At the first U.S. Senate hearing on “Oversight of A.I.: Rules for Artificial Intelligence,” leaders from OpenAI and IBM and deep learning critic Gary Marcus agreed that artificial intelligence needs regulation. Their proposals raised a practical question for lawmakers: how can rules respond to risks while leaving room for different kinds of AI development and use?
Agreement on licensing, disagreement on the regulator
Sam Altman, OpenAI’s CEO, and Marcus called for a new U.S. agency to oversee AI. Christina Montgomery, IBM’s Chief Privacy & Trust Officer, argued that existing agencies could handle the job. All three supported licensing companies that operate AI models above a certain level of risk.
The call for regulation from industry figures surprised some senators. U.S. Senator Dick Durbin remarked, “I can't remember when we last had companies come and plead with us to regulate them.” Montgomery said IBM had long supported what she called precision regulation, with AI rules tied to risk.
The disagreement was less about whether oversight is needed than about how to organize it. A new agency could be built around AI-specific reviews and monitoring. Using existing agencies would place responsibility within institutions already in place, while leaving open how well they could cover the risks.
What a new agency could do
Marcus said existing authorities and lawsuits under current laws do not offer enough coverage and can move too slowly to protect what matters. He described a new agency as a nimble monitor that could review AI systems before and after deployment, follow developments and recall products.
He also called for more investment in AI safety research and greater transparency from companies such as OpenAI. Those proposals connect oversight to what happens throughout a model’s life: its safety before release, its behavior after deployment and the information companies share about it.
Altman proposed licensing models above a certain capability threshold. A company could lose its license if it failed to meet safety standards. Before use, models could be tested for capabilities such as replicating themselves or breaking out of a system, with independent external parties auditing the process.
Altman cautioned that broad regulation could burden small AI startups and the open-source community. A threshold-based system, as he described it, would focus licensing on models with greater capabilities rather than treating all AI development the same way.
Rules based on risk and use
Montgomery’s approach focused on particular uses of AI rather than regulating the technology as a whole. She called for clear definitions of risk by use case, different rules for different levels of risk, and an emphasis on transparency and accountability. She also referred to the EU AI Act in her proposals.
That approach puts the context of a system’s use at the center of the regulatory question. A rule designed around a specific use could account for the risks associated with that setting. The hearing did not settle how to define those categories or how responsibilities would be divided among existing agencies.
Altman and Marcus also pointed to risks including election meddling and other targeted influence. They raised possible dangers associated with the advent of general artificial intelligence (AGI). Altman said licensing should consider what models might eventually be able to do, not only their current capabilities.
Senators press for enforceable rules
Most U.S. senators appeared open to a new AI agency, and Peter Welch and Richard Blumenthal voiced support. Blumenthal warned that “Pandora's box does need more than words like 'licensing' and 'new agency.'” The hearing introduced proposals, but lawmakers still faced the harder task of deciding how to shape them.
Blumenthal said rules must fit the risk, avoid harm, and be effective and enforceable. He stressed that enforcement matters and said a new agency would need adequate funding and capable scientists. Those requirements point to a central test for any oversight plan: whether it has the people and resources to turn safety standards into action.
Altman said OpenAI saw no reason at that time to stop training new models, including in discussion of a possible moratorium such as for GPT-5. Instead, he said the company wanted extensive security testing before release. The debate in the Senate left open how such testing would be defined, who would verify it and what should happen when a system falls short.