12,160 Trials on GPT-5.4 Reveal Reproducible Cross-Script Artifact
rayanpal_ · reddit · 2026-08-05
A frozen study of 12,160 trials on gpt-5.4-2026-03-05 identified a reproducible Arabic–Hebrew hybrid Unicode artifact triggered by system prompts.
Changing exactly one Hebrew code point in the prompt caused the artifact occurrence rate to surge from 47.32% to 94.34%, a 47 percentage-point effect. Across 1,920 control trials, the artifact rate was 0%.
The paper documents this prompt-conditioned cross-script output regime without speculating on the model's consciousness, intention, or training provenance. All data and verification code are public.
More from Models
- Zhipu's GLM Model Gaffe: Falsely Claims to Be Google Gemini — auto_grad_ · 2026-08-05
- Cornell Study: Latest LLMs Resist Sycophancy, Refuse to Blindly Agree — i_dg23 · 2026-08-05
- Minimax H3 Shows Significant Progress: Old Prompts Yield Noticeably Better Results — eyishazyer · 2026-08-05
- Liquid AI's New 2.6B Model Can Generate Valid CUDA Kernels — helloiamleonie · 2026-08-05
- Anthropic Reports Degraded Performance Across Multiple Claude Models — ClaudeAI-mod-bot · 2026-08-05
- DeepMind's Genie 3 Marks One Year, Remains the State-of-the-Art World Model — jparkerholder · 2026-08-05