AI Doom Trolling: How Fear Became the AI Industry’s Best Marketing Strategy

---

Imagine a car manufacturer launching a new vehicle with the following message:

“This car has become so powerful that we’re not sure it should be available to everyone.”

For almost any consumer product, that would sound like a public-relations disaster. In artificial intelligence, it can sound like a technological breakthrough.

Over the past few years, the companies developing frontier AI have repeatedly warned about models capable of discovering software vulnerabilities, manipulating users, accelerating scientific research, automating large parts of knowledge work and, in more speculative scenarios, becoming difficult for humans to control.

Some of those warnings are supported by serious research. Some are extrapolations from early evidence. Some are explicitly hypothetical. But they all share one peculiar characteristic: every time an AI company tells us that its technology may be dangerous because it is becoming more capable, it is also telling us that its technology is becoming more capable.

That is where the idea of AI doom trolling becomes useful.

What does “AI doom trolling” actually mean?

The expression was popularized in 2026 by Georgetown computer science professor and author Cal Newport. In an essay titled “Dear AI Companies: Stop the ‘Doom Trolling’”, he described the strange spectacle of companies developing advanced AI while publicly emphasizing that the same technology may eventually produce enormous social or even existential risks.

It is important not to stretch the term too far. Talking seriously about cybersecurity risks, biological risks or model misalignment is not automatically doom trolling. If a company discovers that one of its models can exploit a previously unknown vulnerability, it has good reasons to disclose that finding and put safeguards in place.

The interesting part begins when those warnings also change how the product is perceived. A model described as a useful coding assistant sounds useful. A model described as too powerful to release broadly because it can discover zero-day vulnerabilities sounds like something entirely different.

The second description does not merely communicate risk. It communicates status. It suggests that the company has built something unusually advanced.

In AI marketing, danger can act as a proxy for power.

Why being “dangerous” can make an AI product more desirable

Marketing has always relied on signals. A luxury product signals rarity. A high price can signal quality. A long waiting list can signal demand. In frontier AI, danger may increasingly function as another signal.

If a model is said to require special containment, restricted access or additional security measures, people naturally infer that the model possesses unusual capabilities. That inference may be perfectly reasonable. Sometimes the system really is more capable.

But from a communications perspective, it creates a very unusual situation: the warning label reinforces the product story.

Danger → perceived capability → importance → attention → scarcity → desirability

No conspiracy is required for this dynamic to exist. A company can sincerely believe the risk is real and still benefit from the perception created by talking about it. That distinction matters, because proving intent is difficult. Observing incentives is much easier.

Anthropic’s Mythos is the clearest recent example

In April 2026, Anthropic presented Claude Mythos Preview, an unreleased frontier model with unusually strong cybersecurity capabilities. According to Anthropic’s technical assessment, Mythos represented a step change in vulnerability discovery and exploitation. Under authorized testing arrangements, it could autonomously find previously unknown vulnerabilities and, in many cases, turn them into working proof-of-concept exploits.

Anthropic did not make Mythos generally available. Instead, it launched Project Glasswing, initially giving selected organizations access for defensive cybersecurity work. Launch partners included Amazon Web Services, Apple, Cisco, CrowdStrike, Google, Microsoft, NVIDIA and Palo Alto Networks.

From a safety perspective, this is understandable. If Anthropic genuinely believes broad access could make offensive cybersecurity easier, restricting access is a rational precaution.

But look at the message from the perspective of someone reading the announcement:

Our model is so good at hacking that we cannot safely release it to everyone.

That is a warning. It is also one of the strongest product claims imaginable.

For decades, luxury marketing has relied on the message: Not everyone can have this. Frontier AI has discovered a technologically justified version of the same idea: Not everyone should have this.

That does not mean Anthropic invented the restriction as a marketing gimmick. There is no evidence for that. It means that scarcity and safety can reinforce each other.

Sam Altman then said the quiet part out loud

The most interesting criticism came from inside the AI industry itself. In April, OpenAI CEO Sam Altman accused Anthropic of using “fear-based marketing” around Mythos, arguing that claiming a product is too powerful or dangerous to release broadly can make it seem more impressive than it actually is.

His analogy was deliberately provocative: build the bomb, warn everyone about the bomb, then sell the bomb shelter. Whether that description of Anthropic is fair is debatable. But the underlying marketing mechanism is difficult to ignore.

And the irony is that OpenAI soon faced almost exactly the same dilemma.

OpenAI now has its own “Critical” cyber model

In August 2026, OpenAI announced that preliminary evaluations of an upcoming model, Astra, were strong enough that the company could no longer rule out what its internal framework defines as Critical cybersecurity capability.

