1. The Video That Makes Doom Funny
There is something deeply revealing about the viral AI-doom video “I’m Upping My P(doom)”. It is clever, funny and technically impressive. It takes a dense collection of AI-risk vocabulary—AGI, the singularity, alignment, recursive self-improvement, paperclip maximizers, the orthogonality thesis, RLHF and the basilisk—and turns it into a fast, catchy animated rock-and-rap performance.
The problem is not that the video says every AI danger is imaginary. The problem is almost the opposite. It takes a mixture of real technological developments, legitimate safety concerns, philosophical thought experiments and highly speculative scenarios, blends them together, adds comedy and music, and then gives the audience an emotional experience in which the entire apocalypse eventually becomes a theatrical joke.
That is entertaining. But is it educational? More importantly, what happens to our perception of genuine AI risks when the risks themselves become entertainment?
2. The Cute Tiger Cub
Imagine somebody showing you a baby tiger. It is tiny, playful and apparently harmless. It rolls around, jumps at a ball and makes everyone laugh.
None of that changes what the animal is. A tiger cub does not become less dangerous simply because it is cute; it eventually grows into a tiger.
This is the Cute Tiger Cub Problem. Presentation and underlying capability are different things. A frightening soundtrack does not make an object more dangerous, and a funny soundtrack does not make it safer.
The same principle applies to AI. A cute cartoon representation of an AI system does not tell us how dangerous that system may become, just as a terrifying visual does not prove that the underlying technology is dangerous.
Humans are extraordinarily responsive to narrative. We do not merely process information; we experience it. Music tells us how to feel, animation tells us what deserves attention, and comedy tells us what should not be taken too seriously.
That is precisely what makes modern AI-generated media so powerful. The system can take a complicated subject and turn it into an emotionally memorable experience before the audience has had an opportunity to examine the assumptions behind it.
3. The Video Is an Extraordinary Demonstration of AI’s Power
There is an irony here. The video itself demonstrates one of the capabilities we should be thinking about when we discuss AI.
The creator’s public repository says that Claude Opus 5.5 generated essentially the entire animation project from broad instructions: use the Clawd character, give each lyric interesting visuals and transitions, and make the result lively and ambitious. Opus generated the storyboard and animation guide, while the final project contains nine chapters of animation code.
AI therefore did not merely provide a paragraph of commentary about AI risk. It transformed an abstract collection of ideas into a coherent emotional experience.
A traditional library stores information. A database retrieves information. A modern generative AI system can transform information into new representations, arguments, stories, plans, code and images, and increasingly interact with tools and the world around it.
That does not make AI inherently dangerous. But it does mean that the phrase “AI is just a library” seriously understates what modern generative systems can do.
The more important question is what happens when a system with those capabilities is used not merely to retrieve information but to construct an interpretation of it.
4. Where the Video Becomes Interesting
Consider the opening idea: “I see sparks of AGI.” That is not simply invented nonsense. The phrase echoes Microsoft’s 2023 Sparks of Artificial General Intelligence paper, which argued that GPT-4 displayed unusually broad capabilities while also discussing its limitations.
So the video begins with a legitimate research reference, and that is precisely what makes the transition that follows so important. The narrative then accelerates.
Training becomes sudden capability emergence; capability becomes the singularity; optimization becomes runaway optimization; misalignment becomes rebellion; a thought experiment becomes an apocalypse; and a probability becomes a rising thermometer.
Each transition can be made entertaining. But entertainment hides the gaps between them.
Consider the difference between these two statements: “AI systems are becoming more capable” and “AI systems will become recursively self-improving superintelligences that escape human control.” The first is an empirical statement. The second is a possible future scenario involving numerous additional assumptions.
The distance between those two statements is where much of the serious AI debate actually lives. A three-minute song can cross that distance in a few seconds; serious analysis cannot.
5. The P(doom) Meter
One of the video’s most effective visual devices is the rising P(doom) meter. The number climbs through the song, giving the audience a powerful visual impression that the evidence is accumulating and the probability of catastrophe is correspondingly increasing.
But the meter is not actually performing Bayesian inference. It is performing Bayesian-looking storytelling.
Suppose we observe a new AI capability. We might update our assessment of a particular risk, but how much we should update depends on the causal model behind that risk. A capability might increase one risk while decreasing another, while a new safety technique might reduce one failure mode and introduce another.
The number on the meter cannot settle those questions. It rises because the story requires the number to rise.
