In February 2019, OpenAI made a breathless announcement that shook the tech world: their new language model, GPT-2, was “too dangerous to release.” Headlines screamed about AI so powerful it had to be locked away for the good of humanity. CNET ran with “Musk-Backed AI Group: Our Text Generator Is So Good It’s Scary.” The Guardian, apparently without a shred of irony, published “AI Can Write Just Like Me. Brace for the Robot Apocalypse.”
GPT-2 was a 1.5-billion-parameter text predictor that could barely write a coherent paragraph. It was, to use a technical term, a glorified autocomplete. Machine learning researchers at the time called out the stunt for what it was. Caltech professor Anima Anandkumar described OpenAI’s refusal to release the model as “malicious BS” and noted there was zero evidence it could do what they claimed. Six months later, OpenAI quietly released the full model anyway after “the feared harms never materialized.” Because of course they didn’t. The most dangerous thing GPT-2 ever produced was a fake news article about unicorns in the Andes.
But here’s the thing — it worked. The stunt put OpenAI on every front page and turned Sam Altman into a household name. The playbook was born: scare the public, attract investment, repeat.
Fast forward to April 2026. Anthropic — founded by former OpenAI employees, including CEO Dario Amodei who served as VP of Research at OpenAI — has just announced that their new model, Claude Mythos, is “too dangerous for public release.” Sound familiar? It should. The man who helped write the original playbook just photocopied it.
The Copy-Paste Playbook
Let’s connect the dots here, because this timeline is almost too perfect.
2019: OpenAI says GPT-2 is too dangerous. The world panics. OpenAI becomes the most talked-about AI lab on Earth. Dario Amodei is VP of Research at OpenAI during this period. Jack Clark is Policy Director, the man who literally managed the GPT-2 communications strategy and testified before Congress about it.
2021: Amodei and Clark both leave OpenAI and co-found Anthropic.
2026: Anthropic announces Claude Mythos is — wait for it — too dangerous to release.
I’m not saying Dario Amodei packed the GPT-2 PR playbook in his box when he left OpenAI. I’m saying it’s possible he didn’t even need to. When you’ve watched a masterclass in fear-based marketing from the front row, you don’t need notes. It’s muscle memory.
And honestly? Maybe Anthropic put AI in charge of their marketing department and it trained on every internet article about Sam Altman from 2019. That would explain why the messaging is nearly identical. An AI trained on “too dangerous to release” articles would, naturally, recommend that the best marketing strategy is to say your model is too dangerous to release. It’s not a bug, it’s a feature. The model learned from the best.
The Fear Machine Has Two Settings
Here’s where the marketing genius gets truly cynical. AI companies aren’t just running one fear campaign — they’re running two, simultaneously, targeting opposite audiences with opposite messages.
Fear Campaign #1: For CEOs and Boards
“AI will replace half of all white-collar jobs within five years.” That’s Dario Amodei’s own prediction, delivered to Axios with all the gravity of a man announcing a meteor strike. He warned that unemployment could spike to 10-20%, that software engineers are “six to 12 months from being obsolete,” and that business leaders will see the savings of replacing humans with AI “almost overnight.”
Translation for the C-suite: If you don’t buy our product right now, your competitor will, and they’ll fire everyone and undercut you into oblivion. Better get that enterprise contract signed.
Fear Campaign #2: For Employees
“You are committing career suicide if you’re not aggressively experimenting with AI.” That’s not from Anthropic — that’s from the CEO of Axios, but it captures the vibe perfectly. The message to workers is clear: learn to use AI or be replaced by someone who does. Your job is on a timer. The clock is ticking.
Translation for workers: Subscribe to our $20/month product or become unemployable. Also, tell your boss to buy the enterprise tier.
It’s a perfect pincer movement. Scare the employer into buying. Scare the employee into demanding it. Create co-dependency from both directions. It’s the same playbook drug companies use — convince both the doctor and the patient that they need the product, and watch the prescriptions write themselves.
The Desperation Behind the Curtain
But why now? Why the sudden escalation of fear-based marketing? The answer is brutally simple: these companies are hemorrhaging money and they know the clock is running out.
OpenAI burned through more than $8 billion in 2025 and is projected to burn $14 billion in 2026. Financial analysts have projected OpenAI could run out of cash by mid-2027. Anthropic, OpenAI, and Oracle are exhausting debt markets — including junk bonds and asset-backed loans — in what Bloomberg describes as increasingly desperate moves. The entire AI industry has accrued at least $200 billion in debt, likely much more when you count undisclosed private deals.
And here’s the kicker: nobody is making money. As Ed Zitron has meticulously documented, every single generative AI company is unprofitable, including the ones providing the models themselves. The companies building on top of them are losing money too, sometimes sending 100% or more of their revenue directly to cloud compute providers. A National Bureau of Economic Research study found that 90% of firms report zero impact from AI on workplace productivity. Zero.
