Reviewing Suspected AI-Generated Text: Evidence, Limits and a Practical Workflow
Fluent wording, few errors or a familiar structure cannot establish who wrote a text. When AI involvement is suspected, ask whether the content is correct, its sources can be checked and the drafting process complies with the applicable rules.
A detector score is a classification result under particular conditions. It does not identify the author, establish AI’s contribution or determine misconduct. A defensible review combines content checks, drafting evidence and discussion with the writer.
AI-text review needs sources and reasoning checks
- Sources
Verify references and DOIs
- Reasoning
Check evidence for each claim
- Provenance
Review drafts and working records
- Conclusion
Preserve detection limitations
Conceptual diagram. Writing-style signals do not establish that an author used AI.
Source: Lê Minh · Conceptual illustration, not software output.1. Four questions often collapsed into one score
| Question | Relevant evidence | What a detector alone cannot establish |
|---|---|---|
| Was AI used? | Tools, scope of assistance and the drafting process. | Authorship identity or the entire writing history. |
| Is the content correct? | Sources, data, calculations and coherent reasoning. | That a low-scoring manuscript is scientifically sound. |
| Was content copied? | Textual overlap, citations and permitted reuse. | That an AI score is a similarity score. |
| Was a rule violated? | Applicable rules, disclosures and relevant evidence. | That every form of AI assistance is prohibited. |
Identify the work being reviewed and the applicable rules at the outset. Grammar assistance, translation, outlining and generating an analysis are different contributions. Specify what is allowed and what must be disclosed rather than applying one label to all assistance.
A sound study may still require disclosure of AI use. A fully human-written manuscript can contain inaccurate references or weak reasoning. Record content quality and tool use as separate questions.
2. What does detector research establish?
Liang et al. (2023) examined detectors using English writing by native and non-native writers and documented false positives affecting the latter under the conditions studied. This raises a fairness concern, but does not measure every detector available in 2026 or establish error rates for Vietnamese text.
Dugan et al. (2024) introduced RAID to test detectors across generators, domains and generation conditions. Their findings show why performance may change outside familiar testing conditions. An advertised accuracy is not a guarantee for the manuscript type being reviewed.
These findings do not make every detector useless. They require attention to evaluation data, language, genre and date. One newspaper excerpt being flagged is an observation, not an estimate of a product’s accuracy.
3. Interpret percentages and labels correctly
Products can return labels, classification scores or the proportion of qualifying text flagged by a model. These are different quantities. For example, Turnitin describes its AI percentage in relation to qualifying prose and distinguishes it from similarity; the percentage is not the probability that a particular person used AI (Turnitin, n.d.).
Read an “80%” result using that product’s definition. Do not translate it into an 80% probability of misconduct or a claim that only 20% was written by the author. Do not average outputs from unrelated detectors without compatible scales and calibration.
No score threshold transfers automatically across products and purposes. An organization using a screening threshold should distinguish screening from a final determination and evaluate the consequences of both false positives and false negatives.
4. A hypothetical example: why high accuracy still needs scrutiny
The following numbers were constructed for teaching and are not measurements of Turnitin or another product. Consider 1,000 texts with known origins: 100 AI-generated and 900 human-written. Suppose a detector flags 80% of the AI texts and incorrectly flags 5% of the human texts.
| Known origin | Flagged as AI | Not flagged | Total |
|---|---|---|---|
| AI-generated | 80 (true positives) | 20 (false negatives) | 100 |
| Human-written | 45 (false positives) | 855 (true negatives) | 900 |
| Total | 125 | 875 | 1,000 |
In this hypothetical set, 45 of the 125 flagged texts were human-written. Overall accuracy of 93.5% coexists with 36% incorrect classifications among flagged texts because human writing is more prevalent. Precision would change with the prevalence of AI texts in the evaluated set.
This calculation uses assumed, known-origin labels. It does not assign an individual authorship probability to a person under review or establish how many people in a real group violated a rule.
