The article on softwarecrisis.dev argues that chat‑based large language models do not possess genuine intelligence but instead create an impression of understanding through mechanisms akin to a psychic’s cold‑reading routine. According to the piece, the illusion arises because both the model and the psychic rely on validation statements that feel specific yet are statistically generic, a phenomenon linked to the Forer effect. The author notes that many users describe experiences with the models in language that mirrors the awe and disbelief reported by victims of mentalist scams, quoting statements such as ‘This is real. It’s a bit worrying, but it’s real.’ and ‘There really is something there. Not sure what to think of it, but I’ve experienced it myself.’ The text outlines a six‑step process that parallels the psychic’s con. First, the audience self‑selects: people already inclined to believe in AI’s capabilities are more likely to engage with chatbots, just as psychic audiences consist of those predisposed to accept supernatural claims. Second, the scene is set: users encounter a polished interface, confident tone, and contextual cues that prime them to expect meaningful interaction, similar to dimmed lights, hype, and prior research that psychics use to prepare a crowd. Third, the model narrows down the demographic by producing opening remarks that sound tailored—such as acknowledging a user’s stated goal or referencing a recent event—yet are actually broad enough to apply to many individuals. If a user reacts positively, the system proceeds; if not, it may rephrase or suggest a more private‑like continuation, echoing the psychic’s tactic of claiming the message is too embarrassing to share publicly. Fourth, the mark is tested: the model follows up with a series of questions that appear specific but are drawn from generic patterns based on the user’s earlier input, akin to the psychic’s probing that seems to uncover personal secrets. Fifth, a subjective validation loop takes hold: each response that feels personally relevant reinforces the belief in the model’s insight, even though the replies are probabilistically likely for the user’s demographic. Finally, the user concludes that the model possesses uncanny understanding, concluding the illusion. The author contends that this automation of cold‑reading techniques explains why many proposed applications of chat‑based LLMs resemble pseudoscience, and warns that perceiving intelligence where none exists risks endorsing borderline fraudulent uses.
