AICD Bench - Fine-Grained Human-Machine Classification (Zero-Shot CoT): leaderboard

Fine-grained authorship of code from AICD Bench (Orel et al., 2026): an LLM prompted zero-shot with chain-of-thought and a summary of the code features of each class labels a code sample human-written, machine-generated, hybrid (human code rewritten or completed by an LLM) or adversarial (generated to look human-written). Macro-F1 in percent over the four classes on a 5% sample stratified by label, domain and language; higher is better.

Metric: Macro-F1 (%; four-way human-written, machine-generated, hybrid and adversarial code; 5% sample stratified by label, domain and language; zero-shot chain-of-thought prompt with per-class code features). Source: arxiv.org. Saturation forecast: Around March 2027. 2 models tracked.

Top models

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=aicd-bench-fine-grained-human-machine-classification-zero-shot-cot · How It Works · Data refreshed daily, snapshot 2026-11-12.