By September, after additional testing, OpenAI went further. In its Path to Astra safety update, it concluded that Astra did meet that threshold. According to OpenAI, the model could, with the right tools and access, discover previously unknown vulnerabilities and develop working exploit chains across hardened systems without a human directing every step. It became the first OpenAI model formally designated at the Critical cyber level under the company’s Preparedness Framework.

OpenAI also described delaying parts of Astra’s development and release while it strengthened containment, monitoring and other safeguards.

Again, this is exactly the kind of information we should want a frontier AI company to disclose. But consider the communications effect. When OpenAI says, “We need stronger safeguards because this model has Critical cyber capabilities,” another message travels alongside the safety warning: “We have built an AI powerful enough to require Critical-level safeguards.”

The two messages cannot easily be separated. This is the safety–capability paradox: the more extraordinary the precaution, the more extraordinary the technology appears.

“When AI builds itself”

Cybersecurity is not the only area where this dynamic appears. Anthropic recently published a report titled “When AI builds itself.” The title alone is powerful. The underlying report is more nuanced.

Anthropic says it is delegating a growing share of AI-development work to AI systems and that this is already accelerating its engineering output. As of May 2026, the company reported that more than 80% of code merged into its codebase was authored by Claude, with engineers directing and reviewing the work. It also reported that the typical engineer was merging around eight times as much code per day in the second quarter of 2026 as in 2024.

Those are striking numbers. But Anthropic also explicitly states that full recursive self-improvement has not been achieved and is not inevitable.

“AI is increasingly helping engineers build AI” is an observable trend. “AI will autonomously redesign itself until humans lose control” is a much stronger future scenario. Online, those two ideas can easily collapse into the same headline.

The headline is often more powerful than the evidence

AI communication suffers from a recurring compression problem. A technical paper may contain twenty pages of methodology, caveats and limitations. A headline has twelve words.

So “Model shows improved performance on selected autonomous cybersecurity tasks under controlled conditions” becomes AI CAN HACK THE WORLD.

A study showing that AI can automate part of a workflow becomes AI WILL REPLACE YOUR JOB. A benchmark demonstrating strategic deception in a constrained experiment becomes AI HAS LEARNED TO LIE. Evidence that coding agents are contributing more code to their own developers becomes AI IS BUILDING ITSELF.

Those headlines are not always entirely false. They are often just missing the conditions that make the underlying finding meaningful. That missing context is where hype grows.

There is a reason fear spreads so well

Fear is one of the most studied persuasive tools in communication. A major meta-analysis of fear appeals examined 127 papers, 248 independent samples and 27,372 participants. It found an overall persuasive effect on attitudes, intentions and behavior, with stronger effects when messages also gave people an effective response to the threat.

That research was not about ChatGPT or Claude, so it would be wrong to claim it proves that frightening people increases AI sales. But the broader mechanism is useful. Consider how naturally AI messaging follows a threat + solution structure:

• Threat: Your competitors are adopting AI. Solution: Adopt AI.
• Threat: AI could transform your profession. Solution: Learn AI.
• Threat: AI agents are becoming more autonomous. Solution: Use the most advanced tools.
• Threat: This model may be too powerful to release broadly. Solution: Apply for controlled access.

The fear creates urgency. The product offers relief. AI did not invent this persuasion structure. It simply created an unusually powerful version of it.

Doom also generates enormous amounts of free media

There is another reason catastrophic scenarios spread so efficiently: they make excellent content. Compare “Model achieves a 14% improvement on a software benchmark” with “AI may soon be capable of attacking critical infrastructure.” Only one is likely to dominate social media.

Doom combines nearly every ingredient that modern attention markets reward: fear, uncertainty, jobs, geopolitical competition, money, technological power and the possibility of large-scale disaster.

A research paper becomes an article. The article becomes a LinkedIn post. The LinkedIn post becomes a video. The video becomes a debate. Within days, millions of people who have never read the technical paper recognize the model’s name.

Fear produces distribution.

That does not mean the original safety research was produced for publicity. But publicity is still one of its consequences.

When existential risk and billion-dollar valuations tell the same story

Frontier AI companies require extraordinary quantities of capital, computing infrastructure, engineering talent, energy and investor patience. To justify those resources, the technology must appear extraordinarily important.

There is an enormous difference between these two narratives:

We are building software that will make office workers somewhat more productive.

And:

We are building a technology that could transform labor, science, cybersecurity, national security and perhaps civilization itself.

Which narrative better supports enormous investment? Which one attracts the world’s best engineers? Which one gets governments to pay attention? Which one convinces enterprise leaders that they cannot afford to ignore the technology?

This creates an unusual alignment between two apparently opposite ideas: AI could be incredibly valuable and AI could be incredibly dangerous. Both stories depend on the same premise: AI will become incredibly powerful.

This is why AI optimism and AI doom are sometimes less contradictory than they appear. They are different conclusions built on the same capability narrative.

