Audit: AI models alter student essay quotes in 1.18% of cases

A check of 27,462 quotes in feedback from six AI models (LLaMA-3.1 8B, Mistral-7B, Mistral-24B, Qwen2.5-70B, Gemini 2 Flash, LearnLM) on student essays found 1.18% of quotes change the student's words. Per-model rates range from 0.22% for LearnLM to 13.2% for Mistral-24B, so model size does not predict fidelity.

• 1.18% of 27,462 checked quotes change the student's words (95% CI 0.83–1.76%)
• Short 4–6 word quotes are altered 2.29% of the time vs 0.68% for long ones
• LearnLM had 0.22% alteration rate; Mistral-24B was worst at 13.2%
• Few-shot prompting had the highest change rate at 4.32% and quoted least