5. Check language, version and manuscript-handling conditions
Read the product requirements before submitting text. When checked on 15 November 2026, Turnitin’s AI Writing Report guidance listed supported languages but did not include Vietnamese. An out-of-scope result should not be presented as validated performance for Vietnamese (Turnitin, n.d.).
Translating a Vietnamese manuscript into English changes the input. A detector then evaluates transformed text. Using this procedure requires a separate assessment of translation’s effect. Likewise, short excerpts, tables, code and reference lists should not be treated as supported long-form prose unless the product documents that capability.
For unpublished or restricted manuscripts, use services authorized to handle that material and check organizational storage/access conditions before upload. Record the product, date, input type and settings so results from different versions are not conflated.
6. A documented review workflow
Define the question
Record permitted AI assistance and the decision contemplated. Do not begin by searching for a tool that returns a high score.
Preserve a comparison copy
Record the text version, source and review date. Compare the same version across checks and identify changes between reviews.
Check verifiable claims
Select important claims, numbers, citations and DOIs against original sources. Record the actual issue: nonexistent reference, mismatched year or unsupported statement.
Inspect reasoning
Compare the question, method, results and conclusions. An interpretation inconsistent with its table requires correction regardless of the detector result.
Discuss the writing process
Ask about source choices, calculations and revisions. Use genuine drafts, notes or existing version history; do not request retrospective manufacture of evidence.
Record findings and limits
Separate verified findings, unresolved questions and requested information. Include detector output, if used, with its scope and limitations.
This is an editorial workflow proposed here, not a validated diagnostic scale. Missing version history does not prove AI use, and a useful discussion does not authenticate every part of a manuscript. Decisions should follow available evidence and the organization’s process.
Turnitin itself cautions that its model can misclassify text and that its output should not be the sole basis for adverse action against a student (Turnitin, n.d.). Translate suspicion into specific questions and facts that the writer can address.
7. A neutral, checkable review record
| Record | Illustrative example | Next step |
|---|---|---|
| Claim to check | The manuscript says a cited study established causality. | Inspect the original design and conclusion. |
| Source information | A DOI resolves, but the title or year differs. | Check metadata and full text; correct the citation where needed. |
| Drafting information | Drafts and notes exist; the scope of AI assistance is unclear. | Request a description of the tool and tasks assisted. |
| Detector information | A version was flagged; language support needs checking. | Retain the report and its limitations without converting its label into authorship. |
Review template: We identified [specific content/source issue] in version [date or identifier]. The result from [product and review date] is recorded within [input conditions]. We request [source material, explanation or relevant records]. The available information does not establish authorship or a violation. Further decisions will follow the organization’s evidence and procedure.
A useful response identifies the work performed, any tools used and genuine sources or drafts retained. Rewording solely to reduce a detector score does not resolve an unclear content issue or scope of tool use.
8. Keep conclusions proportional to the evidence
- Check sources and calculations even when the detector score is low.
- Do not treat style, punctuation or fluency as proof of AI authorship.
- Read the exact product’s score definition, language support and genre limits.
- Distinguish false positives, false negatives and overall accuracy.
- Do not treat multiple detectors as independent votes or interchangeable measurements.
- Give the writer an opportunity to explain checkable facts.
- Match the certainty of a finding to its evidence.
A useful review identifies what needs correction and what remains to be established. Its value comes from checkable sources, reasoning and decisions rather than a percentage that appears to settle authorship.
Further reading
Types of scholarly articles; Manuscript preparation and publication workflow.
References
- Dugan, L., Hwang, A., Trhlík, F., Zhu, A., Ludan, J. M., Xu, H., Ippolito, D., & Callison-Burch, C. (2024). RAID: A shared benchmark for robust evaluation of machine-generated text detectors. In L.-W. Ku, A. Martins, & V. Srikumar (Eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 12463–12492). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.acl-long.674
- Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), Article 100779. https://doi.org/10.1016/j.patter.2023.100779
- Turnitin (n.d.). Using the AI Writing Report. Retrieved November 15, 2026, from https://guides.turnitin.com/hc/en-us/articles/22774058814093-Using-the-AI-Writing-Report