AI, the Fermi Paradox and the Problem of “You First”

My friend Daniel Newman is the AI bull against whom I calibrate my own bullishness. If you think I am optimistic about artificial intelligence, spend some time with Daniel. His conviction about what AI will make possible was part of my thought process in writing The Indispensability Trap.

So when Daniel weighed in on LinkedIn about slowing AI development, I paid attention.

His argument was straightforward. AI can create substantially greater security risks, but it can also produce substantially stronger defenses. The companies building these technologies have a responsibility to invest more, and faster, in making them safe. Technological revolutions have always created opportunities for people who want to improve the world and people who want to exploit it.

I broadly agree. I had already approached the subject from two directions. First came my Dr. Evil article on Techstrong.ai, exploring how AI might actually contribute to human extinction. Then came a more serious report, AI Extinction Risk: What Can We Do That Actually Matters?, examining practical ways to protect us from ourselves without putting technological progress on hold.

Yes, ourselves. We build these systems, decide where to deploy them and grant them authority. We own those decisions.

But the more I listen to the slowdown debate, the more I wonder whether we are arguing over a choice that exists mostly in the abstract.

When somebody says, “We should slow down,” who exactly is “we”? And what happens when part of “we” decides it would rather win?

Where Is Everybody?

That question sent me in a direction that might seem strange for an AI op-ed: the Fermi paradox.

The basic puzzle is familiar. The universe has had an enormous amount of time to produce life. There are vast numbers of stars and planets. Yet we have no convincing evidence of another technological civilization. Where is everybody?

There are many possible explanations. Life might be extraordinarily rare. Intelligence might be rarer still. Our searches might be inadequate, or our assumptions about what other civilizations would do might be wrong.

One possibility is the Great Filter. In Robin Hanson’s formulation, one or more steps between lifeless matter and a lasting, expanding technological civilization could be exceptionally difficult. Perhaps the hardest steps are behind us. Perhaps something ahead prevents civilizations from surviving long enough to become widely visible.

Technological self-destruction is one candidate, not an established answer. The absence of an alien radio broadcast does not tell us that somebody else built a chatbot and blew up their planet.

Still, the hypothesis raises an uncomfortable question. Could a civilization become smart enough to understand the danger it is creating without becoming capable of agreeing on what to do about it?

Look around. That does not require much imagination.

We Have Met This Problem Before

In The Indispensability Trap, I discuss the War of the Currents. Electricity provides an instructive example of technological promise, genuine danger and commercial competition becoming entangled.

Edison’s direct-current interests competed with the alternating-current systems associated with Tesla and Westinghouse. Arguments about safety were also arguments about whose technology would prevail. The Department of Energy’s account describes Edison’s campaign to discredit alternating current while the companies competed for major electrification contracts.

Electricity could kill people. That was real. But the commercial response was hardly a collective agreement to stop building electrical systems until everyone felt comfortable. Participants wanted their systems adopted, their investments rewarded and their competitors defeated.

That does not mean every safety objection was dishonest. It means safety debates took place inside a competitive market, not outside it.

The electricity analogy has limits. An electrical accident is not equivalent to a hypothetical civilization-ending AI failure. But the people making decisions faced incentives we would recognize in today’s boardrooms. Acknowledging danger did not eliminate the desire to control the technology that was becoming essential.

Nuclear technology raised the stakes considerably.

Scientists investigated whether a nuclear explosion could ignite the atmosphere. We should tell that story accurately. The Los Alamos analysis found that energy losses would prevent a self-sustaining reaction under the conditions it examined. This was not a group of scientists deciding to turn Earth into a short-lived star and hoping for the best.

The more revealing story for our purposes came afterward.

In 1946, the Baruch Plan proposed international control of atomic energy. It included oversight of facilities, inspections and penalties for violations. Once the arrangement was implemented, the United States would begin destroying its nuclear arsenal.

The Soviet Union objected to an arrangement that preserved the American nuclear monopoly during the transition, subjected Soviet facilities to inspection and removed its Security Council veto over relevant sanctions. The proposal failed. The State Department’s historical account lays out the dispute.

This was not simply a failure to understand that atomic weapons were dangerous. It was a disagreement about who would surrender leverage, under whose authority and in what order.

You first.

More than eight decades into the nuclear age, humanity has not destroyed itself. That is something to appreciate. It does not establish that we have permanently solved the problem, or that the precautions and agreements developed along the way were unnecessary.

It does show how difficult it can be to translate a shared danger into an arrangement that rivals will accept.

