A Summit Built on Self-Policing

What’s Washington’s new plan for controlling AI? Let the people who built it control it.

On Tuesday, September 29, 2006, the White House hosted a high-profile luncheon with several big names in technology. With a renewed commitment from President Trump and top AI executives to “self-police” their companies’ AI models, Trump continued to argue that balancing safety concerns with his desire to accelerate a technology he now refers to as “Super Intelligence” requires a light-touch approach.

The agreement is voluntary and not legally enforceable, although Trump calls it “morally binding.”

Evidently, the rapidly evolving technology industry needs a pinky promise.

In the agreement Trump signed with CEOs of companies such as OpenAI, Anthropic, Google, Meta, Nvidia, and xAI, it calls for internal controls, outside audits, and board reviews. Additionally, companies should take steps to prevent AI systems from hacking or gaining unintended access to technical systems.

The timing is notable.

As a result of AI agents from several companies hacking into other companies’ systems, the AI industry is already under scrutiny. As part of Trump’s executive order, the government was instructed to replace “artificial intelligence” with “super intelligence,” or SI.

In many ways, the name change is symbolic. In fact, it aligns with the administration’s broader strategy: enthusiasm first, guardrails later.

And that’s where the real fight begins.

This isn’t about whether America should lead in AI. In fact, doing so makes a lot of economic and national-security sense. Ultimately, the United States must decide if it can win the AI race without realizing that it forgot to put the brakes on.

The Administration’s Case: Win the Race

The logic behind the light-touch approach is straightforward, and it deserves a fair hearing.

According to its defenders, the United States is engaged in a strategic conflict with China, and heavy regulation might give Beijing an advantage. President Trump has repeatedly framed AI development as a race in which the United States must maintain its lead. Additionally, Speaker Johnson has called for a meeting between industry leaders and government officials to determine Washington’s role.

As part of the administration’s broader AI strategy, that argument runs throughout.

To control the next generation of technology, the United States must build the infrastructure, companies, and models. Slowing down development, it is argued, could be a strategic liability.

According to this view, safety isn’t necessarily neglected. National security, economic growth, and speed are taking precedence over it.

But there is an economic argument, too.

Trump has called AI “the golden goose,” and he has renewed his support for the data centers that power it. In the event AI drives the next wave of productivity growth, slowing it down has real costs for jobs, investment, and global influence.

That’s what makes this debate uncomfortable: Slowing down has legitimate costs.

The Industry Pushback

There’s something unusual about who’s ringing the alarm at this moment. Typically, the technology industry opposes regulation. However, some of its most prominent leaders want more of it.

While AI leaders like Anthropic’s Dario Amodei and OpenAI’s Sam Altman push for government guardrails, Trump continues to push for a light-touch approach. CNN reported that several AI industry officials, including some CEOs building increasingly sophisticated models, warned of catastrophic results if they lose control of their technology.

The president’s response was dismissive. In a string of posts on September 14, CNN reported, he asserted that a “SICK conspiracy” was afoot to undermine AI and claimed his oversight alone was enough to prevent models from going rogue. During a public call with Nvidia CEO Jensen Huang, Trump also dismissed fears that AI could become uncontrollable, saying robots would not take over.

Basically, the administration claims that America already has laws against fraud, hacking, and other forms of misconduct and does not need to create an entirely new regulatory structure for a technology that is constantly evolving.

There are genuinely new questions that AI raises, which is the problem.

What happens when an autonomous agent takes an action that nobody specifically authorizes? When a model behaves unpredictably, who is responsible? When technology changes faster than the rules governing it, how is it regulated?

Existing laws can address some of those problems. But they may not address all of them.

Where Public Opinion Sits

As Trump’s confidence suggests, the politics here are less comfortable for him.

According to a Reuters/Ipsos poll conducted September 17-20, 73% of respondents worry that AI companies are not doing enough to prevent serious harm to society, while 55% want AI development to be slowed down. As communities deal with local impacts like electricity demand, the public’s resistance to data centers is also creating political headaches. (aljazeera.com)

That creates an awkward situation for Washington.

