Blinded expert annotators misclassified 83.7% of LLM-generated errors as human-written¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? In a blinded source-attribution task on 196 synthetic submissions, annotators misclassified 164/196 (83.7%) as human-written, indicating low distinguishability. → Ali Keramati 2026

Evidence¶

Ali Keramati 2026¶

Ali Keramati, Jie Cao, Iman Mohammadi, Mark Warschauer, Yang Shi. (2026). Simulating Students' Java Programming Errors with Large Language Models. arXiv preprint. https://arxiv.org/abs/2606.14113 doi:10.48550/arxiv.2606.14113

q2 · i? · design · r2

Blinded expert annotation study (N=401 single-coded items after a 43-item double-coded pilot) comparing authentic and synthetic errors from the best-performing setting (Claude Sonnet 4 with Self-Refine). The deception rate was "164/196 (83.7%)" on the synthetic items, defined only on those 196 items.

"Annotators misclassified 164/196 (83.7%) of LLM-generated submissions as human-written, indicating a low level of distinguishability"

Discussion¶

Related Claims¶

• LLM-generated errors received significantly higher plausibility ratings than authentic student errors, explained by a narrower, cleaner error-type distribution — related
• Blind expert verification shows no overall preference for human over LLM qualitative coding when sources are judged symmetrically — related
• LLM and human written 7C analyses show no overall difference in behavioral alignment or evidence correspondence, but align less on Communication and Constructive dimensions — related
• LLM-generated erroneous code aligns with authentic student errors at levels that depend on model and prompting strategy, with Claude Sonnet 4 most consistently aligned — related