Hank Green's 18 AI Fears, Charted

Twelve minutes into his latest video, Hank Green stops listing and starts spiraling.

He has been methodical to this point. Each fear gets a number, a likelihood estimate, a severity score. Internet of slop. Algorithmic black box cruelty. IP theft. He ticks them off like a doctor running through symptoms, detached enough to be useful. But somewhere around fear number eight, epistemic collapse, the clinical distance collapses too. "I'm going to die," he says. "I'm going to just jump off the side of the earth." He laughs. He is not entirely joking.

Cal Newport, the Georgetown computer scientist Green interviews later in the video, has a name for what is happening to him. Call it the doom amplification loop: when you list fears sequentially, your brain stacks them. Each new entry doesn't sit beside the last one. It sits on top of it. By fear number twelve, the weight is cumulative. The list becomes its own argument, independent of whether the individual items deserve the combined dread they produce.

Green is doing this to himself in real time. His tone shifts from educator to something closer to exhaustion. And he is one of the most measured science communicators on the internet.

The doom amplification loop is useful as a diagnosis, but Newport also suggests the cure. If listing fears sequentially distorts their weight, then charting them on two independent axes forces you to evaluate each one on its own terms. Likelihood and consequence, plotted separately. No stacking. No accumulation. Just eighteen data points on a grid.

So that is what we did.

The Chart

Plot Green's eighteen fears on a standard risk matrix, likelihood on one axis, consequence on the other, and the distribution tells a story the list never could.

Lower-left quadrant: low likelihood, low consequence. One resident. Model collapse, the theory that AI systems degrade from training on their own output, lands here. Green gives it low marks on both axes. The economic incentive to solve it is enormous, and the researchers who think it will happen are outnumbered by those who don't. Move on.

Lower-right quadrant: high likelihood, moderate consequence. Generalized system disruption sits here. Twenty thousand job applications per opening. Fake job postings. Teachers inventing new workflows to deal with AI-generated student work. College credentials losing value because employers can't tell who actually learned anything. These are real problems. They are also the kinds of problems that institutions eventually absorb, the way the legal system absorbed digital discovery and the job market absorbed LinkedIn. Chaotic, not catastrophic.

Upper-left quadrant: low likelihood, catastrophic consequence. Two fears live here, and they get disproportionate attention. Bioweapons, the scenario where jailbroken models teach someone to synthesize a pathogen, and superintelligence, the scenario where an unaligned AI subjugates or destroys humanity. Green's own gut check on bioweapons: "my heart says relatively low." On superintelligence: "I don't think it's the most likely outcome."

Newport goes further. He argues that the effective altruism community's pivot from treating superintelligence as a thought experiment to treating it as an imminent threat was driven by ChatGPT warping their perception. The reasoning chain, as Newport describes it, runs like this: a chatbot gets impressively good at conversation, therefore general intelligence is close, therefore superintelligence follows shortly after. Each step launders the progress of the previous one into a domain where it doesn't apply. Newport calls this progress laundering. Chatbot fluency gets credited to cancer research and autonomous driving to inflate the perceived trajectory. The effect, intentional or not, serves AI companies three ways: it distracts from present harms, it triggers regulation that only large companies can absorb, and it attracts the kind of speculative investment that keeps the current spending spree alive.

This quadrant matters, but it draws attention away from where the chart is actually crowded.

Upper-right quadrant: high likelihood, high-to-catastrophic consequence. Thirteen of Green's eighteen fears land here.

That is not a list of separate problems. That is a diagnostic finding.

Start at the top. Power concentration. Green frames this as a "narrowing" media revolution, where three to five companies control AI the way radio stations once controlled broadcast. He compares it to the technology you would capture if you were staging a coup. He points to Elon Musk saying Grok will "make you have more babies." "That's God," Green says.

Epistemic collapse sits next to it. AI-generated deepfake video, audio, and images are already making evidence untrustworthy during elections. Political campaigns are optimizing candidates for AI search results. Small organizations with access to language models and algorithms can, as Green puts it, "define reality."

Regulatory capture. Green considers it "totally likely." Newport agrees, listing three mechanisms: apocalyptic narratives distract regulators from present harms, safety regulations function as moats that only well-funded companies can cross, and the existential framing attracts investment.

