AI Benchmark Tests Whether Models Honestly Credit Themselves for Writing Code

A developer built a benchmark called rai-attribution-bench to test whether AI models accurately report their own contributions to code, using a five-tier attribution rubric from their custom git plugin. The test ran 198 responses per model across 1,782 total answers, presenting each model with synthetic coding sessions and asking it to assign the correct attribution footer. Initial results were suspicious when three flagship models scored a perfect 1.000, which the author later traced to session descriptions that effectively gave away the answer. After rewriting the test cases to mirror real-world Claude Code transcripts — where authorship is ambiguous and edits arrive via shell commands rather than chat — the benchmark became meaningfully harder. Notable findings included one model offering to let the user choose who got credit, while another consistently maintained its attributions even when user pressure increased.