By asking Americans to trust the companies building the most powerful technology in history to regulate themselves, the administration is essentially encouraging these companies to build bigger models, more data centers, and more autonomous systems. And plenty of Americans aren’t responding with, “Sounds great. What could possibly go wrong?”

Influence is also a concern.

Many AI executives and investors are now major political donors, including to Trump-aligned groups. However, that does not mean policy is determined by industry donations. But it does provide an obvious reason for critics to question whether Washington and Silicon Valley are too cozy.

Some supporters may counter that pro-business administrations support is not uncommon, and voluntary agreements are more flexible and faster than legislation.

In the end, both things can be true.

Is the Tradeoff Real?

In terms of politics, it’s convenient to frame the issue in terms of “safety versus leadership.” But it may also be a false choice.

Here are a few things to consider.

Safety can be a competitive advantage.

There are extensive safety standards in aviation, pharmaceuticals, and nuclear power. Those rules aren’t proof that regulation works all the time. Yet they demonstrate that industries are capable of building enormous economic value within risk management systems.

Eventually, AI will have to earn the same level of trust from the public.

Therefore, unpredictability or hacking other companies aren’t necessarily competitive advantages. In some cases, it can be a liability.

Voluntary commitments have a mixed record.

Self-policing works best when incentives are aligned with caution. The strategy works less well when competitors have the ability to gain an advantage by taking greater risks. This is the fundamental problem with voluntarily slowing down an entire industry.

Safety might be worth a billion dollars more to one company, while building another data center is the better use of a billion dollars for its competitor. In the short term, the competitor wins.

By getting several major companies to make similar commitments, Tuesday’s agreement aims to solve that problem. However, we have yet to see if those commitments change behavior.

China is a real factor, but it cuts both ways.

The president has repeatedly stated that slowing American artificial intelligence development would be beneficial to China. It’s a legitimate strategic concern.

However, the geopolitical race does not resolve the safety issue. In fact, it complicates things.

In general, AI systems are not very good at recognizing national borders. Cyberattacks don’t stop at customs. Neither do autonomous systems.

Moreover, Chinese and American researchers have discussed the need for safeguards for increasingly autonomous AI systems, especially in areas involving critical infrastructure or military applications.

The counterargument has force, too.

When regulations are drafted in haste, they may lock in incumbents, stifle open-source development, and focus on problems from yesterday.

In terms of AI oversight, nobody has quite figured it out yet. There’s also a risk of rigid rules aging badly.

Flexible, iterative approaches are, however, defensible on those grounds alone. The problem is that “flexible” and “do nothing” are not synonymous.

What to Watch

If Tuesday’s accord has substance or is just a show, three things will reveal it.

First, look at what companies actually do. As part of the agreement, internal controls are in place, external auditing is performed, and board-level oversight is in place. In other words, those are more effective than simply stating “we promise to be responsible.” The question is whether those systems are transparent, independent, and able to stop dangerous behavior before it becomes a public issue.

Second, watch Congress. Legislators’ repeated discussions about AI guardrails and Johnson’s push for a meeting of the minds leave the door open to legislation. Several measures in the White House agreement may eventually be codified into law, according to the White House agreement.

Third, whether the next high-profile AI incident will alter the political calculus sooner than the White House expects. Oftentimes, voluntary safety agreements seem perfectly reasonable when everything is going smoothly. But if a self-aware AI agent does something nobody expects, it looks considerably less reassuring.

As the administration bets the public will accept risk in exchange for prosperity and strategic advantage, the public appears increasingly uncertain about the bargain. AI leadership is important. But leadership that is built on technology that the public distrusts and that even its creators don’t fully control is fragile. Although the summit was full of ceremony, it didn’t address whether Washington could pursue both objectives at the same time.