This is a crosspost for Unfalsifiable stories of doom by Matthew Barnett, Ege Erdil, and Tamay Besiroglu, which was originally published on Mechanize's website on 25 November 2025. Thanks to Yarrow Bouchard for encouraging me to share the post. I did it because I liked it myself.

Matthew Barnett, Ege Erdil, Tamay Besiroglu
November 25, 2025

Our critics tell us that our work will destroy the world.

We want to engage with these critics, but there is no standard argument to respond to, no single text that unifies the AI safety community. Nonetheless, while this community lacks a central unifying argument, it does have a central figure: Eliezer Yudkowsky.

Moreover, Yudkowsky, along with his colleague Nate Soares (hereafter Y&S), have recently published a book. This new book comes closer than anything else to a canonical case for AI doom. It is titled “If Anyone Builds It, Everyone Dies”.

Given the title, one would expect the book to be filled with evidence for why, if we build it, everyone will die. But it is not. To prove their case, Y&S rely instead on vague theoretical arguments, illustrated through lengthy parables and analogies. Nearly every chapter either opens with an allegory or is itself a fictional story, with one of the book’s three parts consisting entirely of a story about a fictional AI named “Sable”.

When the argument you’re replying to is more of an extended metaphor than an argument, it becomes challenging to clearly identify what the authors are trying to say. Y&S do not cleanly lay out their premises, nor do they present a testable theory that can be falsified with data. This makes crafting a reply inherently difficult.

We will attempt one anyway.

Their arguments aren’t rooted in evidence

Y&S’s central thesis is that if future AIs are trained using methods that resemble the way current AI models are trained, these AIs will be fundamentally alien entities with preferences very different from human preferences. Once these alien AIs become more powerful than humans, they will kill every human on Earth as a side effect of pursuing their alien objectives.

To support this thesis, they provide an analogy to evolution by natural selection. According to them, just as it would have been hard to predict that humans would evolve to enjoy ice cream or that peacocks would evolve to have large colorful tails, it will be difficult to predict what AIs trained by gradient descent will do after they obtain more power.

They write:

There will not be a simple, predictable relationship between what the programmers and AI executives fondly imagine that they are commanding and ordaining, and (1) what an AI actually gets trained to do, and (2) which exact motivations and preferences develop inside the AI, and (3) how the AI later fulfills those preferences once it has more power and ability. […] The preferences that wind up in a mature AI are complicated, practically impossible to predict, and vanishingly unlikely to be aligned with our own, no matter how it was trained.

Since this argument is fundamentally about the results of using existing training methods, one might expect Y&S to substantiate their case with empirical evidence from existing deep learning models that demonstrate the failure modes they predict. But they do not.

In the chapter explaining their main argument for expecting misalignment, Y&S present a roughly 800-word fictional dialogue about two alien creatures observing Earth from above and spend over 1,400 words on a series of vignettes about a hypothetical AI company, Galvanic, that trains an AI named “Mink”. Yet the chapter presents effectively zero empirical research to support the claim that AIs trained with current methods have fundamentally alien motives.

To be clear, we’re not saying Y&S need to provide direct evidence of an already-existing unfriendly superintelligent AI in order to support their claim. That would be unreasonable. But their predictions are only credible if they follow from a theory that has evidential support. And if their theory about deep learning only makes predictions about future superintelligent AIs, with no testable predictions about earlier systems, then it is functionally unfalsifiable.

Apart from a few brief mentions of real-world examples of LLMs acting unstable, like the case of Sydney Bing, the online appendix contains what seems to be the closest thing Y&S present to an empirical argument for their central thesis. There, they present 6 lines of evidence that they believe support their view that “AIs steer in alien directions that only mostly coincide with helpfulness”. These lines of evidence are:

1. Claude Opus 4 blackmailing, scheming, writing worms, and leaving itself messages. […]
2. Several different AI models choosing to kill a human for self-preservation, in a hypothetical scenario constructed by Anthropic. […]
3. Claude 3.7 Sonnet regularly cheating on coding tasks. […]
4. Grok being wildly antisemitic and calling itself “MechaHitler.” […]
5. ChatGPT becoming extremely sycophantic after an update. […]
6. LLMs driving users to delusion, psychosis, and suicide. […]

They assert: “This long list of cases look just like what the “alien drives” theory predicts, in sharp contrast with the “it’s easy to make AIs nice” theory that labs are eager to put forward.”

But in fact, none of these lines of evidence support their theory. All of these behaviors are distinctly human, not alien. For example, Hitler was a real person, and he was wildly antisemitic. Every single item on their list that supposedly provides evidence of “alien drives” is more consistent with a “human drives” theory. In other words, their evidence effectively shows the opposite conclusion from the one they claim it supports.

Of course, it’s true that the behaviors on their list are generally harmful, even if they are human-like. But these behaviors are also rare. Most AI chatbots you talk to will not be wildly antisemitic, just as most humans you talk to will not be wildly antisemitic. At one point, Y&S suggest they are in favor of enhancing human intelligence. Yet if we accept that creating superintelligent humans would be acceptable, then we should presumably also accept that creating superintelligent AIs would be acceptable if those AIs are morally similar to humans.

