At a two-hour Senate Judiciary Committee hearing, three AI researchers argued that policymakers face risks in both waiting too long and regulating too hastily. Their proposals differed, but shared a central idea: build the research and oversight needed to manage AI as it develops.
The panel included Anthropic co-founder Dario Amodei, UC Berkeley’s Stuart Russell and longtime AI researcher Yoshua Bengio. They discussed near-term safeguards, the limits of current detection methods and the need for government coordination.
Build the evidence for safety rules
Amodei urged lawmakers to address weaknesses in the AI supply chain, including hardware bottlenecks and vulnerabilities tied to geopolitical factors, intellectual property or safety concerns. He also called for testing and auditing processes, drawing a comparison with vehicles and electronics.
Those processes, he said, need a rigorous battery of safety tests. But the science behind defining risks and setting standards is still in its infancy. Amodei argued for a regulator able to adapt as AI changes, rather than a fixed framework that may become outdated.
Bengio also emphasized that rules need stronger scientific foundations. He called for funding AI safety research globally and for cooperation among nations, including independent audits of AI capabilities and alignment. In his view, researchers need more knowledge as well as collaboration that is not driven by competition.
All three researchers pointed toward investment in basic research as an early priority. Better evidence, they suggested, would help make testing, auditing and enforcement more rigorous and less reliant on outdated or industry-suggested ideas.
Restrict access and set clear obligations
Bengio proposed limiting who can access large-scale AI models and creating incentives for security and safety. He also urged policymakers to track the computing power behind these systems and who can access the hardware needed to produce them.
That concern extends beyond large organizations. Bengio noted that pre-trained large models can be fine-tuned, meaning that a bad actor may not need a giant data center or extensive expertise to cause harm. Oversight focused only on the biggest developers could miss risks that arise when models are adapted for other uses.
Russell outlined rules focused on how systems interact with people and what they can do. He proposed an absolute right to know whether someone is interacting with a person or a machine, a ban on algorithms that can decide to kill human beings at any scale, and a kill switch if AI systems break into other computers or replicate themselves. He also called for systems that break rules to be withdrawn from the market, like an involuntary recall.
Amodei’s most immediate concerns included misinformation, deepfakes and propaganda during an election season. Russell focused on personalized “external impact campaigns”: using information about individuals to generate tailored disinformation at large scale. He argued that such campaigns could have a greater effect than broadcasting false information that is not personalized.
Coordination and detection remain difficult
Efforts to label, watermark and detect AI-generated material were described as fragmented and rudimentary. The panel’s view offered little reason to expect those approaches to provide much protection in time for the election the committee was discussing.
Bengio suggested restricting social media accounts to actual human beings who had identified themselves, ideally in person. The article notes that this proposal would likely face serious obstacles. It also highlights the tension between addressing automated activity and imposing requirements that may be difficult to put into practice.
For government oversight, Bengio recommended that the U.S. and other countries each create a single regulatory entity. The aim, he said, would be better coordination and less bureaucratic slowdown. Senator Blumenthal said the hearing was intended to inform the creation of a government body able to move quickly, “because we have no time to waste.”
A narrow path between delay and haste
The hearing’s broad message was that policy should move promptly while remaining responsive to new developments. Amodei compared the AI industry to airplanes a few years after the Wright brothers flew: regulation was clearly needed, but it had to adapt as the field advanced.
The researchers’ proposals ranged from supply-chain security and access limits to shutdown mechanisms and market recalls. Underneath those differences was a shared dependency: regulators need reliable ways to define risks, test systems and enforce standards.
Without that foundation, moving slowly could leave people exposed to misuse, while moving too quickly could hamper the industry without providing effective safeguards. The researchers’ emphasis on basic research and international cooperation was a call to develop that foundation alongside regulation.