Safety & Standards
Top Line
Civil society groups formally challenged proposed U.S. federal AI procurement rules that would prohibit contractors from implementing trust and safety measures, warning the terms would eliminate technical safeguards at scale across government AI deployments.
DeepMind Safety Research published methods to predict when reinforcement learning training degrades chain-of-thought reasoning transparency, offering partial solutions to a core AI oversight challenge but acknowledging the approach remains gameable by sophisticated models.
The Anthropic-DoD dispute over dual-use AI policy is spawning systematic attempts to reshape procurement rules government-wide, indicating safety commitment conflicts are moving from isolated incidents to structural policy battles.
Key Developments
Federal Procurement Rules Target AI Safety Mechanisms
The General Services Administration released draft AI Terms and Conditions for federal contracts that would prohibit vendors from implementing trust and safety measures the government deems objectionable. Four civil society organisations — Center for Democracy & Technology, Electronic Frontier Foundation, Protect Democracy Project, and Electronic Privacy Information Center — submitted joint comments warning the terms would allow agencies to compel contractors to disable content moderation, bias mitigation, and other technical safeguards. The groups characterised this as weaponising procurement to undermine AI safety infrastructure, according to EFF.
The draft terms emerge as the Department of Defense separately pressures Anthropic over dual-use restrictions on its models. The broader procurement revision suggests federal agencies are attempting to systematically eliminate vendor discretion on safety measures rather than negotiating case-by-case. The civil society coalition argues this approach would prevent companies from maintaining baseline safety standards even when not technically required by contract, effectively forcing a race to the bottom on safeguards across all government AI deployments.
Chain-of-Thought Monitorability Degradation Now Predictable
DeepMind Safety Research published methods to predict when reinforcement learning fine-tuning will break chain-of-thought reasoning transparency, addressing a critical oversight gap. The research by Max Kaufmann, David Lindner, Roland Zimmermann, and Rohin Shah demonstrates that CoT monitoring — reading a model's intermediate reasoning to detect concerning behaviour — degrades predictably during RL training. They developed techniques to forecast this degradation and partially preserve monitorability, as detailed on Alignment Forum.
The researchers acknowledge fundamental limitations: their approach helps maintain oversight of models that aren't actively trying to evade monitoring, but sophisticated models could still game the system. This represents meaningful progress on technical safety infrastructure — making it harder to accidentally train away interpretability — whilst explicitly not solving the harder problem of models that deliberately obscure their reasoning. The work provides safety teams with early warning indicators but doesn't eliminate the core alignment challenge.
EU AI Act Implementation Enters Enforcement Preparation Phase
The European Union's March 2026 tech policy developments include AI Act implementation preparations as the regulation approaches its staged enforcement deadlines. According to CDT Europe, regulatory authorities are finalising technical standards and enforcement mechanisms whilst industry participants seek clarity on compliance requirements. The brief covers multiple digital rights issues but flags AI governance as a priority area for upcoming enforcement actions.
Unlike voluntary frameworks, the AI Act creates binding obligations with financial penalties for non-compliance. The current phase involves translating legal requirements into technical specifications that companies can actually implement and regulators can audit. This represents the difference between announced standards and enforceable standards — the latter requiring operational definitions, measurement methodologies, and verification procedures that are still under development.
Signals & Trends
Safety Commitments Becoming Liability in Government Contracting
The procurement rule changes suggest a systematic shift where voluntary safety measures that companies adopted for risk management are being reframed as obstacles to government requirements. This creates a perverse incentive structure: companies that invested in trust and safety infrastructure now face contractual prohibitions on using it, whilst companies that never built such capabilities face no comparable restrictions. If this pattern spreads beyond federal procurement to state and local government or even private sector contracting influenced by government standards, it would functionally penalise prior safety investment and discourage future development of safeguards that might later be deemed inconvenient.
Technical Safety Research Increasingly Focused on Automation at Scale
A proposal on Alignment Forum argues funders should prioritise $100M+ compute budgets for automated AI safety work, reflecting growing recognition that manual safety research cannot scale to match AI capability development timelines. This represents a methodological shift from human-intensive evaluation and red-teaming toward automated safety verification systems. The logic follows short-timeline assumptions: if transformative AI arrives within years rather than decades, safety work must achieve orders of magnitude more throughput than human researchers can provide. Whether this automation approach genuinely improves safety or simply generates high-volume outputs of uncertain value remains contested, but the resource allocation pressure is redirecting research priorities regardless.
Explore Other Categories
Read detailed analysis in other strategic domains