AI CEOs Were Right: Teach Big AI What’s Not Okay

AI CEOs Were Right: Teach Big AI What’s Not Okay

A dentist held up a tray and let me stare at the tools. Years later, I sat through a congressional hearing where Sam Altman did something similar: he named risks and asked for help. Both moments left you with the odd certainty that someone else had already decided we needed this.

I write to you as someone who pays attention to how power asks for permission. You felt the CEOs’ warnings as a public alarm and, like me, you wondered whether they were sincere or simply covering their bases. Either way, they were right in one narrow sense: they wanted to be told, clearly, what is not okay.

In a hearing room in 2023, Sam Altman told Congress he feared AI could go “quite wrong.”

That admission was not a plea for favors; it was a political gesture aimed at shaping the terms of control. When OpenAI and other frontier labs put risks on the table, they were serving notice: tell us the lines or we will be blamed when they’re crossed. You should read that as both a strategic demand and a reputational lifeline.

Big AI’s public posture—warnings framed as civic responsibility—bought them political capital. It bought them time, too. CEOs like Altman and Dario Amodei warned of an era that would be painful and chaotic, and their language read like a request: hold us accountable, but let us lead the fix. You could call that smart self-preservation.

What did the Trump administration do to Anthropic?

Short answer: it stepped in and hit pause.

After Anthropic released Claude Fable 5, the administration flagged the company as a supply-chain risk and issued export controls that effectively froze new releases for a period. That action wasn’t framed as consumer protection so much as national control—an effort to keep sensitive capabilities from leaking to rivals or adversaries.

On an executive blog and a company memo, the AI labs showed their hands.

OpenAI’s blog about GPT-5.6 reads like someone trying to square public commitments with private plans. They distributed access to VIP customers and told readers they were working with the White House on a repeatable release framework. Anthropic, meanwhile, ran a survey that showed public distrust but still argued for its own vetting process.

There’s a transactional logic here. The companies want rules because rules can freeze competitors in place and let established players shape the enforcement. You should remember that when you see language about “working with government” — it often means protecting market position while appearing cooperative.

Can the government regulate AI models?

Yes, but the path chosen today is not Congress writing new laws. Instead, the administration has used executive authority—export controls, supply-chain designations, and an Executive Order asking firms to submit models for vetting. That creates a regulatory system that depends on administration discretion rather than statute.

In the Oval Office and across federal agencies, the posture shifted from laissez-faire to targeted policing.

The tilt isn’t ideological so much as transactional. Early Trump Administration signals—especially Vice President Vance’s Paris speech—favored minimal regulation so the domestic industry could sprint. Then came the realization that sprinting without agreed guardrails risks national security and political embarrassment.

So the administration did something odd: it demanded oversight when it served its interests and left the broader legal architecture untouched. The result is a stop-and-start system that punishes some firms and nudges others toward private gating and government-friendly release strategies.

Will U.S. regulation slow AI innovation?

Short-term friction is likely. Companies already warn of a near moratorium on releases, and that can harm earnings and R&D plans. But the deeper risk is geopolitical: while U.S. labs temper releases, foreign competitors—especially in China—may press forward. That creates an uneven global playing field where American control could mean loss of advantage.

At town halls, in polls, and in private surveys, people say they don’t trust the companies.

Only 15% of Americans trust AI firms to decide how the technology is developed. Seven in ten oppose data centers in their neighborhoods. And 87% think foreign governments will use AI against the U.S. within two decades. Those numbers explain why the White House felt it needed to act—public fear is political oxygen.

You and I are not being asked whether we approve of the industry’s promises. We’re living the consequences of those promises. The CEOs’ public insistence that their tech could “go quite wrong” bought them attention, but it also shifted blame away from elected leaders unless those leaders could show they were stepping in.

For the companies, the calculus is simple: prefer a visible, repeatable vetting process to the chaos of surprise enforcement. For the administration, the calculus is political and strategic: keep power to decide which tools travel abroad and which remain domestic advantages.

There’s another layer: the labs wanted clear rules so they could sell compliance as a service—access to models that are “approved.” That’s a profitable business model. You should watch for it.

And there’s one image that won’t leave me: the CEOs holding up their tray and waiting for a thumbs-up. It’s a performance of responsibility. The second image is equally sharp: a referee who only shows up to cheer his favorite team and then writes the rulebook midgame. Those two scenes explain the strain between private power and public authority.

So where does that leave you? With a government that paused one of the loudest tools and a tech industry that asked—loudly—to be told what is not okay. But the mechanism for that judgment rides on presidential preference and administrative discretion, not a law passed and debated by Congress.

I’ll say this plainly: I’d rather see a statute than a seasonal policy. Rules made in public hearings, passed by legislators, and contested in courts are harder to weaponize as geopolitical advantage. You should want durable guardrails, not spot checks that change with the next tweet.

Big AI had a point when it asked to be told what is unacceptable. It wanted clarity to avoid blame and to sell compliant access. You want clarity so your rights and safety are protected, not parceled out by political appetite. Who ends up writing the rules—companies, an administration, or Congress—will decide which of those futures we get?