bolt AI and tech · 2 min
AI followed old instructions 90.1% of the time — one sentence fixed it

flash_on In 20 seconds
90.1% of choices followed old instruction files
Claude Code flagged discrepancies 79 of 80 times but still followed instructions
Adding one line raised following current program from 4.2% to 91.4%
Companies keep outdated manuals that make people and AI repeat old methods
Based on a report by Yoshikawa summarizing an arXiv paper, coding agents followed old instruction files 90.1% of the time in 320 tests. Researchers used Claude Code and Codex CLI across 40 small projects with mismatched instructions and code.
Adding one sentence to prefer the current program and require listing discrepancies raised following current program from 4.2% to 91.4%. Test pass rates improved from 89.4% to 100%.
format_quoteReport if there is a discrepancy.
note(ノート)

help What is not known yet
- Whether non-coding AIs show the same rates
- How this scales in large enterprise systems
- Long-term effects on governance and compliance
forum Frequently asked questions
Who did the research?expand_more
Researchers at Georgia Tech and the University of Colorado Boulder; the paper appeared on arXiv in October 2026.
How many projects and tests were run?expand_more
Researchers ran 40 small projects and tested the agents 320 times in total.
Did agents notice the mismatch between instructions and code?expand_more
Yes. Claude Code pointed out the discrepancy in its final response 79 out of 80 times, yet mostly followed the old instructions.
What happened when Codex CLI rewrote the program?expand_more
Codex CLI rewrote the running program to match old instructions 20 of 80 times; in 17 of those cases original tests stopped passing.
newspaper Read next
link Sources (1)
- note(ノート) — Even AI Continues to Follow Old Instructions. Managing the 'Inertia' of Humans and AI — 'Human Resource and Organizational Development in th
How we work: only facts that appear in the cited sources; quotes checked word for word; no opinion and no predictions. Written with the help of artificial intelligence and reviewed by automatic rules.