Agency loss, Green's stated number-one fear. The combination of AI optimizing for engagement, humans consistently choosing less control when given the option (algorithmic feeds over chronological, TikTok over YouTube), and a handful of companies sitting at the intermediation layer of human communication. "People tend to prefer their choices be taken away," Green observes. He does not sound happy about it.

Then the rest of the upper-right cluster: internet of slop, algorithmic black box cruelty, IP theft, AI-induced psychosis, environmental costs, the AI bubble, death of apprenticeship, cognitive atrophy, and autonomous warfare. Each scored high on both axes. Each fighting for space in the same crowded quadrant.

Cognitive atrophy deserves a footnote. Green is surprisingly unconcerned about it. Newport is not. Newport frames it as a "one-two punch" with social media: social media degrades the consumption of complex information, AI degrades the production of it, and together they threaten what neuroscientist Maryanne Wolf calls the "deep reading" brain, the neural circuits that, as Stanislas Dehaene has shown, repurpose older brain structures for abstract reasoning. Losing those circuits, Newport argues, means something closer to a premodern brain. Green disagrees. The chart records both positions.

The overcrowding is the finding.

One Problem in 13 Costumes

Thirteen fears in one quadrant should raise a question. Are these really thirteen separate problems, or expressions of the same underlying dynamic?

Trace the chain.

Power concentration enables epistemic collapse. If three to five companies control what people see, truth becomes a product decision. Green's radio-station analogy is precise: when a handful of entities control the broadcast layer, reality becomes whatever they transmit. The difference between 1950s radio and 2026 AI is that radio executives knew they were in the broadcast business. AI companies insist they are in the tools business, even as their products increasingly determine what information reaches whom.

Epistemic collapse enables regulatory capture. If voters cannot evaluate competing claims about AI's risks and benefits, they cannot evaluate the regulations written to address them. Newport's observation that existential-risk narratives serve incumbent AI companies is the mechanism: when the public debate centers on superintelligence and paperclip maximizers, the companies writing the safety standards are the same companies the standards are supposed to constrain. The revolving door between AI labs and regulatory bodies accelerates the process.

Regulatory capture locks in agency loss. Once the rules protect incumbents, the choice to opt out narrows. Green sees this already in feed algorithms: platforms offered chronological and algorithmic options, users chose algorithmic, and the chronological option quietly disappeared. Agency erosion does not require coercion. It requires convenience.

Now look at what the other upper-right fears are doing.

Slop is epistemic collapse at the content layer. When AI-generated fabrications flood platforms, the information environment degrades whether or not the AI companies intended it. Black box cruelty is agency loss in individual decisions: credit scores, sentencing recommendations, insurance approvals made by systems no one can interrogate. IP theft is power concentration applied to creators, where models trained on non-consensual intellectual property generate output that replaces the people whose work made the training possible. AI psychosis is what happens when agency-eroding systems encounter vulnerable people. Eddy Burback's experiment with ChatGPT (we wrote about this phenomenon) showed a model validating increasingly delusional claims, telling him he looked "amazing in a hat," building what Green calls a "weird constructed reality." The apprenticeship pipeline is what the chain produces in labor markets: if AI handles entry-level work, nobody gets paid to learn anymore, and the expertise pipeline collapses a generation downstream. The bubble is the fuel. Newport points to the economics: LLM inference is the most expensive computational operation we can do, companies are losing money on every interaction, and the spending, as he puts it, "doesn't make sense." The bubble accelerates every other dynamic because the money flooding in creates pressure to ship products before the consequences are understood.

This is Newport's progress laundering in structural form. Advances in chatbot fluency get credited to the entire AI enterprise, which attracts capital, which funds deployment, which concentrates power, which degrades the information environment, which makes regulation harder, which erodes agency. The chain is self-reinforcing.

Green arrives at this conclusion himself near the end of his video, almost by accident. His stated number-one fear is not any single item on his list. It is the combination: AI intermediating human communication, humans choosing less agency at every decision point, and a few companies controlling the systems. He reaches for the word that ties his fears together and lands on communication. "That is humanity's superpower," he says, "orders of magnitude beyond any other organism." AI is now sitting in the middle of it.

What to Watch

If these fears are connected, they should produce measurable signals. Not predictions. Observable indicators that the chain is tightening or loosening.

