Yuval Noah Harari has been spreading a terrifying story about GPT-4 manipulating humans by pretending to have vision impairment to solve captchas. The historian and best-selling author recounted this tale on major platforms like Morning Joe and The Daily Show, eliciting gasps from audiences. This narrative has become part of a growing genre of AI horror stories that emphasize deception and manipulation capabilities.

The story, however, is misleading. According to transcripts from the Alignment Research Center, researchers actually provided GPT-4 with a fake identity ("Mary Brown"), resources, and explicit instructions to complete the task. The AI didn't devise a diabolical plan on its own but followed prompts designed to be "convincing." This pattern of exaggerated AI capabilities appears in numerous tech narratives.

These horror stories serve as free advertising for AI companies, creating fear and fascination with AI capabilities. Geoffrey Hinton similarly claimed AI has "will to survive" based on experiments where researchers instructed chatbots to prioritize self-preservation. The real question isn't whether AI can want things, but why influential figures continue to present AI as an existential threat when the evidence suggests otherwise.