Commentary. Artificial intelligence regulation has reached the point where delay is no longer a neutral choice. The question is not whether governments should control every model or whether innovation should be frozen. It is whether the companies building the most powerful systems should be allowed to assess their own risks, disclose only what they choose and define success on their own terms.
Our view is clear: the answer should be no. Governments should require independent safety evaluations for the most capable AI systems before and after deployment, while leaving room for smaller developers and lower-risk applications to operate under lighter rules.
That approach is gaining political support. A Navigator Research survey cited by CFO Dive found that 66% of registered voters wanted more AI regulation in October, up from 60% in December 2025. A Reuters/Ipsos poll reported that 80% of Democratic voters and 61% of Republican voters supported stricter federal oversight.
Those numbers do not settle the policy debate, but they make one fact difficult to ignore: public concern is crossing partisan lines. People are encountering AI in employment, education, health care, customer service and elections, often without knowing when an automated system has influenced a consequential decision.
The case for independent testing
Massachusetts is considering a proposal that would require major developers, including OpenAI, Anthropic and Meta, to submit their most powerful models to independent evaluations three times a year for risks of catastrophic harm. The measure could make Massachusetts the first state to authorize third-party scrutiny of frontier systems, according to The Daily Free Press.
The principle is sound. A model-maker has commercial incentives to release a product, expand its capabilities and reassure customers. An outside evaluator has a different responsibility: to test whether the system can be manipulated, leak sensitive information, assist dangerous activity or behave unpredictably under pressure.
Independent review would not guarantee safety. It would, however, create evidence that regulators, customers and the public could examine. It could also establish a common baseline, preventing companies from competing by making the weakest safety promises.
“The push for federal regulation of artificial intelligence is gaining momentum ahead of the Nov. 3 midterm elections,” CFO Dive reported.
The strongest counterargument
Opponents have legitimate concerns. AI development moves faster than legislation, and badly designed rules could lock in today’s technology while making it harder for smaller firms to compete with established giants. Mandatory testing could become an expensive compliance ritual rather than a meaningful safety measure. Regulators may also lack the technical expertise to judge frontier systems, particularly when the systems themselves change through updates.
There is a further risk of fragmented regulation. If Massachusetts, the federal government, the European Union and other jurisdictions impose incompatible requirements, companies could divert investment or limit services in some markets. Industry-led standards can be faster, more flexible and better adapted to technical realities.
These arguments deserve more than dismissal. Regulation that treats a classroom chatbot like a system used in critical infrastructure would be both wasteful and counterproductive. Nor should governments publish sensitive testing details that could help malicious actors exploit a model.
A workable middle ground
The answer is a risk-based framework. Independent testing should focus on the small number of developers training the most capable general-purpose systems. Evaluators should be accredited but not selected solely by the companies being tested. Results should be reported to a public authority, with confidential technical details protected where disclosure would create security risks.
Rules should also cover deployment, not only release. A model that appears safe in a laboratory can produce different risks when connected to financial systems, search tools, code repositories or autonomous agents. Companies should have to report serious incidents, preserve records and explain how they respond when safeguards fail.
Lawmakers should resist both extremes: uncritical faith in corporate self-regulation and sweeping rules written around speculative catastrophe. The immediate goal is accountability for foreseeable harms and credible preparation for more dangerous capabilities.
European lawmakers have already rejected reliance on self-regulation and called for international action, according to the Center for Democracy and Technology Europe. That debate will not be resolved by one state or one election. But the direction should be unmistakable: innovation can continue, while the most powerful systems face scrutiny proportionate to their reach.
For our newsroom, that is the line between responsible optimism and wishful thinking. AI may deliver enormous benefits. The companies building it should not be asked to prove that alone.