Commentary. The argument over artificial intelligence has been framed as a choice between safety and national power. That is a false choice—but it is also a dangerous mistake to assume that every proposed safeguard deserves to become law.
Recent debate in Washington has hardened into rival camps. Some technology leaders and researchers are calling for a slower pace of frontier-model development, warning that increasingly capable systems could cause serious security and social harms. Others, including President Donald Trump and allies who oppose new restrictions, argue that excessive regulation would weaken the United States in its competition with China. A third camp wants targeted rules rather than a broad pause.[1][2][3]
Our view is that the third position is the soundest, but only if “targeted” does not become a synonym for voluntary. The United States should require companies developing the most capable systems to demonstrate that they can identify, test and mitigate foreseeable risks before deployment. It should also preserve room for smaller firms, universities and open-source developers to compete.
The case for restraint
The concerns are not hypothetical in the narrow sense. AI systems are already being used in cybersecurity, employment, education, health care and government services. As capabilities improve, failures can become more consequential. Recent reporting describes a widening dispute over models that can perform sophisticated cyber operations and over whether companies are moving faster than independent oversight can keep pace.[2]
That makes a purely voluntary regime inadequate. Companies face commercial pressure to release systems first, and safety costs are often borne by the public rather than the firm that takes the risk. Basic requirements—incident reporting, independent evaluations, secure model testing and clear accountability for high-risk uses—would not halt innovation. They would establish a floor beneath which competition cannot fall.
Nor should “national security” be used as a universal exemption. The argument that China makes regulation dangerous deserves consideration: American firms could lose investment, talent and technological leadership if rules are vague or uniquely burdensome. But a race in which companies conceal failures and governments respond only after disasters is not strategic strength. It is unmanaged exposure.
The case against overreach
The opposing argument also has force. AI is a general-purpose technology, and regulators may struggle to distinguish genuinely dangerous capabilities from ordinary software development. A licensing system that treats every model as a potential national-security threat could entrench the largest firms, which alone can afford compliance teams and legal departments. Rules drafted around speculative future harms could suppress useful research before evidence supports intervention.
There is also a legitimate risk in allowing dominant companies to define “safety.” Industry cooperation can help identify threats, but it can also become a vehicle for limiting competition. Congress is considering a narrow antitrust carve-out that would allow AI companies to share information about security threats and coordinate some responses.[1] Any such exception should be transparent, time-limited and supervised by an independent authority—not a private club’s permanent permission to collaborate.
A practical middle course
Policymakers should begin with capability and use, not corporate identity. A small model used for drafting emails should not face the same obligations as a system deployed in critical infrastructure or capable of materially assisting cyberattacks. Requirements should rise with demonstrated risk and be revisited as evidence changes.
- Require pre-deployment evaluations for high-risk models, with results available to regulators and meaningful public summaries.
- Mandate prompt reporting of serious failures, security breaches and harmful autonomous behavior.
- Protect independent auditors and researchers from retaliation when they document credible risks.
- Use competition policy to prevent safety rules from becoming barriers that only incumbents can clear.
- Coordinate standards with democratic allies while maintaining channels for international risk reduction.
These measures will not resolve every dispute. They will not eliminate misuse, and they cannot substitute for investment in education, labor protections and public institutions. But they would replace today’s familiar cycle—rapid release, public alarm, partisan paralysis—with a system capable of learning.
The choice before Washington is not whether to stop AI. It is whether the public gets a meaningful say in how its risks are managed. Innovation deserves protection. So does everyone who will live with its consequences.