaluation reports to understand safety performance across models |
| Policy Researchers | Track CAISI’s policy evolution path from “voluntary” to “mandatory” |
Risk Factors
- CAISI’s specific evaluation standards and processes are not fully public
- “Voluntary” framework may transition to mandatory requirements under political pressure
- Safety testing may slow innovation pace, affecting US AI competitiveness
- Chinese AI models in international markets may face additional compliance requirements