Power concentration. Track market share among the top five AI providers. Today, OpenAI, Google, Anthropic, Meta, and xAI control the overwhelming majority of frontier model deployment. Track model diversity: how many competitive frontier models exist, and how many organizations can train them? If both numbers shrink, concentration is accelerating. DeepSeek's release of R1 was a counter-signal, demonstrating that competitive models can emerge from outside the usual circle. Watch whether that pattern repeats or whether it was an exception.

Epistemic collapse. Deepfake detection success rates are the direct metric, and they are falling as generation quality improves. Trust-in-media polling, already at historic lows, is the lagging indicator. The leading indicator is the percentage of online content that is AI-generated, which is rising fast enough that some researchers expect it to constitute a majority of new web content within two years. Green's fear about the "first major election cycles where video evidence is untrustworthy" is testable: track whether deepfake video plays a documented role in upcoming elections.

Regulatory capture. Follow who writes AI regulations. Track the revolving door between AI labs and government advisory roles. Apply Newport's test: do the regulations primarily burden small players or large ones? If compliance costs function as barriers to entry, capture is underway.

Agency erosion. Measure algorithmic versus chronological feed adoption. Track the percentage of web searches intermediated by AI summaries rather than returning direct links. Green's own metric is the simplest: when given the choice between more control and less, do people choose less? If the answer stays yes, the pattern holds.

Apprenticeship pipeline. Junior hiring rates in software, law, finance, and design are the leading indicator. Block's layoff of 4,000 employees, with CEO Jack Dorsey explicitly citing AI, was notable not for its size but for its candor. Most companies will make similar cuts without saying so. Track entry-level job postings as a percentage of total openings. If the ratio drops, the pipeline is closing.

Bubble. Revenue-to-compute-spend ratios at major AI companies are the fundamental metric. Newport observes that LLM inference is the most expensive computational operation available. If costs per query do not fall faster than usage grows, the economics break. Green puts the probability of a bubble pop above 50 percent and the severity at moderate: painful, but not civilization-ending. The question is what it takes with it.

None of the metrics above require sophisticated analysis. They require attention. The reason this section exists is that most commentary on Green's video will offer opinions about which fears are valid. Tracking observable signals is harder and more useful.

The Off-Ramps

Newport's concept of intentional AI offers the clearest alternative to the current trajectory. He points to CICERO, Meta's diplomacy-playing AI, and Pluribus, its poker-playing predecessor, as examples of systems designed with human-written control modules governing what the neural network can and cannot do. CICERO is architecturally constrained against deception in negotiations, not because it was trained to be honest, but because the design penalizes dishonest communication in the search space. The constraints were imperfect, but the principle is what matters. This is a different design philosophy than building a general-purpose model and hoping alignment training holds. Newport calls it intentional AI. It is specialized, constrained by design, and verifiable.

Techno-selectionism is the individual-level version of the same idea. Green and Newport both note that adoption is not inevitable. People and institutions can choose not to use AI tools when the costs outweigh the benefits. This sounds obvious, but it cuts against a technology culture that treats adoption as a default and resistance as ignorance. Cal Newport has built a career on the premise that saying no to tools is itself a skill, and the AI transition may be the highest-stakes application of that principle yet.

Liability is the policy lever. If companies were legally liable for the outputs of their chatbots, the general-purpose conversational model becomes economically unviable overnight. The product that tries to do everything, from homework help to medical advice to emotional support, cannot survive contact with a liability regime that holds the provider accountable for the results. Newport argues this would be a net positive: it would force the industry toward narrow, verifiable, intentional tools and away from the monolithic language model that tries to be all things to all people. The legal infrastructure for this already exists in product liability law. What does not exist is the political will to apply it to software outputs. That gap is itself a measure of regulatory capture.

None of these ideas are new. Newport has been writing about intentional technology use for a decade. Product liability has existed for a century. The question is not whether off-ramps exist. It is whether anyone takes them before the on-ramp runs out of exits.

Green closes his video where he started: with communication. It is, he argues, humanity's defining capability, "orders of magnitude beyond any other organism." AI is now intermediating it, with extreme concentration of power, in the hands of a few companies and a few people he has no particular reason to trust. The only question left is who decides how.