AI Detection Tools - Considerations and Limitations

Overview

As generative AI tools like ChatGPT become more prevalent in education, so are AI detection tools to assess the authenticity of student work. However, these tools vary significantly in accuracy, reliability, and ethical implications.

Accuracy and Reliability

Detection tools often yield inconsistent results. While Turnitin claims 99% accuracy with a 1-4% false positive rate (Turnitin, 2023), recent studies (Elkhatat et al., 2023; Weber-Wulff et al., 2023) report much lower real-world success rates (exposing “serious limitations of the state-of-the-art AI-generated text detection tools and their unsuitability for use as evidence of academic misconduct”). Edited or paraphrased AI text can easily bypass detection (Krishna et al., 2023; Problems with AI Detectors – San Diego Legal Research Center).

Bias and Equity

Detectors are more likely to flag non-native English writers, due to linguistic patterns that resemble AI output (Liang et al., 2023). This raises concerns about fairness and inclusivity in academic assessment (Perkins et al., 2024).

Transparency and Trust

Many tools operate as 'black boxes,' offering little insight into their processes (Mienye et al., 2025). This opacity limits students' ability to understand or challenge the results.

Institutional and Ethical Implications

False positives could trigger investigations that cause undue stress and time burdens (USA Today, 2023 ; NY Times, 2025 ). Additionally, legal issues abound surrounding intellectual property and privacy (e.g., FERPA violations) and discrimination (e.g., Title VI, the ADA).

In short, the use of these detectors can result in false accusations, particularly when evaluating writing from multilingual and non-native English speakers, which results in anxiety and extra work for faculty and students, particularly students.

Statement on AI Detection Platforms from the CSU Directors of Academic Technology

“Generative artificial intelligence (AI) is a rapidly developing technology with significant implications for traditional college education. In the short term, it is important to keep in mind that the ability to detect AI-generated content is currently not 100% reliable. Unlike traditional forms of plagiarism detection—which document the original source of the plagiarized text as evidence of academic dishonesty—no such evidence currently exists for AI detection platforms. While some AI detection platforms claim up to 99% accuracy, even a 1% potential false-positive or false-negative rate presents a considerable challenge to enforcing academic integrity because the instructor cannot "prove" their case to an independent observer. Additionally, a recent study (Liang et al., 2023) from researchers at Stanford University indicated that AI detectors are susceptible to bias against non-native English writers, increasing the risk of false-positives for already marginalized learners. Suggested strategies include constructive discussions between students and faculty, as well as clear communication about expectations related to AI use for assignments and assessments via course syllabi and assignment instructions.”

Additional Resources

• AI detectors - Part 1: Accuracy, deterrence, and fairness in practice
  • Quote 1: "There is no AI-generated text that a human could not have written; human-written text and AI-generated text are not mutually exclusive sets."
  • Quote 2: "AI detectors offer no reliable standard of evidence, no meaningful transparency, and no guarantee of fairness."