Dev Criticizes LLM Surveys: Models Are Text Samplers, Not People
gerardsans · x · 2026-10-26
Developer Gerard Sans argues that using frontier AI models for surveys is "straightforwardly dangerous": their training data comes from the already low-quality internet, further degraded by fiction, movies, and marketing copy, plus a heavy skew toward commercially motivated technical material. Using this as a survey base introduces unknown variables that make reliable forecasting or surveying impossible—"playing data roulette with your eyes closed," confusing textual fluency with scientific validity.
In a follow-up, he stresses AI literacy: research should stop casual anthropomorphism. AI is a text sampler, not a person; it has no psychology, only data-compression artefacts—compressing your inbox doesn't yield a model of your psyche.
More from AGI Musings
- Posting has power-crept into a basic skill; attention is the only scarce resource — tokenbender · 2026-10-26
- Dunking on those who believe in AI doom yet trust hardware verification plans — jessi_cata · 2026-10-26
- AI adoption up to 78% but developer trust down to 33%, Stack Overflow surveys show — agh_murad · 2026-10-26
- Nation-state vulns now have shorter half-lives, raising crisis instability — jeremiecharris · 2026-10-26
- Sci-fi plot nearing reality: AI tunes film scenes to viewers' biometric reactions — IgorBrigadir · 2026-10-26
- Professor catches student cheating with AI who left the prompts in the paper — venturetwins · 2026-10-26