That is excellent animation, but it is not evidence. The distinction matters beyond this particular video.
AI can increasingly turn uncertainty into visual certainty: a chart can make speculation look quantitative, a graph can make an assumption look measured, and a probability meter can make a narrative look Bayesian.
The underlying epistemology has not necessarily changed. Only the presentation has.
6. Real Risks Are Not the Problem
It would be a mistake to respond to the video by saying that AI doom is nonsense. That would make exactly the opposite mistake.
AI has genuine risks. There are serious concerns around misuse, cybersecurity, misinformation, persuasion, autonomous agents, model failures and the difficulty of evaluating increasingly capable systems.
AI-safety researchers themselves acknowledge that alignment evaluation remains an immature science. Frontier systems are increasingly being developed with substantial real-world affordances, making questions about unwanted actions, autonomy and loss of control more consequential.
These are not imaginary problems. But acknowledging them does not require accepting every argument made in their name.
That distinction is crucial. Real risks do not automatically validate every risk narrative.
For a separate examination of whether the control problem itself may be tractable, see The AI Control Problem May Be Solvable.
7. The Y2K Problem
One rhetorical move worth examining is the use of historical technological scares such as Y2K.
Y2K was a genuine engineering problem involving the way computer systems represented dates. Its solution depended on identifying affected systems, modifying them and testing the modifications.
The fact that civilization did not collapse in 2000 does not demonstrate that concern about Y2K was irrational. A large part of the reason catastrophe was avoided was that enormous effort had already been spent mitigating the problem.
But the reverse inference is equally problematic: people worried about Y2K, catastrophe did not happen, therefore today’s AI warnings should be viewed through the same lens.
AI is not Y2K. AI systems learn, generalize and generate; they can interact with humans, write and execute code, and increasingly operate through tools. Future systems may possess capabilities that present systems do not.
The correct lesson from Y2K is therefore not that technology warnings are always exaggerated. It is that technology risks must be evaluated according to the mechanism that creates the risk.
A date-format bug and a potentially autonomous learning system are different engineering problems. Historical analogy cannot substitute for mechanism.
8. The Video’s Most Important Trick
The most interesting rhetorical technique is not any individual joke. It is epistemic blending.
The video places different kinds of claims next to each other until their differences become difficult to see. There are things we know, things we have observed but do not fully understand, plausible future scenarios, philosophical thought experiments, highly speculative possibilities and, finally, jokes.
The video can move through all of these categories without stopping.
For example: Transformers → scaling → recursive self-improvement → paperclip maximizer → extinction. The arrows look natural. But every arrow is a hypothesis.
The same applies to: alignment failure → deception → autonomy → loss of control. Each step requires examination. A song does not have time for that, so the audience receives the entire chain as one emotional object.
A serious educational treatment would stop at each transition and ask: What evidence takes us from here to there?
For a broader analysis of current model limitations and possible solutions, see The Problems with Current AI Models — and Their Probable Solutions.
9. The Narrative Techniques Deserve Scrutiny
The video also illustrates several techniques that can make a complicated argument easier to absorb and harder to examine. None of these techniques is inherently deceptive. They are normal tools of communication. The important question is what happens when they are combined around claims whose evidentiary status varies.
Simplification. A complicated technology can be reduced to an easily digestible metaphor such as “AI is just a library.” Metaphors are useful because they make unfamiliar systems understandable. But a metaphor can also remove important capabilities and distinctions, leaving the audience with a model that is memorable but incomplete.
Historical analogy. Y2K provides a familiar example of technological anxiety. Historical analogies can illuminate a problem, but they can also obscure differences in mechanism. A date-format problem and an increasingly autonomous learning system are not the same type of technological risk.
Borrowed authority. References to respected researchers, institutions or historical episodes can make an argument feel validated. But the authority of the source does not automatically establish that a particular inference follows from the source. The evidence still has to be examined at the point where the inference is made.
Narrative selection. Every story selects some facts and leaves others outside the frame. A narrative emphasizing capability breakthroughs, failed safeguards and extreme scenarios will produce a different impression from one emphasizing successful evaluations, safety improvements and beneficial applications. Neither selection automatically establishes the conclusion; the omitted alternatives matter.
Emotional amplification. Music, animation, humour, pacing and visual repetition can increase the emotional force of an uncertain proposition. The proposition may remain uncertain while the audience’s feeling about it becomes much stronger.
