Uloženo v:
| Hlavní autor: | |
|---|---|
| Médium: | Recurso digital |
| Jazyk: | angličtina |
| Vydáno: |
Zenodo
2026
|
| Témata: | |
| On-line přístup: | https://doi.org/10.5281/zenodo.20100527 |
| Tagy: |
Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
| _version_ | 1866901362414977024 |
|---|---|
| author | Mei, Hangyu |
| author_facet | Mei, Hangyu |
| contents | <p>An open letter to UNESCO's Section for Higher Education proposing the replacement of automated AI-detection regimes in academic institutions with a process-based verification framework.<span class="Apple-converted-space"> </span></p> <p>The letter argues that current AI-detection tools (Turnitin AI, GPTZero, Copyleaks, Sapling, Winston AI, and others) suffer from two compounding failures: (1) technical non-falsifiability — no detector establishes a physically verifiable signature of machine origin, and the author demonstrates from personal practice that heavily AI-assisted work can reliably produce near-zero detection scores; and (2) inverse incentive structure — the regime systematically punishes pedagogically encouraged AI use while a paid grey-market economy enables wholesale fraud to pass undetected.</p> <p>A controlled empirical battery (companion dataset, DOI 10.5281/zenodo.20094765) supplies the load-bearing evidence: 26 canonical human texts written between 81 BCE and 1962 CE, run through five contemporary detectors. One commercial tool classifies 92% of these pre-LLM canonical texts as AI-generated; another classifies 100% of modern human academic writing as AI but 0% of the same canonical corpus. These two failure patterns cannot be reconciled by any threshold adjustment and falsify the institutional claim that detection scores constitute physical evidence of authorship.</p> <p>The proposed alternative anchors academic integrity on the person rather than on the text, via three operational components: (a)post-submission oral verification (a scaled-down thesis-defense model); (b) argument-level overlap detection (semantic-embedding and argument-graph methods comparing submitted reasoning structures against existing literature, rather than surface text); and (c) institutional adoption of the publication-system standard already used by Nature, Science, IEEE, ACM, and the major Elsevier and Springer titles.</p> <p>The letter also addresses a structurally identical failure at the other end of the academic pipeline: predatory first-authorship practices in research-intensive doctoral programs. The same students whose intellectual labour is misread as AI-generated at the entry end are systematically dispossessed of first-authorship at the exit end. Documented cases (Tao Chongyuan, 2018; Lu Jingwei, 2018; Xiangya 2024;<span class="Apple-converted-space"> </span></p> <p>Huazhong Agricultural University 11-student joint denunciation, 2024; Chang'an University 1,000-page filing, 2021) establish the pattern.</p> <p>The letter concludes with a request for UNESCO to convene an international working group on assessment-method transition, building on the principles laid out in the 2023 Guidance for generative AI in education and research.</p> <p>Bilingual archive: English (primary) plus Chinese companion translation. Technical Report TR-AIDET-2026-01 and Evidence Annex are bundled.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_20100527 |
| institution | Zenodo |
| language | eng |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Field Observations on AI-Detection Practices in Academic Institutions and a Proposal for Process-Based Academic Integrity Verification Mei, Hangyu academic integrity, AI detection, generative AI, higher education, UNESCO, education policy, process-based verification, oral defense, argument overlap detection, GPT detector bias, non-native English writers, Turnitin AI, GPTZero, Sapling, Winston AI, open letter, anachronistic false positives, predatory authorship <p>An open letter to UNESCO's Section for Higher Education proposing the replacement of automated AI-detection regimes in academic institutions with a process-based verification framework.<span class="Apple-converted-space"> </span></p> <p>The letter argues that current AI-detection tools (Turnitin AI, GPTZero, Copyleaks, Sapling, Winston AI, and others) suffer from two compounding failures: (1) technical non-falsifiability — no detector establishes a physically verifiable signature of machine origin, and the author demonstrates from personal practice that heavily AI-assisted work can reliably produce near-zero detection scores; and (2) inverse incentive structure — the regime systematically punishes pedagogically encouraged AI use while a paid grey-market economy enables wholesale fraud to pass undetected.</p> <p>A controlled empirical battery (companion dataset, DOI 10.5281/zenodo.20094765) supplies the load-bearing evidence: 26 canonical human texts written between 81 BCE and 1962 CE, run through five contemporary detectors. One commercial tool classifies 92% of these pre-LLM canonical texts as AI-generated; another classifies 100% of modern human academic writing as AI but 0% of the same canonical corpus. These two failure patterns cannot be reconciled by any threshold adjustment and falsify the institutional claim that detection scores constitute physical evidence of authorship.</p> <p>The proposed alternative anchors academic integrity on the person rather than on the text, via three operational components: (a)post-submission oral verification (a scaled-down thesis-defense model); (b) argument-level overlap detection (semantic-embedding and argument-graph methods comparing submitted reasoning structures against existing literature, rather than surface text); and (c) institutional adoption of the publication-system standard already used by Nature, Science, IEEE, ACM, and the major Elsevier and Springer titles.</p> <p>The letter also addresses a structurally identical failure at the other end of the academic pipeline: predatory first-authorship practices in research-intensive doctoral programs. The same students whose intellectual labour is misread as AI-generated at the entry end are systematically dispossessed of first-authorship at the exit end. Documented cases (Tao Chongyuan, 2018; Lu Jingwei, 2018; Xiangya 2024;<span class="Apple-converted-space"> </span></p> <p>Huazhong Agricultural University 11-student joint denunciation, 2024; Chang'an University 1,000-page filing, 2021) establish the pattern.</p> <p>The letter concludes with a request for UNESCO to convene an international working group on assessment-method transition, building on the principles laid out in the 2023 Guidance for generative AI in education and research.</p> <p>Bilingual archive: English (primary) plus Chinese companion translation. Technical Report TR-AIDET-2026-01 and Evidence Annex are bundled.</p> |
| title | Field Observations on AI-Detection Practices in Academic Institutions and a Proposal for Process-Based Academic Integrity Verification |
| topic | academic integrity, AI detection, generative AI, higher education, UNESCO, education policy, process-based verification, oral defense, argument overlap detection, GPT detector bias, non-native English writers, Turnitin AI, GPTZero, Sapling, Winston AI, open letter, anachronistic false positives, predatory authorship |
| url | https://doi.org/10.5281/zenodo.20100527 |