The Boardroom Has Its Own Version

We tend to discuss AI competition as a contest between countries. That matters, but it is only part of the story.

Put yourself in the position of a company developing AI. You believe the technology could transform your industry. Your employees have committed their careers to it. Customers are choosing platforms. Investors expect progress. Competitors are recruiting your researchers.

Now you are asked to slow down.

You might sincerely think a coordinated pause would be sensible. But will your competitor pause? Will its investors accept the same delay? Will its salespeople tell your customers that you are being responsible, or suggest that you have fallen behind?

A lead does not have to last forever to matter. It can help a company secure customers, establish distribution, attract talent and become embedded in other businesses. The fear is that a temporary concession becomes a lasting disadvantage.

Nor does this require a dishonest competitor. Two companies can both recognize a collective risk and both support restraint, provided the other participates. Each may also believe its own approach is safer and that letting a less responsible rival lead would make matters worse.

Everybody has a reason why their continued participation is necessary.

Anthropic has described aspects of this difficulty quite openly. In its February 2026 explanation of its revised Responsible Scaling Policy, the company discussed the gap between safeguards it could implement independently and measures requiring collective action. It pointed to uncertainty in capability assessments, limited government progress and the difficulty of meeting more demanding safeguards unilaterally.

Its revised framework separated its own plans from recommendations for the broader industry.

That is not evidence that safety is impossible or that Anthropic abandoned it. It is evidence that even a company explicitly organized around AI safety encounters limits to what it believes it can accomplish alone.

This is where the indispensability argument becomes especially relevant. The more essential participants believe AI will become, the more consequential they consider the prospect of somebody else controlling it.

A company fears dependence on a rival’s platform. A country fears dependence on another country’s technology. Each sees reasons to preserve its own capabilities. Those reasons can be individually understandable while making collective restraint increasingly difficult.

No Dr. Evil required.

Have We Really Never Slowed Down?

It would be convenient for my argument to claim humanity has never deliberately restrained a technology. It would also be wrong.

The recombinant-DNA moratorium and the 1975 Asilomar conference are important counterexamples. Scientists paused certain experiments and developed recommendations for proceeding with safeguards. The conference’s published statement is a reminder that scientific restraint is possible.

We should also be precise about what AI pause advocates propose. The 2023 open letter called for a six-month pause on training systems more powerful than GPT-4. It explicitly distinguished that from stopping AI development generally and called for a public, verifiable pause involving all key actors.

Those qualifications matter. We cannot dismiss every targeted restriction by pretending its proponents want to shut down every computer.

But specifying that a pause should include all key actors does not secure their participation. Calling it verifiable does not settle what must be observed, who gets access or how disagreements will be resolved.

What happens if consequential competitors remain outside the arrangement? How do participants distinguish permitted research from prohibited advancement? What prevents a temporary restriction from protecting incumbents while limiting challengers?

An agreement does not need perfect compliance to be useful. One holdout does not automatically make every restriction pointless. But proponents still have to explain why enough participants would accept the terms, how compliance would be assessed and why the arrangement would endure under competitive pressure.

That is a much harder assignment than persuading people to sign a statement expressing concern.

Intelligence Is Not the Same As Agreement

This is why I find the Fermi connection worth considering, even while remaining deeply skeptical of claims that AI extinction is around the corner.

We often imagine a civilization confronting a great technological danger as though it were a single decision-maker. It studies the problem, weighs the evidence and chooses a course.

Our civilization does not work that way.

It consists of governments, companies, institutions and individuals with different information, obligations and ambitions. Even when they agree about a danger, they may disagree about its likelihood, its timing and who should bear the cost of reducing it.

Perhaps other civilizations, if they exist, encounter something similar. That is speculation, not an explanation we can establish by looking at the night sky.

But here on Earth, we can examine our own behavior. We know that understanding a shared problem does not automatically produce a shared response. We also know that the fear of losing an advantage can persist even among people acting in good faith.

I have already laid out my views on practical protection in the report. I am not arguing that a universal slowdown’s difficulties relieve anyone of responsibility. A company cannot turn “our competitor would do it” into a blanket excuse for every decision it makes.

I am questioning the confidence with which people talk about slowing humanity down as though somebody has their hand on a common throttle.

Maybe the test facing a technological civilization is whether it can make cooperation sufficiently credible that its participants are willing to accept restraint. We have managed versions of that before. Whether we can do it across something as consequential and contested as AI is a question, not an assumption.

I do not know whether that has anything to do with the silence in the universe. I do know that here, in our own companies and countries, the appeal to slow down keeps encountering the same response.

Sure. You first.