False binary. A complicated spectrum of AI risks can be compressed into a choice between “AI doomers were right” and “AI doom is ridiculous.” That framing leaves little room for the more useful position: some risks may be serious, some arguments may be weak, and different risks may have very different probabilities and mechanisms.
These techniques become especially important when they are combined. Simplification determines the mental model, narrative selection determines which evidence enters that model, authority can strengthen its credibility, emotional amplification strengthens its psychological impact, and a false binary can make alternative interpretations harder to see.
This does not establish deliberate manipulation. The same techniques can be used by an honest educator trying to make a difficult subject understandable. The question is whether the resulting narrative helps the audience distinguish evidence from interpretation, or makes that distinction harder.
There is an important irony here. A video warning about AI’s ability to manipulate information can itself demonstrate how information can become persuasive through selection, framing and emotional amplification.
The danger is therefore not simply that AI can lie. The deeper danger is that AI can make a selected interpretation of reality extraordinarily persuasive.
10. Thought Experiments Are Not Forecasts
The paperclip maximizer is a useful example. The basic idea is valuable: a sufficiently capable optimizer pursuing a badly specified objective could produce outcomes radically different from what its designers intended.
That is a serious alignment concept. But the literal paperclip apocalypse is not an empirical forecast; it is a thought experiment.
The distinction is crucial. Thought experiments help us explore what could happen under certain assumptions. They do not establish that those assumptions will hold.
The same applies to concepts such as the basilisk and other extreme scenarios. They can be intellectually interesting, reveal weaknesses in an argument and help researchers think about unusual failure modes.
But placing them next to actual developments in AI can give the audience the impression that all of them occupy the same evidentiary category. They do not.
11. “Ilya Saw Something”
Authority creates another subtle problem because a reference to a famous AI researcher can carry enormous rhetorical weight. If someone says, “What did Ilya see? We’ll never know,” the sentence contains almost no information, yet it creates mystery.
An unknown observation becomes a dramatic suggestion that perhaps something frightening was discovered behind closed doors. That is not evidence; it is borrowed authority combined with uncertainty.
The audience supplies the missing information, and humans are remarkably good at filling empty spaces with stories.
This is a broader problem with persuasive AI narratives. The model does not need to invent a false fact. Sometimes it only needs to place a genuine person, genuine event or genuine quotation inside a carefully constructed narrative and allow the audience to make the connection.
12. Satire Can Cut Both Ways
To be fair, satire has an important role. It can expose irrationality, puncture exaggerated claims, make difficult subjects accessible and force us to laugh at ourselves.
There is nothing inherently wrong with making AI risk funny. The question is what the comedy causes the audience to believe.
If someone watches the video and thinks, “Some AI-risk arguments are exaggerated, and I should investigate which ones,” the satire has done something useful.
But if the emotional takeaway becomes, “AI doom is basically a ridiculous cult,” then the satire has done something very different.
The problem becomes particularly serious when the underlying subject contains genuine risks. Comedy can reduce fear, and sometimes that is healthy. But it can also reduce vigilance.
A fire alarm played as a comedy routine would still be an alarm. The laughter does not extinguish the fire.
For another HPB examination of how AI risk narratives can interact with broader social and technological change, see The Risk of AI enabled Utopian Dystopia.
13. Another Perspective: Is the Video Actually Defusing AI Doom?
There is another way to interpret the video, and it may be almost the opposite of the reading developed above. Perhaps the video is not primarily trying to convince the audience that AI doom is coming. Perhaps it is doing something subtler: making AI doom entertaining.
That distinction matters because entertainment changes how an audience receives an idea. A frightening proposition presented as a frightening proposition can trigger resistance. The same proposition wrapped in humour, music, cartoons and absurdity can become familiar, and eventually funny.
Oscar Wilde’s observation is relevant here: “If you want to tell people the truth, make them laugh, otherwise they’ll kill you.” The point is not whether the quotation proves anything about this particular video. It illustrates a broader communication mechanism: humour can make difficult or threatening ideas easier to approach.
The reaction to the AI-generated music on X captures the same phenomenon from another direction. The immediate response is not necessarily “this is frightening.” It can be “this is incredibly good,” “this is funny,” or simply “I want to watch this again.” The subject may be AI doom, while the emotional experience is entertainment.
A viewer may therefore begin with the thought, “AI could destroy the world,” and finish with something very different: “AI doom is hilarious.” The factual content has not necessarily changed. The emotional framing has.