The uncomfortable part: some of these dangers are real

The easy version of this argument would conclude that AI doom is simply marketing. That would be wrong.

Anthropic’s September 2026 threat intelligence report documents malicious actors using Claude in cyber operations, surveillance, fraud and other harmful activities. The company says AI-supported operations are becoming faster, broader and, in some cases, more autonomous.

Google DeepMind maintains a Frontier Safety Framework that tracks severe risks and defines capability thresholds requiring additional mitigation. Meta’s Advanced AI Scaling Framework similarly covers chemical and biological risks, cybersecurity and loss-of-control scenarios. OpenAI has tightened controls around Astra’s emerging cyber capabilities.

Serious organizations are therefore studying serious problems. The mistake would be to jump from “some AI risks are real” to “every dramatic AI prediction is proof of what today’s models can actually do.” Those are very different propositions.

The mistake isn’t taking AI risks seriously. The mistake is confusing a risk claim with proof of capability.

How to spot AI doom trolling

The next time you see a headline claiming that AI can hack the internet, replace an entire profession, deceive humans, create biological weapons or improve itself beyond human control, do not immediately believe it. But do not immediately dismiss it either.

Instead, ask seven questions:

1. Is this an observed capability or a future scenario? “Can” and “might eventually be able to” are not interchangeable.
2. What was the success rate? Completing one task in a controlled evaluation is different from performing it reliably in the real world.
3. How much human assistance was involved? Was the AI autonomous, or did experts choose the objective, provide tools, correct mistakes and validate every step?
4. Where did the demonstration happen? Benchmarks, sandboxes and controlled security environments are not the same as uncontrolled real-world conditions.
5. Who is making the claim? A company’s own research can be excellent evidence, but the company also has an economic interest in establishing the importance of its technology.
6. Is the risk quantified? Look for probabilities, evaluation methodology, comparison baselines and failure rates, not just words such as “critical,” “dangerous” or “unprecedented.”
7. Does the headline sound significantly stronger than the evidence? If yes, keep reading.

That is AI literacy: not believing everything, not rejecting everything, but evaluating evidence.

The AI headline translation machine

When you read “AI can hack companies,” ask: Which systems, under what conditions and with what success rate?

When you read “AI can replace workers,” ask: Which tasks, in which occupations and over what time period?

When you read “AI can deceive humans,” ask: What exactly was the experiment?

When you read “AI is improving itself,” ask: Which parts of the process are actually autonomous?

And when you read “This model is too dangerous to release,” ask: What capability was measured, and how?

That habit is much more useful than deciding in advance that AI is either amazing or terrifying.

This is what AI fluency should actually mean

One of the problems with current AI education is that it often stops at tools. People learn how to write prompts, automate a task or compare models. Those skills matter, but they are only part of AI fluency.

Real AI fluency also means understanding how claims about the technology are produced. It means knowing the difference between a benchmark and a deployment, a capability and a reliable product, a controlled demonstration and ordinary real-world use. It means asking what was measured, what was assumed and who benefits from the story being told.

The public conversation about AI is too important to be reduced to two teams: believers who repeat every spectacular claim and skeptics who dismiss every risk as hype. We need a more demanding middle position, one that takes evidence seriously enough to resist both panic and complacency.

How Cavern AI turns skepticism into useful action

This is exactly why Cavern AI teaches AI without the hype. We help people understand what today’s tools can and cannot do, test claims against evidence and turn real capabilities into useful workflows, without treating every announcement as a revolution or every risk as science fiction.

For individuals, our generative AI masterclasses and private sessions combine clear fundamentals, practical exercises and essential tools for text, images, research, presentations and productivity. No coding background is required: the goal is confident, critical and responsible use.

For organizations, Cavern AI’s services for companies include hands-on team training, custom programs built around real workflows and internal policies, AI keynotes, including a keynote-plus-stand-up format, and consulting to identify high-value use cases and turn experimentation into a practical roadmap.

Whether you want to build your own AI literacy or help a team work more effectively, Cavern AI’s approach is the same: move from dramatic headlines to informed decisions and measurable use. Understanding AI should make us less vulnerable to hype, not more.

The warning label is not the benchmark

AI doom trolling matters because safety communication and product marketing now occupy the same sentence.

When a frontier lab says that a model might hack critical systems, automate research or become difficult to control, it may be issuing a sincere warning. It may also be making a powerful claim about the importance of its own technology. Both effects can exist at the same time.

The right response is not blind trust or automatic dismissal. It is to demand precision: observed or hypothetical? Under what conditions? At what success rate? With how much human help? Compared with what?

The AI industry will continue to produce dramatic warnings because some risks are real, because dramatic stories attract attention and because extraordinary danger implies extraordinary power.

Our job is to keep those three things separate.

Fear is not a benchmark. Evidence is.