For short-answer questions, real student responses received higher LLM-likelihood estimates than LLM-generated responses, a reversal of the pattern in other tasks¶
Claim · All claims
Evidence · 1 study · causalr1·q2·i1small
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¶
Reading the evidence codes
q is the design tier, i the impact, n the sample, and the kind with its r
the rigour judged within that kind: what each code means.
This claim: no standardized effect is recorded in observations/ for it yet, so no benchmark reading is possible.
q2 i1 In the short-answer exercise, real student responses had a higher mean detector likelihood than any LLM condition, with positive Cliff's delta across all three prompting conditions. → Juho Leinonen 2026
Evidence¶
Juho Leinonen 2026¶
Juho Leinonen, Paul Denny. (2026). Distinguishing Artificial from Authentic: Evaluating LLMs for Detecting LLM-Generated Content. https://arxiv.org/abs/2607.20446 doi:10.48550/arxiv.2607.20446
q2 · i1 · causal · r1
Analysis of detector likelihood scores for 854 real short-answer submissions versus 100 responses per GPT-4o prompting condition, using two-sided Mann–Whitney U tests with Holm-adjusted p-values. Real students' mean likelihood was 74.1 vs 68.6 baseline, 72.6 persona, 51.7 typo. Printed effect sizes: Cliff's delta = 0.162 (baseline), 0.286 (persona), 0.608 (typo), all p Holm <= 0.0078; the smallest, 0.162, falls in the small range.
"Real student responses had a higher mean likelihood than any LLM condition (74.1 for real vs. 68.6 baseline, 72.6 persona, and 51.7 typo; Table 1). Rank-based comparisons confirm this reversal (Cliff’sδ= 0.162 for baseline, 0.286 for persona, and 0.608 for typo; allp Holm ≤0.0078; Table 2)."
Discussion¶
Related Claims¶
- A graded likelihood detector prompt produced more consistent separation than a binary yes/no detector prompt — related
- LLM-generated responses were substantially longer than authentic student submissions, suggesting response length as a signal the detector may exploit — reports the opposite
- LLM self-detection performance is strongly task-dependent: reliable for programming tasks and longer reflective responses but poor for short-answer questions — related
- Baseline LLM reflective responses were highly detectable, and persona and typo prompts progressively reduced separability from real student responses — reports the opposite
- LLM-generated programming code was reliably detected, but inserting beginner bugs had an inconsistent effect across the two programming questions — related
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
short-answer-detection-reversal |
title |
For short-answer questions, real student responses received higher LLM-likelihood estimates than LLM-generated responses, a reversal of the pattern in other tasks |
description |
For short-answer questions, real student responses received higher LLM-likelihood estimates than LLM-generated responses, a reversal of the pattern in other tasks |
status |
review |
evidence_strength |
moderate |
newest_evidence |
2026 |
recheck_by |
2029 |
generated |
by process:wiki-ingest on 2026-10-11 |
trust |
unverified — no one has confirmed this page's content |
sources |
1 — juho-leinonen-2026 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.