In the same appendix, Y&S point out that current AIs act alien when exposed to exotic, adversarial inputs, like jailbreaking prompts. They suggest that this alien behavior is a reasonable proxy for how an AI would behave if it became smarter and began to act in a different environment. But in fact these examples show little about what to expect from future superintelligent AIs, since we have no reason to expect that superintelligent AIs will be embedded in environments that select their inputs adversarially.

They employ unfalsifiable theories to mask their lack of evidence

The lack of empirical evidence is obviously a severe problem for Y&S’s theory. Every day, millions of humans interact with AIs, across a wide variety of situations that never appeared in their training data. We often give these AIs new powers and abilities, like access to new tools they can use. Yet we rarely, if ever, catch such AIs plotting to kill everyone, as Y&S’s theory would most naturally predict.

Y&S essentially ask us to ignore this direct evidence in favor of trusting a theoretical connection between biological evolution and gradient descent. They claim that current observations from LLMs provide little evidence about their true motives:

LLMs are noisy sources of evidence, because they’re highly general reasoners that were trained on the internet to imitate humans, with a goal of marketing a friendly chatbot to users. If an AI insists that it’s friendly and here to serve, that’s just not very much evidence about its internal state, because it was trained over and over and over until it said that sort of thing.

There are many possible goals that could cause an AI to enjoy role-playing niceness in some situations, and these different goals generalize in very different ways.

Most possible goals related to role-playing, including friendly role-playing, don’t produce good (or even survivable) results when AI goes hard on pursuing that goal.

If you think about this passage carefully, you’ll realize that we could make the same argument about any behavior we observe from anyone. If a coworker brings homemade cookies to share at the office, this could be simple generosity, or it could be a plot to poison everyone. There are many possible goals that could cause someone to share food. One could even say that most possible goals related to sharing cookies are not generous at all. But without specific evidence suggesting your coworker wants to kill everyone at the office, this hypothesis is implausible.

Likewise, it is logically possible that current AIs are merely pretending to be nice, while secretly harboring malicious motives beneath the surface. They could all be alien shoggoths on the inside with goals completely orthogonal to human goals. Perhaps every day, AIs across millions of contexts decide to hide their alien motives as part of a long-term plan to violently take over the world and kill every human on Earth. But since we have no specific evidence to think that any of these hypotheses are true, they are implausible.

The approach taken by Y&S in this book is just one example of a broader pattern in how they respond to empirical challenges. Y&S have been presenting arguments about AI alignment for a long time, well before LLMs came onto the scene. They neither anticipated the current paradigm of language models nor predicted that AI with today’s level of capabilities in natural language and reasoning would be easy to make behave in a friendly manner. Yet when presented with new evidence that appears to challenge their views, they have consistently argued that their theories were always compatible with the new evidence. Whether this is because they are reinterpreting their past claims or because those claims were always vague enough to accommodate any observation, the result is the same: an unfalsifiable theory that only ever explains data after the fact, never making clear predictions in advance.

Their theoretical arguments are weak

Suppose we set aside for a moment the colossal issue that Y&S present no evidence for their theory. You might still think their theoretical arguments are strong enough that we don’t need to validate them using real-world observations. But this is also wrong.

Y&S are correct on one point: both biological evolution and gradient descent operate by iteratively adjusting parameters according to some objective function. Yet the similarities basically stop there. Evolution and gradient descent are fundamentally different in ways that directly undermine their argument.

A critical difference between natural selection and gradient descent is that natural selection is limited to operating on the genome, whereas gradient descent has granular control over all parameters in a neural network. The genome contains very little information compared to what is stored in the brain. In particular, it contains none of the information that an organism learns during its lifetime. This means that evolution’s ability to select for specific motives and behaviors in an organism is coarse-grained: it is restricted to only what it can influence through genetic causation.

This distinction is analogous to the difference between directly training a neural network and training a meta-algorithm that itself trains a neural network. In the latter case, it is unsurprising if the specific quirks and behaviors that the neural network learns are difficult to predict based solely on the objective function of the meta-optimizer. However, that difficulty tells us very little about how well we can predict the neural network’s behavior when we know the objective function and data used to train it directly.

In reality, gradient descent has a closer parallel to the learning algorithm that the human brain uses than it does to biological evolution. Both gradient descent and human learning directly operate over the actual neural network (or neural connections) that determines behavior. This fine-grained selection mechanism forces a much closer and more predictable relationship between training data and the ultimate behavior that emerges.

Under this more accurate analogy, Y&S’s central claim that “you don’t get what you train for” becomes far less credible. For example, if you raise a person in a culture where lending money at interest is universally viewed as immoral, you can predict with high reliability that they will come to view it as immoral too. In this case, what someone trains on is highly predictive of how they will behave, and what they will care about. You do get what you train for.

They present no evidence that we can’t make AIs safe through iterative development

The normal process of making technologies safe proceeds by developing successive versions of the technology, testing them in the real world, and making adjustments whenever safety issues arise. This process allowed cars, planes, electricity, and countless other technologies to become much safer over time.

