By Watson Scott Swail, President/CEO & Senior Research Scholar, Educational Policy Institute

AI deserves serious safeguards—but fear, corporate self-interest, and geopolitical competition make a simple slowdown far harder than it sounds.

There is no shortage of anxiety about artificial intelligence. Industry leaders, former employees, lawmakers, and researchers are warning that AI is advancing faster than our ability to understand or govern it. Some are calling for stronger safeguards. Others are asking whether development itself should slow down.

The debate intensified this week when Dario Amodei, CEO of Anthropic—the company behind Claude—urged the industry to “slow the pace” of frontier AI development. His call followed warnings from former Anthropic researcher Jacob Coxon, who said that Anthropic and OpenAI were “racing straight to self-improving superintelligence and gambling with our lives.” OpenAI CEO Sam Altman, Google DeepMind chief Demis Hassabis, and Elon Musk quickly endorsed the idea of slowing the pace or strengthening outside evaluation.

Why the warnings matter

The concern centers on artificial superintelligence, or ASI: systems that could outperform humans across many cognitive tasks and operate with increasing autonomy. The danger is not necessarily a humanoid robot. It is software with access to critical systems, financial networks, laboratories, communications infrastructure, or cyber tools—and the ability to act faster than people can respond.

That deserves a reason for pause. AI systems are already being used for phishing, impersonation, deepfakes, surveillance, and software attacks. More capable models could magnify those threats. The hard question is not whether AI creates risk. It clearly does. The question is whether the people building it can manage those risks while competing to release the next, more powerful model.

Scale makes that urgency difficult to dismiss. ChatGPT and Gemini now reach enormous global audiences, although published figures often measure different things—weekly users for one platform, monthly users for another. Business adoption is also widespread. Whatever the precise comparison, generative AI is no longer a niche technology. It is becoming part of the infrastructure of work, search, education, and everyday communication.

Can the industry regulate itself?

This is where my skepticism begins. Companies rarely regulate themselves to the point of sacrificing their strategic advantage. They manage risk, certainly, but usually in ways that protect the company and its shareholders. That does not make Amodei’s proposal insincere. It does mean we should ask whether voluntary safeguards are enough—and whether calls to slow down could also reinforce the position of firms that are already ahead.

My cynical first thought was one-upmanship: if a slowdown became law or an international norm, would it constrain Anthropic as much as it constrained smaller or less advanced competitors? Safety and competitive advantage are not mutually exclusive motives. Both can be true at the same time.

Government regulation is necessary, but it will not be quick. Congress does not move at the speed of frontier-model development, and international agreements move even more slowly. Any serious framework would need participation from the United States, China, India, Russia, Europe, and other major technology centers. A rule observed by only some competitors could reward those operating outside it.

That is the uncomfortable logic of an AI arms race: everyone may benefit from restraint, yet no one wants to be the only participant who restrains itself. Meaningful oversight will therefore require more than promises. It will need independent testing, incident reporting, enforceable safety thresholds, and consequences for companies that evade them.

Too much hype?

It is still difficult to separate credible risk from apocalyptic theater. When I heard a former Anthropic researcher warn that AI could cause human extinction, my first reaction was, “Here we go again.” We have a long history of greeting unfamiliar technologies with predictions of disaster. Sometimes those fears are exaggerated. Sometimes they identify a real danger before the rest of us are ready to see it.

That history cuts both ways. Repeated doom-and-gloom warnings create a natural aversion to alarmism, but dismissing every warning can be just as reckless as accepting every prediction. AI is not a microwave oven or a cell tower. It is a general-purpose technology that can write software, imitate people, process enormous stores of information, and increasingly act through other systems. The comparison should make us cautious about panic—not complacent about risk.

The benefits are real, too

Any honest discussion must also account for what AI can do well. It is accelerating research, helping analyze medical images, supporting drug discovery, improving clinical-trial data review, and giving researchers new ways to model proteins, vaccines, and treatment options. During COVID-19, AI and machine-learning tools helped with tasks ranging from vaccine-candidate analysis to clinical-data processing. They did not create decades of mRNA science overnight, but they did help researchers work faster.

Remember the diagnostic tools, like the Tricorder, that Dr. “Bones” McCoy used on Star Trek? We are closer to that future with AI than without it. No physician can read every new journal article or evaluate every possible treatment pathway unaided. AI can help synthesize evidence and surface patterns, provided that clinicians remain responsible for judgment and that systems are tested for accuracy, bias, privacy, and safety.

A better question

AI poses real risks, and sensible guardrails are necessary. But regulation will be slow, international cooperation will be uneven, and the companies calling for restraint may have strategic interests of their own. The choice is not between blind optimism and panic. It is between governing the technology deliberately and allowing competitive pressure to govern it for us.

We should demand independent evaluation, transparency about failures, security standards for high-risk systems, and clear accountability when AI causes harm. At the same time, we should protect the medical, scientific, educational, and economic uses that make the technology so valuable. The goal should not be to stop progress. It should be to make progress survivable, trustworthy, and broadly beneficial.

The challenge is not choosing between blind optimism and panic. It is building safeguards without surrendering the benefits that make AI so important.

Postscript: Full disclosure: I used AI to help research this article, proofread it, and test whether the argument flowed logically. I could have hired an editor, but that would have taken more time and money than I was willing to spend. Copilot was fast, effective, and essentially free. So I have to ask: am I part of the problem, too?