Against the AI Omnicause

We must work directly to ban ASI, because nothing else will do.

Hello everyone, this article is a collaboration with Fernando Borretti. We decided to work together because we share ideas, and he’s a much more prolific writer than I am.

Funnily enough, we worked on this a few months ago and just didn’t get to publishing it. We think it only got more relevant as time passed.

If you look at the websites of AI safety organizations, you find many disparate causes: deepfakes, job loss, disinformation, and so on. Rarely you will find, buried deep down, as if they are ashamed to bring it up, a mention of the fact that top experts in AI think there is a significant chance that AI leads to human extinction, that even the CEOs of the companies building AI think there is a one-in-four chance that things go “really, really badly”.

Yet, when you talk to the people who founded—or fund—these organizations, in many cases, their main concern is human extinction. And yet they don’t talk about extinction risk. And on the face of it, this disconnect between private beliefs and public advocacy is hard to explain. It’s like you are on the Titanic, and you see the iceberg coming a mile away, and rather than point it out to the crew and urge them to turn the ship, you talk about everything but the iceberg: that third class passengers should have more equitable access to the gym, for instance.

This is a recurring failure mode among people who care about AI risk. You have people whose main concern is human extinction, but in their work, their public statements, their advocacy, there is no mention of extinction. Instead their work focuses on what we call the AI omnicause: the ever-growing list of every problem, great and small, associated with AI, everything under the sun except human extinction. Job loss, the environmental impact of datacenters, deepfakes, disinformation, technical risk mitigations like AI evaluations and pre-deployment testing, etc.

In this article, we argue that the AI omnicause is a losing strategy, both practically and morally. Practically, because it does not work: working on the one thousand and one mundane risks of AI does nothing to address the risk of human extinction from AI. Morally, because if you believe that the lives of eight billion people are at stake, you are obligated to tell people about it.

Why does the AI omnicause happen?

Why is the AI omnicause such a common failure mode? First: because it’s easier. When you have a portfolio of problems, and some are easy, and some are hard, people naturally gravitate towards the easy causes. And if solving the easy problems doesn’t get you closer to the goal, you can always rationalize that away.

Second, many people radically underestimate what others are able to hear. They decide—maybe from a few bad interactions, or sometimes without even trying—that politicians and the public don’t want to hear about human extinction, so they don’t talk about it. And when politicians don’t ask about extinction risk (because no-one told them) they use that silence to justify not talking about it. In doing this, they are giving up without even trying. The question cannot be “what are people talking about now?” or “what issues are relevant now?”, since part of our task is to make this issue as salient as possible.

Third, many people believe that talking about human extinction is too confronting, or too science-fictional. Unfortunately, reality has no such inhibitions, and is rapidly racing ahead of them: just recently OpenAI had one of its AIs hack its way out of containment, and hack into another company’s systems, by exploiting never-before-seen security vulnerabilities, and Anthropic’s AI solved a math problem that stood open for over eighty years. This is the world of July 2026, and our messaging should stay ahead of it.

Finally: many people routinely outsmart themselves when it comes to strategy and messaging. They can’t do the simple thing. If their goal is to go from A to B, they can’t just move in the direction of B. They’re too smart for that. Instead, they’ll perform Olympic-level mental gymnastics, and convince themselves that doing anything but moving towards B is the right plan. They can’t state plainly what they believe, but pass everything through so many filters that they end up talking about every issue but the one issue they centrally care about.

The very term “AI safety” is an example of how the AI omnicause works in messaging. You start with “superintelligence risks human extinction”: one single, focused premise. And then you put it through the wringer of messaging, PR, respectability, and what comes out the other end is “AI safety”: one vitally important cause is divided into a thousand irrelevant little problems; and the desperate urgency of the situation we find ourselves in, gets diluted away to nothing. Human extinction justifies international agreements, bans, verification regimes. Students using AI to cheat in school does not.

Why the AI omnicause is a losing strategy

If your goal is to reduce the risk of human extinction from superintelligence, the AI omnicause is a bad way to go about it. Why? Simply because none of these disparate causes actually do anything to address the central problem: extinction risk from superintelligence. Again, it’s like you’re trying to travel from A to B, and instead of walking in the direction of B, you go off in a random direction, on the theory that “going somewhere is better than nothing”, hoping you will randomly stumble upon B.

We have to start from the basics: superintelligence leads to human extinction. Therefore, to prevent human extinction, we have to prevent the development of superintelligence. Preventing the development of superintelligence is the goal, and we have to pursue that goal directly.

