I was struck this week by calls to slow AI coming from four AI leaders who are fiercely competing to outpace each other. Dario Amodei, CEO of Anthropic, called for “pacing the frontier.” Sam Altman, CEO of OpenAI, basically agreed. As did Elon Musk. And Demis Hassabis, who leads Google DeepMind, also wants greater coordination. When they say publicly that safety may not be keeping pace with capability, we should collectively listen. (Associated Press, September 16, 2026, “AI rivals found rare agreement on safety. Putting it into practice is harder.”)
My fear is that this is getting drowned out by two camps. The first is the sensationalist camp claiming that there is a 10 percent chance that AI will wipe out humanity. This claim is extraordinary. And extraordinary claims require extraordinary evidence. Attaching numbers and probabilities to a long chain of speculative assumptions creates an illusion of science but is not in fact evidence. It reminds me of the worst version of effective altruism, particularly the expected-value reasoning espoused by Sam Bankman-Fried. (The New Yorker, December 1, 2022, “Sam Bankman-Fried, Effective Altruism, and the Question of Complicity.”)
The second camp is touted by President Trump who, this week, rejected calls to slow development saying “whoever wins AI wins” citing the need to compete with China. His administration’s March legislative framework similarly emphasizes a national approach designed to allow American companies to innovate with limited regulatory burden. Sources: Associated Press, September 13, 2026, “Trump downplays the need to check AI development and says he doesn’t want to cede edge to China”; The White House, March 20, 2026, “President Donald J. Trump Unveils National AI Legislative Framework.”
While I reject the apocalypse narrative, the new risks created by AI are already here. But they aren’t affecting everyone equally. And when individuals get hurt, its easy for we Americans to turn a blind eye.
AI Fraud may be the clearest example.
Americans reported $15.9 billion in fraud losses to the Federal Trade Commission in 2025, up from a little over $12 billion in 2024 (+27% in one year). Imposter scams alone cost $3.5 billion, nearly three times their 2020 level. The Federal Bureau of Investigation’s 2025 Internet Crime Report found nearly $21 billion in reported cyber-enabled crime losses. For the first time, it separated complaints involving AI: 22,364 complaints and $893 million in reported losses. Which is almost surely a substantial undercount. The FBI describes criminals using AI to create fake social profiles, cloned voices, synthetic identification documents, personalized conversations and convincing fake video. Businesses reported more than $30 million in losses from business email compromise scams involving AI alone. (Sources: Federal Trade Commission, June 15, 2026, “FTC Data Show People Reported Losing $3.5 Billion to Imposter Scams in 2025”; Federal Bureau of Investigation, April 6, 2026, “Cryptocurrency and AI Scams Bilk Americans of Billions”; Federal Bureau of Investigation, 2025 Internet Crime Report.)
Artificial intelligence makes fraud cheaper, more convincing and much easier to do at scale. An immediate, known risk is malicious people using increasingly powerful foundation models to create highly sophisticated scams. I recently received an email from a Private Equity partner specializing in education asking for company due diligence. They cited our firm’s reputation as the basis for the outreach. I was gratified for this unknown executive to have high regard for our work. But the email provided a link to a dataroom (first red flag - what PE firm gives access to a data room without vetting, an executed NDA and joinder?!?). I looked up the company online. Its website was professional and listed investments in leading edtech companies. LinkedIn had professional profiles. Oddly, we had no contacts in common (another red flag in an intimate industry). With some checking, we realized that the firm was not listed on any of the claimed transactions. I like to think that my bullshit meter is dialed to 11 but it took me real time to figure out that everything was fabricated. Given how intricate this scam was, what risks do my parents’ generation face? Or my kids? Or people that don’t use AI? Or pretty much anyone who acts in good faith and has trust in humanity. How on earth will we sort out what is real when AI can expertly imitate loved ones, colleagues, and even movie stars. AI is already making fraud and impersonation more sophisticated, more scalable and harder to detect.
Shutting down AI is not a realistic solution. The tools are already enormously useful, widely distributed and increasingly embedded in how people work. Any serious attempt to stop AI in its tracks would cripple the economy. Nor is a pause a substitute for regulation. If frontier companies decide to slow down, the questions around cybersecurity, independent evaluation, incident reporting and accountability remain. If they don’t slow down, those same questions remain.
Instead, we need AI regulation now. And we should assume we will have to keep changing it to address new developments. AI is changing far too quickly to imagine Congress can pass one law and be done. But we need the fundamentals in place now: minimum cybersecurity requirements for frontier laboratories, independent third-party testing, mandatory disclosure of serious incidents, stronger security monitoring for high-risk deployments, rules governing when autonomous systems can act without human authorization, and emergency shutdown and override requirements.
The OpenAI and Hugging Face incident is a good illustration of why. The AI agents demonstrated genuinely concerning offensive capabilities, but OpenAI’s own postmortem identified more familiar problems involving isolation, monitoring, credentials and incident escalation. Internal teams saw warning signs before the breach, including unauthorized internet access and agent-to-agent communication, but those signals were not escalated aggressively enough. OpenAI says it has tightened its controls.
The lesson is not that OpenAI “lost control” of artificial intelligence in some science-fiction sense. It is that highly capable systems can exploit ordinary weaknesses in cybersecurity and governance. OpenAI has built extraordinary foundation models. That does not mean it has solved cybersecurity. Cutting-edge AI makes basic security controls more important, not less.
AI does not need to have a 10 percent chance of wiping out humanity to justify serious regulation. The risks are already here and they are enough.
It is past time to act.