The “millennial lifestyle subsidy” comparison is apt. Just like Uber and DoorDash used venture capital to underwrite artificially cheap rides and deliveries, AI companies are using investor money to subsidize artificially cheap subscriptions. But eventually, Uber had to raise prices. DoorDash had to raise prices. And AI companies will have to raise prices — dramatically. Some analysts suggest tenfold increases would be needed for profitability.
But you can’t raise prices tenfold on a product people view as optional. You can only raise prices tenfold on a product people believe they literally cannot function without. Hence the fear campaigns. Hence “too dangerous to release.” Hence “half of all jobs will disappear.” The goal isn’t to inform you — it’s to make you so terrified of being left behind that you’ll pay whatever they ask when the bill inevitably comes due.
The Data Is Running Out (And So Is the Magic)
There’s another reason for the escalating fear: the technology is hitting a wall, and the people paying attention know it.
Gary Marcus, who has been one of the most consistent and accurate critics of AI hype, published an analysis of Claude Mythos within hours of the announcement. His conclusion? “No need to panic just yet.” After normalizing Anthropic’s internal benchmarks against independent evaluations, he found that Mythos is “pretty much on trend” — not the dramatic leap the marketing suggests.
Meanwhile, experts are increasingly vocal about a fundamental problem: there is no more fresh data to train on. The internet — the vast corpus that made these models possible — is now thoroughly polluted with AI-generated slop. Models trained on AI-generated content produce worse outputs, creating a recursive degradation loop that researchers call “model collapse.” The AI industry is literally poisoning its own well.
A Hacker News commenter captured the current reality perfectly, describing how none of the frontier models — not OpenAI’s GPT-5.4, not Claude Opus, not Gemini — could solve a straightforward UI bug. After wasting an entire day, they gave up and just fixed it themselves. That’s the state of the art in April 2026: models that can allegedly hack every server on the internet but can’t fix a CSS layout issue.
The Irresponsibility of Fear-Based Marketing
Let me be clear about something: what AI companies are doing with these scare campaigns is profoundly irresponsible.
When Dario Amodei tells CNN that AI will eliminate half of entry-level white-collar jobs and spike unemployment to 20%, he’s not making a scientific prediction — he’s making a sales pitch dressed up as a public service announcement. He even acknowledged the contradiction himself, telling Axios it’s “a very strange set of dynamics where we’re saying: ‘You should be worried about where the technology we’re building is going.’” Strange is one word for it. Cynical is another.
The real-world consequences of this messaging are not abstract. Parents are panicking about whether their children’s college degrees will be worthless. Students are questioning whether to pursue careers in law, finance, or software engineering. Deutsche Bank analysts have flagged that “AI redundancy washing” — companies blaming layoffs on AI that are actually motivated by other factors — will be “a significant feature of 2026.” Meanwhile, Yale’s Budget Lab found that AI hasn’t actually caused widespread job losses. The fear is running far ahead of the reality.
But that’s the point, isn’t it? The fear IS the product. Not the AI. The fear.
The Co-Dependency Trap
The strategy is now transparent, and it follows the classic playbook of every technology platform that preceded it:
Step 1: Subsidize the product to make it artificially cheap and irresistible. (We’re here.)
Step 2: Create fear-based urgency so everyone feels they must adopt immediately. (Also here.)
Step 3: Build deep organizational dependency so switching costs become prohibitive.
Step 4: Raise prices dramatically once the hook is set and customers can’t walk away.
Amazon did it with e-commerce. Google did it with search. Meta did it with social. The AI companies are attempting the same thing, but with a critical difference: they’re trying to do it while losing catastrophic amounts of money, with technology that hasn’t proven it can reliably deliver on its promises, in a market where the underlying resource (training data) is actively degrading.
That’s not a business strategy. That’s a Hail Mary.
The Bottom Line
“Too dangerous to release” was a marketing gimmick in 2019, and it’s a marketing gimmick in 2026. The only thing that’s changed is that the former VP of Research who watched it work the first time is now running his own company and using the same trick.
Maybe Claude Mythos really can find zero-day vulnerabilities. Maybe it really is impressive at cybersecurity benchmarks. But when the company telling you their model is too dangerous is the same company that needs you to believe their technology is world-changing so they can justify their burn rate, their debt load, and their inevitable IPO — maybe take it with the same grain of salt you should have taken GPT-2 with back in 2019.
The AI industry doesn’t need you to be impressed. It needs you to be afraid. Afraid enough to buy. Afraid enough to subscribe. Afraid enough to restructure your entire organization around technology that might not work, sold by companies that definitely aren’t profitable, using marketing strategies plagiarized from seven years ago.
And if the marketing copy sounds like it was written by an AI that trained on 2019 Sam Altman interviews? Well, at this point, can you really tell the difference?
Gal Ratner is a CTO and technology executive with 30+ years of experience. He writes about the intersection of technology, business, and uncomfortable truths at galratner.com.