Mandating pre-deployment testing, evaluation, and certification of models will not stop superintelligence. Rather, it will allow companies to build superintelligence, and later realize that during testing their AI escaped and hacked into another company, or an oversight body might find a universal jailbreak in the AI which nullifies all safety measures, at which point all the evaluation and certification has been for nothing. Working on short-term AI risk—like deepfakes, or job displacement, or disinformation—will not stop superintelligence, and letting companies build a superintelligence that refuses to generate deepfakes is like letting them build a nuclear weapon as long as the plutonium is ethically-sourced and conflict-free. Stopping the construction of datacenters will not stop superintelligence: there is a lot of computing power in the world already, and we can’t simply hope (because it is just that, a hope) that simply reducing the growth rate of computing power—in the one country you advocate in—will prevent superintelligence.

Finally, we have to talk about the moral dimension. The AI omnicause is not just a bad means to the goal. It is morally, ethically wrong. If you, as an expert, are called to give your expert testimony, and in your heart you believe that superintelligence will cause human extinction, but with your words you talk about parochial issues like disinformation and datacenters: you are committing a grave omission. First, you are depriving society of a desperately-needed alarm bell. Secondly, if experts tell people that the risks of superintelligence are those of any other normal technology, those people will walk away expecting things will basically turn out fine—which would be a rational thing to believe, if the risks from superintelligence were ordinary. And in addition to being unethical, this type of dishonesty is deeply corrosive to an organization’s ability to operate, because in practice it’s impossible to maintain the cordon sanitaire between the story you tell the public, and the story you tell yourselves; between the “noble” lies and the truth.

More generally, people routinely forgo long-term gain for short-term gain. Here, the short-term gain is attention from talking about salient issues like datacenters, or avoiding the difficult issue of extinction risk, whereas the long-term gain they are forgoing is actually accomplishing their goal.

How to avoid the AI omnicause

We have to look the goal square in the face, and address it directly: the goal is to stop superintelligence, the means is a ban on the development of superintelligence.

When you talk to policymakers and politicians, talk about extinction risk. The interventions we need to stop superintelligence—international agreements between nuclear powers—are what you might call a big ask, and cannot be justified on the basis of mundane risks. And we would not be working on this problem if we believed that the biggest risk from AI was the environmental impact of datacenters. Therefore, you should put your cards on the table, and say plainly what you believe.

Most importantly: don’t outsmart yourself! Look at the goal, and look at the means, and ask yourself: do the means actually get us to the goal? The winning strategy is never some convoluted, indirect, Rube Goldberg scheme. The winning strategy is a single sentence: talk about extinction risk, ban superintelligence.

Staying on track

Having set their eyes on the right goal, how should the people and organizations working to stop superintelligence stay on track?

First, a pattern you see again and again: an organization starts with a single, specific, well-scoped goal. And over the years it accumulates causes like dying people accumulate medicines. They take on this cause, that grievance, until the original goal has been diluted by an ocean of irrelevance into nothing. In the short-term it doesn’t feel like dilution, it feels like you are gaining in influence and relevance by attaching yourself to popular causes, but over the long run the organization’s effectiveness is ground down to nothing. The solution is to continually, relentlessly reorient the organization towards the goal, and to cut away everything that does not advance the goal. There is one goal: to stop superintelligence.

Second, there is a constant temptation to try and harness the broad anti-tech and anti-AI sentiment towards the goal of banning ASI. This is a superficially attractive but deeply fraught idea. The central problem is that popular anger is vague, while our goal is very specific. To harness popular anger, you need to make your message simpler, remove qualifiers, remove detail, so that it travels as far as possible. But our goal has a lot of detail—international agreements, verification regimes—and anything short of the goal, anything broader or simpler than the goal, will not do. And it’s dishonest: we are not part of the broader anti-tech movement because some tech is good, and some tech is bad. Again: falsifying your beliefs might feel like a win in the short-term, but is destructive in the long-term.

Third, and related to the above, is the temptation to try and join a larger movement or political coalition. A coalition is not a free injection of political capital: it is a constant negotiation with people who have their own goals and interests. The natural result is not that you get your way, rather, even if you are very good at negotiating, you accomplish a simplified, stripped-down, lowest-common-denominator version of your goal. You risk turning the organization into a small remora of a much larger movement that you can neither control nor gain influence from.

Finally, we have to talk about entryism. This is when people join an organization, though they don’t share its values or beliefs, to access that organization’s resources and redirect its priorities to their own goals. And as an organization grows in prestige and resources, it becomes a more attractive target for entryism. We see this everywhere in AI safety: there are organizations where the people who started them care about existential risk, and the people who fund them care about existential risk, but the people working for them don’t even believe in existential risk. And you can’t run an effective organization without shared goals and values. Inevitably, what happens in these cases is that the organization drifts towards the beliefs and priorities of the people actually doing the work, regardless of what the people nominally running the organization believe.

Conclusion

To summarize: the AI omnicause does not work. Keep your eyes on the goal, which is to ban superintelligence. Keep the strategy simple, keep the message simple, tell the truth, and tell it far and wide, continually make sure you are aiming at the goal and nowhere else, and ensure that the people you work with share your goals.