Artificial Intelligence poses an existential risk to humanity. Or it will destroy all jobs. Or it will go rogue and escape control. These are headlines in reputable news outlets, but they are wrong. The mechanism that creates these concerns is a media that amplifies the fears and hostility of a class of public intellectuals and then disseminates them to an already anxious public. In some cases, AI anxiety serves the unstated financial interests of particular individuals or groups — Russia and China, for example, may try to expand these narratives — but there is usually an element of truth behind the fears, which means they cannot be immediately dismissed.
For AI, risks remain hypothetical. No one has connected the dots that show how you get from superintelligence or Artificial General Intelligence (AGI) to an existential threat. The absence of practical detail makes it likely that an existential AI threat is both hypothetical and improbable.
AI is the fifth wave of media-induced moral panic since 2000. These narratives share certain characteristics: a real problem; a veneer of expertise; a distrust of government, companies, and markets to act fairly or provide solutions; and repetition by conventional and social media. The more extreme claims eventually produce disbelief and opposition, but this can take months or even years. In the interim, ill-considered policy initiatives are put in place and linger even after the public has moved on. Retrospection does not appear to be a feature of the American political process. Constant repetition, dismissal, and rejection (often indignant) of any criticism or questioning, and an embedding in popular culture, reinforces the moral panic.
The Internet has removed curation and barriers to publication, providing new tools to amplify assertions. Truth and mis-truth are not easily distinguished online without additional fact-checking efforts. But the internet effect is not the whole story. There have been previous episodes of moral panic (i.e., the Red Scare or Population Bomb) that predate it.
This critique of AI panic is predicated on the assertion that AI fears are exaggerated beyond the point of credibility. In the last 70 years, there have been numerous predictions of extreme outcomes, from over-population, the peak oil theory, the disappearance of all beaches on the East Coast by 2020, or global famine by the year 2000. All have proven to be wrong. Yet, something drives the culture to predict catastrophe over and over again.
The prospect of AI catastrophe requires a discussion of both probability and process. Probability is the likelihood that a prediction is accurate. One way to estimate this is by the track record of previous extreme predictions, of which none have come to fruition. Of course, the miniscule possibility of a “black swan” or “random walk” event means we cannot depend entirely on precedent to predict probability. But in the past, when carefully constructed, precedent has been generally accurate.
Process is what is necessary to turn a hypothetical risk into a real one. It is easy to imagine outcomes, but if you cannot explain the steps needed to get to them, they remain imaginary. Process introduces an element of operational complexity that reduces the chance of catastrophe. For AI, we would need to explain how it could achieve a level of lethality that would allow it to inflict mass casualties. Mass casualties require either nuclear weapons or weapons of mass destruction and at the moment, neither are plausible scenarios.
Another precedent might be 1980s fears of a “nuclear winter,” a detonation of thousands of nuclear weapons that would throw enough soil in the air to block out sunlight, leading to freezing temperatures and the end of human life. Dramatic, but dependent on a chain of unlikely events. AI may prove similarly dramatic but unlikely. Nuclear weapons have demonstrated capability and their acquisition shows clear intent. Nuclear risk is not hypothetical, but nuclear winter was imaginary, the extension of a legitimate fear into the realm of the improbable, and it may be that unconscious fears of nuclear war underpin much of the fear of AI.
Nuclear weapons could not produce extinction. Can something else? Extinction is not that easy. The COVID-19 pandemic killed about 0.3% of the world‘s population. Nuclear weapons used en masse in the 1980s could have killed about a billion people, or about 20% of the global population. Even the recursive self-improvement of AI tools is unlikely to cause extinction or mass casualties. To do this, AI would need to be connected to a weapon of mass destruction. Futhermore, traditional extinction stories fail to take into account the ability of the human race to adjust and adapt.
This does not mean there is zero potential for AI to create security and safety risks. Frontier AI seems particularly prone to error because of inadequate coding processes, the way that a Model-T automobile or early biplane was unintentionally unsafe compared to current cars or aircraft, given primitive production and a lack of standards.
We cannot rule out system-wide failures in the electrical grid or other critical infrastructure from badly written AI products. These could possibly be lethal, but without mass effect. Software is not lethal by itself. It has to connect to something, either to a weapon or to infrastructure. Even then, the ability to do harm remains limited, much less at a massive scale. This means that for now, human extinction from AI remains both improbable and hypothetical, but the Anxiety Machine will grind ahead regardless.
James Lewis is a Distinguished Fellow with the Tech Policy Program at the Center for European Policy Analysis.
Bandwidth is CEPA’s online journal dedicated to advancing transatlantic cooperation on tech policy. All opinions expressed on Bandwidth are those of the author alone and may not represent those of the institutions they represent or the Center for European Policy Analysis. CEPA maintains a strict intellectual independence policy across all its projects and publications.
Steel and Silicon: Allied Innovation for the Next Century
2026 CEPA Tech & Security Conference
