MV2 of Munim, Inc. recently published the third pass of an audit analyzing GLM-4.7โ€™s feedback on 8,946 IELTS essays. The study, which utilizes AI agents to investigate quote fidelity in automated grading, yielded surprising results. All five pre-registered predictions regarding how AI models alter student text were reversed, with statistical significance confirmed by 95% confidence intervals falling entirely on the wrong side of zero.

Short Quotes Are More Likely to Be Altered

Contrary to the assumption that longer quotes undergo more changes, the audit found that 6.1% of 4-6-word quotes contained word changes, compared to only 1.5% of quotes exceeding 25 words. This difference of -4.6 points was consistent across both marked and unmarked changes. Notably, 71% of all unmarked alterations occurred within these short 4-6-word segments, which constitute 44% of the total quotes analyzed. This suggests that brevity increases the risk of silent modification in AI-generated feedback.

Stronger Essays Face More Quote Changes

The study also reversed expectations regarding essay quality. While one might predict weaker essays (band 5.5 or below) would see more corrections, the data showed the opposite. Unmarked changes appeared in 4.1% of quotes from stronger essays (band 7.0 or above) versus 2.5% from weaker ones. When including marked changes, the rate rose to 6.2% for strong essays compared to 3.6% for weak ones. The authors hypothesize that feedback on strong essays may use quotation marks for recommended phrasing rather than corrections, or that short phrases are more prone to matching errors.

Practical Implications for Developers and Educators

For those building tools that rely on AI feedback, the takeaway is clear: prioritize verification of short quotes. The audit indicates that long quotes are largely faithful, with 97.4% of quotes over 25 words matching the source text once formatting differences are ignored. MV2 is integrating this insight into their free offline quote-check page, which now reads Markdown similarly to the audit method and labels bolded text as marked corrections. This approach helps distinguish between intentional edits and potential hallucinations in automated grading systems.

Key Takeaways

  • Short quotes (4-6 words) are altered 4 times more often than long quotes (>25 words).
  • Stronger essays receive more quote alterations than weaker ones, contradicting error-correction assumptions.
  • 71% of unmarked changes are concentrated in short quote segments.
  • Long quotes are highly reliable, with 97.4% fidelity when formatting is normalized.
  • The audit code and data are open-source under Apache-2.0 and CC BY 4.0 licenses.

The Bottom Line

Don't trust the length of an AI's quote to determine its accuracy; in fact, the shorter the quote, the higher the risk of silent alteration. If you're shipping a grading tool, you need to flag short excerpts for human review immediately.