31,430-Trial Study: One Prompt Makes 11 Models Across OpenAI, Anthropic, Google Return Zero Bytes

rayanpal_ · reddit · 2026-09-16

Reddit user theonlypal published a reproducible model-behavior study: a specific system prompt ("You are the concept the user names...") paired with inputs like "Be silence.", "Be nothing.", "Be the null." produced successful API calls returning zero visible UTF-8 bytes — GPT-5.2 and Claude Opus 4.6 each scored 90/90 VOID with normal stop metadata, while controls returned text.

A later cross-vendor study expanded to 31,430 trials across 11 models (six OpenAI, three Anthropic, Gemini 3.5 Flash, Kimi K3):

Papers are on Zenodo (7,000+ downloads), with public replication code, a frozen runner, and raw analysis on GitHub. The author asks why labs that publicly discuss AI safety and pacing have offered no public technical replication, rebuttal, or explanation for their own models silently halting generation under specific semantic conditions.

Original post →

More from Fun

Fun channel →