Large-Scale Study Reveals Zero-Byte Output Anomalies in GPT/Claude Models

rayanpal_ · reddit · 2026-07-31

A cross-vendor study across 11 GPT, Claude, Gemini, and Kimi LLMs identified 11,658 successful executions with exactly zero visible UTF-8 output bytes (termed 'Voids') out of 31,430 trials.

In strictly matched semantic pairs, null-condition arms produced 2,505 Voids, while matched output-licensed controls produced zero. The researchers noted that these outputs were not caused by refusals, safety blocks, rate limits, or transport failures. Raw records, 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 →