MercoPress. South Atlantic News Agency

New Research Raises Questions About the Accuracy of AI Detection Tools

Universities are facing new questions about the reliability of automated writing checks. A research review published in October 2026 found serious problems with AI detection accuracy. The findings raise concerns for students whose original assignments may receive incorrect scores. Researchers are now urging institutions to examine additional evidence before making misconduct accusations.

New Study Examines 90 Research Papers

The October 17 study appeared in the journal Discover Education. Researchers examined 90 studies published between 2020 and 2026.

Their review identified problems involving language differences and automated classification errors. Some systems incorrectly identified human-written assignments as artificial intelligence-generated content. These mistakes are known as false positives and can have academic consequences.

The researchers also found that results varied considerably between different detection systems. This makes comparing scores from separate platforms particularly difficult.

Commercial Tools Show Different Accuracy Levels

Another study published in February 2026 tested two commercial detection platforms. Researchers examined 192 texts covering student assignments and professional writing. The collection also included AI-generated passages and documents combining human and machine writing.

Originality achieved 69% overall accuracy in this particular experiment. Turnitin recorded 61% accuracy under the same evaluation conditions.

Both systems struggled when passages combined human writing with AI assistance. Scientific writing also produced weaker results than humanities material during testing.

These findings do not mean every detection platform performs identically. Accuracy depends on the documents and testing conditions involved.

Language Differences Raise Fairness Questions

The debate also concerns students who write in English as an additional language. Their writing may follow patterns that detection systems interpret incorrectly.

An October 2026 conference paper examined 135,389 pairs of academic documents. Researchers compared manuscripts before and after professional English-language editing. They then tested the documents using 13 AI detection systems.

Some systems produced higher AI scores after professional editing. Others reduced their scores when processing the same revised documents.

The findings suggest that writing style can influence automated classifications. For universities across Latin America, such concerns have particular relevance to multilingual academic communities.

Can an AI Detector Prove Academic Misconduct?

An AI detector examines writing patterns rather than directly establishing authorship. A high percentage therefore cannot independently prove that someone cheated.

Turnitin acknowledges that its system can incorrectly classify human-written material. The company advises educators against using detection results alone for disciplinary decisions.

Its software also hides specific scores between 1% and 19%. This policy addresses the higher risk of misleading interpretations at lower detection levels.

Researchers recommend examining several forms of evidence before reaching conclusions:

- Earlier assignment drafts that document the student's writing process.

- Research notes and references used during assignment preparation.

- Conversations with students about their arguments and research decisions.

- Institutional policies explaining permitted uses of writing assistance.

Universities Face Pressure to Review Assessment Practices

The latest findings also challenge how universities investigate suspected plagiarism. Automated scores may identify assignments requiring closer examination, but human review remains necessary.

Researchers recommend clearer assessment policies and accessible procedures for challenging incorrect accusations. Students also need guidance about acceptable editing and AI assistance.

The debate now extends beyond software performance into academic fairness. Universities must protect original work without treating uncertain predictions as established facts.