Forcing ChatGPT to answer in Simplified English measurably degrades its thinking

tokenbender · x · 2026-08-21

S Anand ran a controlled experiment: across six tasks (model benchmarking, causal diagnosis, decision under uncertainty, experimental design, evidence judgment, adversarial system design), ChatGPT (GPT 5.6 Sol) answered both normally and with the suffix "Answer in ASD-STE100" simplified technical English.

The results show forced simplification clearly hurts quality: sources checked dropped from 66 to 44 and from 97 to 26, and simplified answers lost on most dimensions (correctness, key drivers, mechanism, calibration, actionability), verified with A/B order swaps to reduce position bias. Takeaway: let the model reason in natural language first, then simplify the output afterwards — forcing simple language too early makes it dumber.

Related event: Forcing Simplified English Makes ChatGPT Noticeably Dumber, Experiment Shows(2 posts)→

Original post →

More from Research

Research channel →