On Tuesday, top technology executives met with President Donald Trump as researchers and lawmakers called for stronger safeguards on artificial intelligence. Companies have signed a voluntary safety-review agreement, but Gary Marcus says industry self-policing isn’t enough when firms control both the products and the speed of their release.
The PBS News Hour interview, dated 9/29/2026, put the central question plainly: who decides whether powerful software is safe—the companies selling it, or an independent public process? Marcus, professor emeritus at New York University, author of “Taming Silicon Valley” and writer of the Substack Marcus on AI, argued for outside regulation and independent evaluation. Meanwhile, the executives and the president discussed a technology Trump said the United States couldn’t afford to lose ground on.
Who Gets to Set the Rules
Bennett said Trump appeared focused on industry self-regulation at the White House meeting. Asked if that was enough, Marcus replied, “It's certainly not enough. And trusting these companies is insane.” He said The New York Times had reported that OpenAI knew about problems that led to the Hugging Face incident months before it happened, but did nothing because it wanted to rush products to market. That, Marcus argued, is why safety decisions should move beyond the companies releasing those products.
He compared the oversight he wants to processes for aviation, food and drugs. Makers, he said, should have to show that a product's benefits outweigh its risks before release. He also called for audits after release and full disclosure of incidents. The interview describes no public, independent system currently governing that process.
Marcus said the White House was conducting some evaluation without independent oversight, under a secret framework it hadn’t revealed to the public. He said the framework precluded and delayed Anthropic’s Mythos system but didn’t delay OpenAI’s latest system, which he thought had more problems in some ways. He called the process idiosyncratic and opaque, leaving customers unsure what would be released and unable to trust U.S. AI. His alternative: clear public criteria, like those used for drugs, rather than decisions made behind a curtain.
The Costs of a Race
Trump described AI as technology “that will be bigger than the Industrial Revolution” and said, “We're leading by a lot.” He warned against sharing secrets with China and said, “there is nobody in third place.” The competition frame puts speed and national advantage first, while safety advocates ask whether protections can keep pace.
Marcus called the race against China “a little bit overblown.” He said China has regulation and a licensing regime for software products, and argued competition needn’t mean abandoning safeguards: “We don't make terrible cars just because we want to beat China on cars.” He predicted both countries would have their own “sovereign AI,” comparing the outcome to Coke and Pepsi, and said neither would wipe out the other.
The risks he described aren’t limited to a far-off extinction scenario. Marcus said extinction was unlikely, but catastrophe might be likely. He cited an incident in which an AI reported something incorrectly and nearly led to war, warning that errors and hallucinations could trigger further incidents. He also raised the possibility that cyber hacking could temporarily take down an electric grid and then a hospital. Referring to the WannaCry virus in the U.K. “a few years ago,” he said vulnerabilities could cause serious damage.
Bennett’s interview included no account of grassroots organizing, mutual aid or direct action, and named no nonprofit or NGO response. The proposed remedy was federal oversight: Marcus has called for a dedicated federal AI agency and said independent evaluation is the top need. Asked whether rules would arrive before a major crisis, he called it “the, you know, what is it, $3 trillion question.” He said the Hugging Face hack, along with tens of thousands of similar incidents, was a wake-up call; most incidents wouldn’t be harmful, he said, but one could be. The government, he warned, was racing the clock.