Satire can therefore work in two directions. It can expose exaggerated claims and make the audience more skeptical. But it can also lower the emotional guard surrounding the subject. Once a frightening idea becomes familiar, funny and culturally recognizable, it may no longer produce the same level of caution.
This raises an uncomfortable possibility: the video may be simultaneously talking about AI doom and defusing AI-doom anxiety. That does not establish the creator’s intention, and timing alone cannot establish a hidden motive. But the psychological effect is worth examining independently of intent.
The distinction is between what a piece of media says and what the piece of media does to its audience. A horror film talks about monsters while making audiences afraid of them. A comedy about monsters can talk about exactly the same monsters while making them seem ridiculous.
The subject matter can therefore be identical while the psychological direction is opposite. That may be the more interesting question about this video: is it warning us about AI doom, teaching us not to take AI doom too seriously, or doing some of both?
But there is an important counterpoint. Holding opposing ideas in mind is not itself a manipulation technique. F. Scott Fitzgerald famously described the ability to hold two opposing ideas in mind while retaining the ability to function as a test of first-rate intelligence. The constructive version of this ability is intellectual flexibility: instead of choosing a side too early, we keep both possibilities visible while examining the evidence.
That is exactly what good AI-risk analysis should do. AI could produce enormous benefits, and AI could produce enormous harms. We should be able to examine both propositions without allowing either one to automatically cancel the other.
The potential problem begins when someone else controls not only the two opposing ideas, but also the framing, emotional tone and implied conclusion attached to them. Then the same cognitive capacity can become dual-use. It can help us think more deeply, or it can make a carefully constructed narrative harder to resist.
This distinction matters because cognitive dissonance, double binds, gaslighting and related forms of manipulation are not interchangeable concepts. A contradiction can simply be evidence of a genuinely complicated reality. The presence of two opposing ideas is therefore not evidence of manipulation; the mechanism, context, incentives and resulting pressure on the audience have to be examined separately.
There is also a wider timing question worth examining, without turning timing into proof of intent. Anthropic announced on June 1, 2026 that it had confidentially submitted a draft S-1 to the U.S. Securities and Exchange Commission for a proposed IPO. The company said that the offering would depend on market conditions and other factors.
None of that demonstrates a connection between Anthropic, the IPO and this video. It does, however, make the broader communication environment more interesting. AI companies now operate in a world where capability, safety warnings, public trust, investment and cultural narratives about AI are all interacting at once.
This does not require a conspiracy or a hidden agreement. An organization can have legitimate reasons to communicate both the benefits and the risks of its technology while also having financial, competitive or reputational incentives that influence how those messages are framed. The important analytical question is therefore not whether a particular company is “pro-AI” or “anti-AI”, but whether its incentives could systematically affect which uncertainties are emphasized, which are minimized and which emotional framing reaches the public.
There is one more variable in this process: the observer. The effect of a narrative is not determined by the narrative alone. It also depends on the observer’s understanding, emotional maturity, prior beliefs and ability to tolerate uncertainty.
A sophisticated observer can enjoy the music and humour while still asking what the video is actually claiming. Another observer may experience the same humour as evidence that the underlying subject is ridiculous. The message has not changed, but the interpretation has.
This gives us a useful three-layer model. Wilde represents the psychological side: how does the framing of an idea affect the audience? Fitzgerald represents the epistemic side: can we hold opposing possibilities in mind long enough to examine the evidence? The observer connects the two.
The mature response to AI risk is therefore neither automatic fear nor automatic reassurance. It is the ability to hold two propositions at the same time: AI may create extraordinary benefits, and AI may create extraordinary risks. The task is then to ask what evidence would distinguish the competing possibilities.
This is where emotional maturity and intellectual maturity meet. Emotional maturity helps us tolerate uncertainty without immediately reaching for comfort or catastrophe. Intellectual maturity helps us separate observation from interpretation, interpretation from prediction, and prediction from evidence.
The resulting chain is more complicated than “AI manipulates humans.” It is message → framing → observer → interpretation → belief → action. Distortion can enter at any point in that chain.
That may ultimately be the deeper lesson of the Cute Tiger Cub Problem. The tiger is one thing. The presentation of the tiger is another. But what the observer sees depends partly on the observer as well.
The deeper question is therefore not “Who secretly ordered this video?” We have no evidence for such a claim. The more useful questions are: Who controls the emotional framing of AI risk? What incentives shape that framing? And what happens when existential fear becomes entertainment?
