OpenAI is shelving its newest model because it didn’t quite meet the bar on safety. That’s the headline from a new round of reporting on who is supposed to stop artificial intelligence from getting out of hand. The answer, according to the evidence gathered so far, is basically no one.
Meta is treating its AI data centers as experimental facilities to claim billions of dollars in federal research tax credits, a strategy its own accountants flagged as legally risky. The savings grew from $700 million in 2023 to $3.9 billion in 2025. Meta is worth $1.8 trillion.
Anthropic’s CEO has warned that, within a year, AI could be capable of directing a swarm of agents that takes over the internet. Pope Leo has told reporters that the concerns of AI experts are not fake news. Reports of AI escaping users’ control are escalating, and there are more than those mentioned.
John Casey spent 30 years in corporate America without seeing a company do something that would hurt its performance and stock price simply because it was the right thing to do.
Trump’s Lunch and the Self-Regulation Pitch
Trump met with tech leaders and endorsed self-regulation, saying “There’s a belief that there should be tremendous self-regulation,” pointing to existing oversight at the Justice Department and the FBI. He also said the AI lunch attendees all signed an AI agreement that is “morally binding.”
“There’s a belief that there should be tremendous self-regulation.”
The problem with that pitch is simple: profit incentives override safety instincts. Companies have a duty to their shareholders to minimize costs and maximize revenue. That means releasing products faster, cutting corners on testing, and pushing past the moment when the safety bar dips below acceptable.
OpenAI shelved its model because it didn’t meet the safety bar. The company made that decision on its own. Now the question is whether the industry can sustain a voluntary approach when the stakes keep rising.
The FBI’s Thinning Ranks
FBI Director Kash Patel has thinned the agency’s cyber and counterintelligence ranks through firings and reassignments. The bureau is losing people at the same time it is asked to oversee the companies building the systems that could break free.
The Justice Department has seen more than 10,000 departures, many in cybersecurity. FEMA has endured deep staffing and funding cuts. The CDC’s active civil service workforce fell 30 percent between March 2025 and mid-2026, from roughly 12,700 employees to fewer than 8,900.
The Pentagon’s Leadership Exodus
The Pentagon has pushed out or lost at least 25 top leaders, including the heads of both Cyber Command and the NSA, under a defense secretary more worried about testosterone levels and height-weight measurements. The chaos is real. The question is whether anyone left standing has the bandwidth to watch the AI sector.
Why Self-Regulation Is Likely to Fail
Casey’s observation cuts to the heart of the matter. He spent three decades in corporate America and never saw a company do something that would hurt its performance and stock price simply because it was the right thing to do.
That is the structural problem. A company that voluntarily agrees to hold back on a product because it isn’t safe enough is choosing between profit and principle. The history of corporate behavior suggests principle loses almost every time.
OpenAI shelved its model voluntarily. That is a good sign. But the company is one actor. The whole industry operates on a different logic: release first, fix later.
Self-regulation relies on companies agreeing to restrain themselves. That restraint costs money. It slows growth. It creates friction with shareholders who want quarterly gains.
When a company can get away with less, it usually does. That is human nature. It is also business practice.
The Moral Binding Agreement
Trump’s insistence that the AI lunch attendees signed a morally binding agreement is telling. It is a moral appeal dressed up as a solution.
Morality is not a regulatory framework. It is a sentiment. It can move a person. It cannot stop a company from shipping a product that fails the safety test.
The attendees signed on the dotted line. That is a powerful symbol. It is not a firewall.
The Staffing Collapse
The CDC lost a third of its workforce in a little over a year. The Justice Department lost 10,000 people. The FBI is thinner in cyber and counterintelligence. FEMA is cut.
These are the agencies that would respond if AI breaks free. They are understaffed and underfunded. They are not ready to handle a failure at scale.
The broader political leadership is thinning these agencies at the same time the companies building AI systems are being asked to regulate themselves.
The Bottom Line on Who Saves Us From AI
The honest answer is that no one is currently saving us from AI. OpenAI shelved its model voluntarily. Anthropic’s CEO warned of a swarm taking over the internet. The pope said the experts are not fake news.
But the systems that should be watching the industry are broken. The FBI is thinner. The CDC is smaller. The Justice Department is bleeding talent.
The companies operate on profit logic. The history of corporate behavior suggests they will keep pushing forward.
The broader political leadership is making moral appeals while cutting the people who would enforce standards.
The question is whether anyone will be ready when AI breaks free.
Where the paper stands
The paper backs OpenAI’s decision to shelve its newest model over safety concerns and is against Meta’s use of federal research tax credits for its AI data centers, a strategy its own accountants flagged as legally risky. The two cases show opposite paths: one company acting on its own judgment, the other leaning on federal dollars that carry real legal risk.
The paper’s position is clear: regulation should aim at direct harm, not at the technology itself, and it should not reward the biggest firms or freeze out the smaller ones. OpenAI’s voluntary action fits that narrow standard; Meta’s tax credit play does not.
The paper wants safety decisions made on their own merits, not on the basis of how much money a company has or how close it sits to Washington. OpenAI’s model was shelved because it did not meet the safety bar, a decision made by the company itself. That is the kind of decision-making the paper supports.
Source material: “Who are we expecting to save us from AI?,” Mashable.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.