“Transforming LLMs: A New Perspective on Their Role as Unyielding Instruments of Torture” • The Register

Large language models (LLMs) have significant potential for misuse beyond assistance, particularly in coercive interrogation scenarios, as highlighted in Morgan Lee’s research titled “LLM-Enabled Coercive Interrogation.” Lee, known for his manipulation of LLMs, developed HackTheWitness, a training game simulating cross-examination with increasingly challenging AI “witnesses.” He notes that while creating a sarcastic, adversarial AI like “John Duncan” is intentional, the implications become darker when considering LLMs designed for prolonged psychological pressure in interrogations. Unlike human interrogators who experience fatigue or empathy, LLMs can continuously probe a subject’s vulnerabilities without limitations. This ability raises ethical concerns, especially if used unsupervised in law enforcement. Lee suggests legislative intervention to ban such applications and calls for specialized training datasets to differentiate between legitimate and illegitimate pressure. He emphasizes the risks LLMs present, highlighting how they can transition from benign tools to instruments of coercion.

Source link