AI system prompts are now mostly AI-generated: suggestive evidence for RSI?

Disclaimer: this is a vibe-coded analysis written up by Opus 5.5, so dive into the slop only if you're really interested in this:

BLUF: In spring 2026, the share of newly written system-prompt text that is AI-authored jumped from about half to about 90% at Anthropic, and from about a quarter to about 95% at OpenAI. That includes the safety rules. The text that remains human-written is mostly terse interface reference, product facts, legacy passages not yet rewritten and, at OpenAI, agent permission and sandbox rules.

• Method: ~220k words of first-appearance text (only what each version adds) from official and leaked Anthropic and OpenAI prompts, June 2024 to September 2026, scored with the Pangram detector and checked against known human and AI text in the same genre.
• Anthropic: 45% → 89%. The break falls between Opus 4.6 and Opus 4.7. A check that needs no detector agrees: zero em-dashes in new text across all 18 official versions before April, then they appear.
• OpenAI: 23% in 2025 → 93–95% in the 2026 Codex app prompts.
• Safety rules: 46→82% at Anthropic, 29→96% at OpenAI.

Code, detector labels, figures and a draft paper: github.com/HaukeHillebrandt/sysprompt-ai-authorship