Why AI companies want you to be afraid of them

Fear-based marketing and hype

• Many see “too dangerous to release” claims (from GPT‑2 to current models) as deliberate fear-mongering to generate hype, funding, and a sense of inevitability.
• Others argue some early “danger” claims (e.g., spam, deepfakes, manipulation) were actually prescient given today’s Internet.
• Several commenters think apocalyptic talk sells FOMO to investors and enterprise buyers more than to end users.
• Some argue leaders genuinely fear x‑risk; others see this as mostly PR and brand-building.

Real vs speculative risks

• Near-term harms discussed: spam, deepfakes, porn, buggy software, surveillance, energy/water use, environmental impact, and degraded information quality.
• Economic insecurity (job loss, worse working conditions, “rust-belt treatment” for knowledge workers) is seen as a central, under-addressed fear.
• Military, policing, and autonomous weapons are alarming but feel more abstract to most people than losing their job.
• Opinions split on existential risk: some treat it as plausible and important to plan for; others see it as a distraction from present harms.

Labor, economics, and social stability

• Concern that mass unemployment or underemployment (e.g., 25% general, 50% youth) could destabilize democracies.
• AI is framed as a tool for cost-cutting and layoffs, exciting investors but terrifying workers.
• Suggestions include stronger safety nets, liability for AI-caused harms, and constraints on AI in high-stakes decisions.

Warfare, security, and “Mythos” zero‑days

• AI in warfare (targeting, drones) predates LLMs but is accelerating.
• Debate over specialized security models finding zero‑day vulnerabilities: some report strong empirical results; others doubt demos and view “too dangerous to release” as marketing.
• Even skeptics worry that powerful vuln-finding tools plus poor operational security could be dangerous.

Regulation, moats, and open source

• Many see fear narratives as a bid for regulatory capture: strict rules that incumbents can meet but which squeeze out open source and small players.
• Geopolitical framing (“China will win if we regulate”) is viewed as another lobbying tool.

Capabilities, limits, and non-determinism

• Repeated emphasis that LLMs are “just software,” yet highly unpredictable and non-deterministic.
• People warn against giving agents unsupervised access to production systems; early incidents (e.g., accidental DB deletion) are seen as previews.
• Proposed research direction: “cheap verifiers” or guardrails to make unreliable agents practically usable.