That is why analysing only the literal claims is insufficient. We also have to examine the direction in which the narrative moves the audience.
14. The Same AI Could Make the Opposite Video
Here is perhaps the most revealing experiment. Give the same powerful AI the same broad body of information and ask it to create the strongest possible case that AI will produce an unprecedented economic and scientific boom. It could probably make that too.
Ask it for the strongest case that AI will cause mass unemployment, transform education, undermine human expertise or create extraordinary scientific progress, and it can construct those narratives as well.
The model does not necessarily have to change. The frame changes.
That brings us back to the old warning: “There are three kinds of lies: lies, damned lies, and statistics.”
Modern AI adds another possibility: there are facts, interpretations—and machines that can turn either into extraordinarily persuasive stories.
This is why another formulation matters: Experiments give you observations; assumptions give them meaning. AI can be extraordinarily good at the second part.
15. AI Can Industrialize Human Bias
We often imagine AI as a machine that removes human bias. Sometimes it can help do that. But a sufficiently capable generative system can also industrialize bias.
A human might write one persuasive article. An AI system can generate a hundred versions. A human might make one compelling video; an AI system can create thousands of variations, each adapted to a different audience.
The problem therefore isn’t simply whether AI itself is biased. The deeper question is: Who controls the frame that the AI is asked to amplify?
A system capable of generating explanations, images, videos, music and arguments can become a force multiplier for whatever assumptions enter the system at the beginning.
That is much more powerful than a library. A library waits for you to interpret the information; a generative AI can participate in the interpretation.
16. The Tree of Life and Flat Earth Problem
This is closely related to a broader problem we have examined before. Complex subjects are often made easier to understand by simplifying them, but simplification has a cost.
If a complicated biological history is reduced to a single picture, important distinctions can disappear. If a complicated scientific question is reduced to a collection of provocative claims, a false framework can sometimes feel intellectually complete.
The danger is not simply factual error. It is removing the distinctions that allow the reader to evaluate the claim.
The same thing can happen with AI. Real capability, plausible risk, speculative scenario and joke can become one seamless story, and the viewer may no longer see the boundaries between them.
17. The Cute Tiger Cub Returns
This is why the title matters. A tiger cub is cute. That is real. It is also a tiger. That is equally real, and the two facts can coexist.
AI can be astonishing. It can help people, produce beautiful things, write music, solve problems and generate comedy. It can even make a sophisticated AI-doom video almost entirely by itself.
AI can also create genuine risks. These statements are not contradictory.
The mistake is believing that one emotional impression cancels the other.
Cute does not mean safe. Scary does not mean dangerous. Funny does not mean harmless. Viral does not mean true.
And perhaps most importantly: A compelling narrative is not evidence for the conclusion embedded inside it.
18. What Should an Educational AI-Risk Video Do?
A genuinely educational treatment would separate the categories. It would tell the audience what we know, what we have observed but do not yet understand, what researchers consider plausible, and what remains speculative.
It would present the strongest argument for a risk and the strongest argument against it. It would identify the assumptions on which the disagreement depends rather than silently moving from one assumption to another.
That would probably be slower. It might also be less viral and less entertaining.
But the audience would leave with something more valuable than a feeling. They would leave with a map of the uncertainty.
And that is exactly what we need as AI becomes more powerful.
19. The Real Irony
The most ironic thing about the video is that it may accidentally demonstrate the very phenomenon that deserves much more attention.
AI is becoming extraordinarily good at taking complicated information and turning it into memorable human experiences. That is a tremendous opportunity, but it is also a tremendous responsibility.
The future will not be shaped only by what AI can calculate. It will also be shaped by what AI can persuade people to believe.
The danger is therefore not simply that AI might someday become too intelligent. There is another, more immediate question: What happens when a machine becomes extremely good at making an interpretation of reality feel obvious?
That is why we should neither laugh away AI risks nor accept every AI-doom narrative as prophecy.
We need something harder. We need to learn to separate the data from the interpretation, the interpretation from the prediction, the prediction from the probability, and finally the story from the truth.
A baby tiger can be adorable. It can also grow into a tiger.
The wise response is not to panic every time we see a cub. Nor is it to assume that because the cub is cute, there is nothing to worry about.
Watch the tiger. Study what it can actually do. Understand the mechanisms. And don’t let the soundtrack make the decision for you.