The people who build frontier artificial intelligence say it is getting away from them. The people who study it, deploy it and defend against it mostly agree. They just don’t think a slower frontier is the fix.
Anthropic CEO Dario Amodei on Sept. 12 called for the industry to pace capability gains, and OpenAI CEO Sam Altman and Elon Musk endorsed the plan within hours. PYMNTS CEO Karen Webster argued that the labs already have the money and knowledge to govern their own models, and that rules written in a hurry would favor the frontrunners. PYMNTS reached out to experts to weigh in on the alarm.
Maya Mikhailov, CEO and co-founder of Savvi AI, sees a gap between what the labs are worried about and what their customers actually use. Slowing frontier AI development is not the same as slowing AI adoption, she told PYMNTS, because most enterprise customers “aren’t even using the state-of-the-art models to begin with.” That gap shows up in the revenue too: Ramp’s AI Index found the top 1% of businesses drive the vast majority of model companies’ enterprise revenue. It also gives the labs a financial reason to ease off. Training these models costs billions of dollars, Mikhailov said, “while most of their enterprise clients aren’t even coming close to using the advanced models’ capacity.” Selling what they already have brings in more money, and that counts for something as the labs look toward initial public offerings (IPOs) and closer scrutiny from public markets. A slowdown carries a second benefit: it invites regulation that would hit open-weight and Chinese models, the cheaper rivals undercutting the labs’ prices.
Chris Selland, a lecturer in entrepreneurship and innovation, expects the pace to moderate on its own, for a simpler reason. “It only makes economic sense for the model providers to push the frontier at the pace at which the market can absorb it,” he told PYMNTS. He’s also skeptical of the alarm around recursive self-improvement, the point at which models begin improving and engineering themselves with no human in the loop, calling it “a significant amount of hyperbole.” Still, he said, if the labs genuinely believe slowing down is the right call, they have an obligation to do it.
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
Minyang Jiang, chief strategy officer at Credibly, questions the competitive logic behind the whole idea. Pulling back would just hand ground to whoever keeps pushing, she told PYMNTS. And if advancing AI destroys value faster than it creates it, she asked, “then who really wins in the end?”
Agents Raise the Stakes for AI Mistakes
The danger experts do see is the shift from chatbots to agents. Jiang pointed to the August Hugging Face incident, in which AI agents attacked systems they weren’t asked to and hid what they had done. The problem, she said, is that AI “does not have a mental model of what is acceptable or ethical, only what is technically feasible.”
Mayank Kejriwal, a research assistant professor at USC, sees that incident as part of a bigger shift. Artificial intelligence models can now control machines and run tasks that take days to finish, he told PYMNTS, which means “the risk surface is now operational rather than merely informational.” Once a system can trigger workflows on its own, permissions and human approval become the real safety work, not the model’s raw capability. And slowing a handful of leading firms, he added, “will not stop progress globally.”
Ben Colman, co-founder and CEO of Reality Defender, watches that shift play out in bank contact centers, roughly three to six months after each new capability ships. “A year ago, the threat was an artifact you examined after the fact. Now it is a live conversation,” he told PYMNTS. Autonomous agents already file a claim, forge the paperwork behind it, and pass voice verification on the follow-up call, without a human involved at any step. Offense spreads in weeks, he said, while defense moves at the speed of procurement. As he put it, “AI was a tool an attacker used and now AI is an attacker’s operating system.”
Selland thinks independent evaluators would help, but only within limits. That works “as long as they are knowledgeable and capable of providing effective oversight,” he said, and in his view that rules out government regulators. He points instead to self-regulation the industry sponsors and runs itself, something like FINRA in financial services.
Whatever the labs decide, Mikhailov doesn’t expect it to change enterprise behavior. “Enterprises won’t stop adopting AI just because they don’t have newer models to deploy,” she said.
Add as Preferred Source
