Geng Report: Are juvenile offenders psychologically different from juvenile criminals? A preliminary investigation (Frontiers in Psychology, DOI: 10.3389/fpsyg.2026.1704471) — Highly Suspect

Academic fraud report · Geng Detector

Summary

This AI-assisted forensic review of Liu et al., "Are juvenile offenders psychologically different from juvenile criminals? A preliminary investigation" (Frontiers in Psychology, DOI: 10.3389/fpsyg.2026.1704471, published 26 May 2026) rates the paper HIGHLY SUSPECT. The study is survey-based with no imaging, so detection relied on internal consistency of statistical reporting; all 7 findings were verified verbatim against the source PDF (two machine-verification misses were traced to line-break hyphenation). Major issues include: a 180° contradiction on the direction of the affective empathy group difference between Results and Discussion; two different F values (0.87 vs. 0.76) for the same total-empathy group effect; a Methods passage mislabeling cognitive and affective empathy as subdimensions of self-control; an abstract-level subgroup gender claim with no corresponding subgroup analysis reported; residence percentages for the juvenile offender group summing to 93.34% (≈6.66% unaccounted); pooled statistics from m=50 multiple imputations reported as ranges (e.g., F = 3.5–5.4, p = 0.02–0.05) that conflict with Table 1 single values; and an inconsistent significance-marker legend in Table 1. Confidence is high that the internal contradictions exist as reported; limits include that no single item is individually conclusive and benign explanations cannot be fully excluded without author-provided raw data. Automated Benford/digit tests were excluded as inapplicable. Final misconduct determination requires institutional investigation.

Verdict

🟠 HIGHLY SUSPECT (original verdict maintained after re-verification). All 7 reported findings were confirmed verbatim against the source PDF. Individually, many items could reflect carelessness, but their combination — directional reversal, inconsistent statistics, and variable misattribution — cannot be fully dismissed. This is not a formal finding of misconduct; final determination requires institutional investigation.

Key findings

Directional self-contradiction on empathy: Results report juvenile criminals scored significantly higher than juvenile offenders on affective empathy (F = 3.5–5.4, p = 0.02–0.05); the Discussion states the opposite (criminals' affective empathy significantly *lower* than juvenile offenders, citing Huang and Su, 2010). Both cannot be true.
Two F values for the same effect: total-empathy group effect reported as F = 0.87, p = 0.39 in text but F = 0.76 in Table 1.
Variable misattribution: Methods first treats self-control as approximately normal (Shapiro–Wilk 0.98–0.99, parametric analysis), then writes "for self-control and its subdimensions (cognitive empathy and affective empathy)" — calling empathy subdimensions of self-control — creating internal contradiction in the analysis plan. (Machine verification miss traced to line-break hyphenation "affec-tive".)
Unreported subgroup analysis: Abstract and conclusion claim "Female juvenile offenders reported lower personal belief in a just world but higher affective empathy than their male counterparts", and Methods promises one-way ANOVA / t-tests on demographic variables among juvenile offenders, but Results contain only full-sample ANCOVA / rank-ANCOVA gender main effects (e.g., PBJW B = 2.59; affective empathy F = 5.5–7.4, p = 0.007–0.02).
Residence percentages do not sum: juvenile offender group: 39.33% + 52.81% + 1.20% = 93.34% (≈6.66% unaccounted); control group sums to exactly 100.00% (27.06% + 65.68% + 7.26%). Rounding cannot explain the gap.
Pooled statistics reported as ranges: affective empathy effects given as F = 3.5–5.4, p = 0.02–0.05; F = 5.5–7.4, p = 0.007–0.02; group×age F = 2.0–3.1, p = 0.07–0.10, versus Table 1 single values F = 4.48*, 6.45, 2.56. Rubin-rule pooling of m=50 imputations should yield single pooled values; several ranges straddle the significance threshold.

Table 1 legend error**: footnote reads "\* p < 0.05, \*\* p < 0.01, \* p < 0.001" (third marker should be \*\*\*), inconsistent with in-text markers such as F = 60.20\*\*\*.

Evidence highlights

Automated evidence synthesis (not a misconduct verdict): posterior 99.7% (prior 5%), 95% credible interval [96.8%, 100%], BF = 6.59e+03; assumptions of conditional independence may overstate joint strength.
Highest-weight evidence: [T1] AI-generated text features (0.15/1.0, logLR=1.97, 23.6%); [C1] causal claims under cross-sectional design with acknowledged unremovable confounding (logLR=1.79, 19.2%) — treated as disclosure, not misconduct; Finding 2 (logLR=1.59, 16.2%); Findings 7 and 6 (logLR=1.08 each).
Excluded clues: [S1] Benford and [S2] last-digit tests inapplicable to controlled/reliability-type values (0.93, 0.86, 0.90, 0.72, 0.77, 0.73); [S3] inter-column differences inapplicable. Table 1 Self-control age effect F = 0.00 is consistent with text (linear p = 0.99, nonlinear p = 0.56) and was not listed as a finding.

Notes

Verification confirmed two machine "not found" flags (Findings 3 and 4) were artifacts of line-break hyphenation/sentence splitting; the text is present.
No image forensics applicable (questionnaire study).
Recommended actions: request raw data and imputation/analysis code from authors (empathy direction, F = 0.87 vs. 0.76, pooled affective-empathy statistics, subgroup gender analyses, residence percentage gap); consider contacting the journal regarding Table 1 vs. text statistical consistency and Methods variable attribution.
AI-assisted report for academic discussion; final determination rests with official investigations.