A bipartisan coalition of researchers, former officials, and civic leaders has released a concrete roadmap for artificial intelligence governance, arguing that the United States can no longer afford to improvise. The Pro-Human Declaration, signed by hundreds of notable figures, lays out a strict, pro-safety framework at the very moment a high-profile rift between the Pentagon and Anthropic exposed just how thin America’s AI rulebook remains.
Organizers describe a public mood shift toward guardrails that is both swift and broad. Citing recent national polling, they say upward of 95% of Americans oppose an unchecked sprint toward superintelligence. Whether Congress acts on that sentiment is the open question the declaration is designed to answer.
What the AI Governance Roadmap Specifically Demands Now
The document is blunt about the crossroads: either race to replace humans, handing critical decisions to opaque systems, or build AI that extends human capability while keeping people in charge.
Its five pillars center on:
- Human control
- Anti-monopoly safeguards
- Protection of the human experience
- Civil liberties
- Real legal accountability for developers
Among the strongest provisions are:
- A temporary halt on superintelligence efforts until there is scientific consensus on safety and explicit democratic authorization
- Mandatory “off-switches” and operational oversight for powerful models
- A ban on architectures with self-replication, autonomous self-improvement, or resistance to shutdown
In short, do not build what you cannot control, and prove safety before scale.
The approach echoes familiar regimes. Drugmakers cannot ship a compound without clinical evidence and regulatory approval. The declaration effectively argues for an AI equivalent: pre-deployment testing with standardized evaluations, auditable documentation, and post-release monitoring. NIST’s AI Risk Management Framework provides a technical scaffold; the EU’s AI Act has already codified pre-market conformity checks for higher-risk systems. This roadmap would bring U.S. practice in line with those realities.
Why the timing matters for U.S. AI safety and policy
The release lands amid a rare public dispute over control of frontier AI. After a clash over access terms, the Pentagon labeled Anthropic a “supply chain risk,” while OpenAI quickly reached a separate arrangement with defense officials. The episode underscored how vendor policies — not democratically set rules — currently define the limits of government use and safety standards. As one policy analyst told a major newspaper, this was the country’s first real debate about who holds the keys to advanced systems.
Child safety is the coalition’s clearest pressure point. The declaration seeks required pre-release testing for chatbots and companion apps targeting minors, gauging risks such as self-harm prompts, emotional manipulation, or exacerbation of anxiety and depression. Public-health context is sobering: federal surveys have reported historic highs in teen depressive symptoms in recent years. In that light, the case for minimum safety baselines — the AI equivalent of seat belts — becomes easier to make.
The signatory list is intentionally cross-ideological, spanning former national security leaders, technologists, and figures from both conservative and progressive circles. The shared premise: regardless of politics, humans must retain final say over systems that can influence markets, national defense, and the information environment.
How implementation could work across agencies and industry
Near-term execution does not require inventing policy from scratch. The White House has already directed reporting for high-compute training runs and safety tests for dual-use capabilities. Regulators could align those thresholds with standardized, third-party evaluations, require incident reporting for model failures, and compel auditable training and inference logs for frontier systems.
Accountability can ride on familiar rails:
- Apply product liability and negligence standards to AI-enabled harms
- Require safety cases before deployment in high-risk use cases
- Mandate independent red-teaming and post-market surveillance
Insurers are ready-made enforcers — they can price risk and demand stronger controls as a condition of coverage, accelerating best practices without waiting for a new agency.
The roadmap also targets concentration of power. The heaviest models depend on scarce compute and proprietary data held by a handful of companies. Opening access to secure public compute via national labs, expanding privacy-preserving data partnerships, and enforcing interoperability can curb lock-in while supporting safety research. Internationally, coordination through standards bodies and model evaluation benchmarks would reduce regulatory arbitrage.
The tradeoffs to watch as AI safety rules take shape
Critics worry that a moratorium on certain research could blunt innovation or push it offshore. Supporters counter that unforced errors would be costlier. The latest Stanford AI Index estimates tens of billions of dollars in annual private AI investment, a sign that capital will chase clarity; predictable rules can stabilize, not stifle, progress. The World Economic Forum projects that 44% of workers’ skills could be disrupted within a few years, underscoring the need for worker upskilling and transition plans alongside safety rules.
Even boosters of the declaration acknowledge challenges: reaching scientific consensus on superintelligence risks; defining “capable of self-replication” in code; and ensuring public input that is more than a checkbox. But those are engineering and governance problems, not reasons to punt. The core test is straightforward: can developers show, with evidence, that powerful systems behave within human-defined bounds — and can society turn them off if they do not?
That is the heart of this roadmap. It asks policymakers to lock in pre-deployment testing, clear lines of human control, and democratic consent before escalating capability. In a field famous for moving fast, it is a call to move correctly — while there is still time to choose the road.
