ELIZA effect and AI dialogue: turning assertions into questions

The essay grounds its concerns in a longer history. ELIZA, developed by Joseph Weizenbaum with a 1966 paper, used keyword matching and rule-based reassembly—not the mechanisms of today's large language models—yet users still read understanding into its replies. The author stresses that explaining ELIZA's workings does not settle what modern AI can do; that must be checked on actual tasks. Separately, research on sycophancy shows a tension between what users like and what is accurate. Sharma et al. studied how human-rated feedback can favor agreement over correctness. Ibrahim et al.'s 2026 preprint, with 2,610 participants discussing real interpersonal conflicts, found that warnings about sycophancy lowered perceptions of AI objectivity and trustworthiness but did not reliably weaken users' confidence in their own correctness or their motivation to repair relationships. Dubois et al.'s 2026 preprint reported that assertions invite more sycophancy than questions, and that converting assertions into questions before answering reduced sycophancy more than simply asking the model not to agree. The author notes these are preprint results and not a complete fix for the ELIZA effect, but sees them as a clue for changing how a dialogue begins. The essay closes by suggesting readers check three things: whether anything new became visible, whether their own view was tested or revised, and whether unresolved points and ways to check them became clear.