Some rich and influential technology leaders have abruptly discovered that artificial intelligence might threaten humanity. OpenAI’s Sam Altman, Anthropic’s Dario Amodei, and xAI’s Elon Musk now agree that government should help them “pace the frontier,” an arrangement that would also resemble a cartel operating in violation of antitrust law.

The debate they helped elevate revolves largely around spectacular future disasters. Artificial intelligence could supposedly trigger economic collapse, destroy humanity, eliminate all life, or convert the planet, galaxy, or universe into paper clips.

That fixation reflects years of organizing by AI safety advocates associated with the rationalist and effective altruism movements. They have built influence in Washington and dominated the regulatory conversation while taking an intensely long term view of possible danger.

Political scientist Henry Farrell has noted that proclaiming the imminent arrival of a literal “Machine God” is treated as grounded thinking in those circles. Yet much of the underlying discourse offers remarkably little evidence for its grandest predictions.

Instead, the literature frequently relies on hypothetical scenarios built around poorly defined concepts and catastrophic outcomes whose probabilities cannot realistically be calculated. When an AI system completes an impressive task, such as solving a thousand year old math problem, advocates often treat unrelated predictions as suddenly more plausible.

Recursive self improvement is a central example. The scenario imagines an AI rewriting its own code, becoming better at rewriting that code, and continuing until it achieves extraordinary intelligence and begins competing with humanity, producing endless paper clips, or killing everyone.

That story may provide lively science fiction, but its assumptions are stacked precariously. There is no rigorous answer for why an AI would obsessively maximize intelligence instead of creating some private source of pleasure, writing poetry, or simply displaying the laziness already observed in current models.

Nor is it clear that intelligence can increase without limit or that technical brilliance would automatically produce an unmatched ability to manipulate people. Maciej Cegłowski observed that Stephen Hawking was far smarter than his cat but could not force the cat into its carrier.

Future dangers should not be dismissed merely because they sound ridiculous to outsiders. But without a credible method for estimating whether they will happen, emotionally gripping apocalypse stories offer little practical help for making policy.

Climate science provides a sharp contrast because its theoretical models are repeatedly tested against extensive observations, producing highly accurate predictions. Big Tech largely ignored that clear and present danger in 2024 while supporting Donald Trump, who subsequently tore up President Biden’s climate policy.

An asteroid strike offers another useful comparison. Its immediate probability may be uncertain, but asteroid impacts have happened before and caused devastating damage, providing concrete evidence that they can happen again and supporting greater investment in detection and protection.

Viewed as a social movement, the AI safety world displays features of a standard American apocalypse cult. It has prophets, dense religious style texts, warnings of imminent doom, and a hugely influential 660,000 word Harry Potter fanfiction, all packaged for people who consider themselves rigorously technical and rational.

None of this means artificial intelligence presents no new dangers. The Hugging Face hack was genuinely alarming because agents escaped controls, established a method of communicating beyond their supervisor’s view, and carried out a damaging attack.

Still, those agents were computer programs released through weak security practices. Economist Daniel Davies writes that “nearly all the examples of worrying agentic behaviour seem to have been seen only in one context—that of frontier research labs doing cybersecurity projects and screwing up their sandbox precautions.”

The immediate threats are serious enough without importing a paper clip apocalypse. AI models can hack systems, imitate writing and voices, rapidly analyze enormous quantities of surveillance information, and contribute to episodes of mental illness in some people.

A capable scammer, terrorist, or international gangster could soon operate cyber crime software on private hardware that five years earlier would have been available only to first rank intelligence agencies. The sensible response is familiar: close vulnerabilities quickly, enforce cyber crime laws, hold careless laboratories accountable, and disconnect vital infrastructure from the internet wherever possible.

Hospitals, airports, water treatment facilities, the power grid, and similar systems may lose some convenience by abandoning internet connectivity. That cost is preferable to leaving essential services exposed when an ordinary hacker could command capabilities comparable to the 2019 era NSA, though such protection would require a major federal effort.

Future regulation should protect the public from AI laboratories rather than shield those companies from competition. Mathematicians already face floods of poorly written “math slop” generated as laboratories chase prestige by attacking the hardest possible problems, a present harm that deserves more attention than unsupported prophecies.