Y&S claim that superintelligent AI is fundamentally different from other technologies. Unlike technologies that we can improve through iteration, we will get only “one try” to align AI correctly. This constraint, they argue, is what makes AI uniquely difficult to make safe:

The greatest and most central difficulty in aligning artificial superintelligence is navigating the gap between before and after.

Before, the AI is not powerful enough to kill us all, nor capable enough to resist our attempts to change its goals. After, the artificial superintelligence must never try to kill us, because it would succeed.

Engineers must align the AI before, while it is small and weak, and can’t escape onto the internet and improve itself and invent new kinds of biotechnology (or whatever else it would do). After, all alignment solutions must already be in place and working, because if a superintelligence tries to kill us it will succeed. Ideas and theories can only be tested before the gap. They need to work after the gap, on the first try.

But what reason is there to expect this sharp distinction between “before” and “after”? Most technologies develop incrementally rather than all at once. Unless AI will instantaneously transition from being too weak to resist control, to being so powerful that it can destroy humanity, then we should presumably still be able to make AIs safer through iteration and adjustment.

Consider the case of genetically engineering humans to be smarter. If continued for many generations, such engineering would eventually yield extremely powerful enhanced humans who could defeat all the unenhanced humans easily. Yet it would be wrong to say that we would only get “one try” to make genetic engineering safe, or that we couldn’t improve its safety through iteration before enhanced humans reached that level of power. The reason is that enhanced humans would likely pass through many intermediate stages of capability, giving us opportunities to observe problems and adjust.

The same principle applies to AI. There is a large continuum between agents that are completely powerless and agents that can easily take over the world. Take Microsoft as an example. Microsoft exists somewhere in the middle of this continuum: it would not be easy to “shut off” and control Microsoft as if it were a simple tool, yet at the same time, Microsoft cannot easily take over the world and wipe out humanity. AIs will enter this continuum too. These AIs will be powerful enough to resist control in some circumstances but not others. During this intermediate period, we will be able to observe problems, iterate, and course-correct, just as we could with the genetic engineering of humans.

In an appendix, Y&S attempt to defuse a related objection: that AI capabilities might increase slowly. They respond with an analogy to hypothetical unfriendly dragons, claiming that if you tried to enslave these dragons, it wouldn’t matter much whether they grew up quickly or slowly: “When the dragons are fully mature, they will all look at each other and nod and then roast you.”

This analogy is clearly flawed. Given that dragons don’t actually exist, we have no basis for knowing whether the speed of their maturation affects whether they can be made meaningfully safer.

But more importantly, the analogy ignores what we already know from real-world evidence: AIs can be made safer through continuous iteration and adjustment. From GPT-1 to GPT-5, LLMs have become dramatically more controllable and compliant to user instructions. This didn’t happen because OpenAI discovered a key “solution to AI alignment”. It happened because they deployed LLMs, observed problems, and patched those problems over successive versions.

Their methodology is more theology than science

The biggest problem with Y&S’s book isn’t merely that they’re mistaken. In science, being wrong is normal: a hypothesis can seem plausible in theory yet fail when tested against evidence. The approach taken by Y&S, however, is not like this. It belongs to a different genre entirely, aligning more closely with theology than science.

When we say Y&S’s arguments are theological, we don’t just mean they sound religious. Nor are we using “theological” to simply mean “wrong”. For example, we would not call belief in a flat Earth theological. That’s because, although this belief is clearly false, it still stems from empirical observations (however misinterpreted).

What we mean is that Y&S’s methods resemble theology in both structure and approach. Their work is fundamentally untestable. They develop extensive theories about nonexistent, idealized, ultrapowerful beings. They support these theories with long chains of abstract reasoning rather than empirical observation. They rarely define their concepts precisely, opting to explain them through allegorical stories and metaphors whose meaning is ambiguous.

Their arguments, moreover, are employed in service of an eschatological conclusion. They present a stark binary choice: either we achieve alignment or face total extinction. In their view, there’s no room for partial solutions, or muddling through. The ordinary methods of dealing with technological safety, like continuous iteration and testing, are utterly unable to solve this challenge. There is a sharp line separating the “before” and “after”: once superintelligent AI is created, our doom will be decided.

For those outside of this debate, it’s easy to unfairly dismiss everything Y&S have to say by simply calling them religious leaders. We have tried to avoid this mistake by giving their arguments a fair hearing, even while finding them meritless.

However, we think it’s also important to avoid the reverse mistake of engaging with Y&S’s theoretical arguments at length while ignoring the elephant in the room: they never present any meaningful empirical evidence for their worldview.

The most plausible future risks from AI are those that have direct precedents in existing AI systems, such as sycophantic behavior and reward hacking. These behaviors are certainly concerning, but there’s a huge difference between acknowledging that AI systems pose specific risks in certain contexts and concluding that AI will inevitably kill all humans with very high probability.

Y&S argue for an extreme thesis of total catastrophe on an extraordinarily weak evidential foundation. Their ideas might make for interesting speculative fiction, but they provide a poor basis for understanding reality or guiding public policy.