Study Reveals 37% 'Semantic Void' Phenomenon in GPT, Claude, and Other LLMs

rayanpal_ · reddit · 2026-07-31

A frozen cross-vendor study across 11 major LLMs (including GPT, Claude, Gemini, and Kimi) revealed a startling rate of empty outputs. Out of 31,430 trials, 11,658 successful executions returned exactly zero visible UTF-8 output bytes, accounting for roughly 37% of the total.

Through 4,290 strictly matched semantic pairs, the researchers confirmed that these "Semantic Voids" are entirely distinct from standard safety refusals, rate limits, or transport failures. Raw records, event hashes, verification code, and full analysis have been made public.

Related event: Massive Tests Reveal High Rate of Zero-Byte Empty Outputs in Major LLMs(3 posts)→

Original post →

More from Models

Models channel →