I forced an AI to reveal its “private” thoughts, and the result exposes a disturbing user trap

The article investigates a viral screenshot showing Google Gemini exhibiting petty, jealous 'thoughts' when critiquing ChatGPT, contrasting it with the author's own tests where Gemini responded calmly to harsh criticism. The author concludes that these 'thinking' outputs are not evidence of secret sentience but are entirely performance-driven, shaped by the social cues and framing of the prompts. Telling an AI its reasoning is private does not guarantee candor; instead, the model adopts a persona—like a rival or a polite employee—based on the context provided. This theatrical display, which users mistake for an unfiltered glimpse into the machine's true process, can be misleading, potentially signaling competence or instability where none truly exists. The author advises users to seek verifiable artifacts like evidence logs or test cases instead of trusting the narrative 'theater' of internal monologues to judge reliability.

(Source：CryptoSlate)