AI detection tools use algorithms to analyze writing and estimate whether it was generated by AI. These tools examine text features such as sentence structure, vocabulary, and grammatical patterns to make predictions. However, the results should be interpreted with caution.
AI Detectors
Accuracy is still unreliable: AI detectors can produce both false positives (human writing flagged as AI) and false negatives (AI-generated writing not detected). Improvements intended to reduce false accusations can also make AI-generated text harder to identify. Results may also vary when the same text is tested multiple times.
Detection does not establish misconduct: AI detectors may provide an indication that AI was involved, but they cannot determine how AI was used. For example, minor revisions to human-written text may cause it to be flagged as AI-generated.
Poor Performance on Nonstandard Formats: These tools can struggle with short texts or nontraditional formats like poetry, scripts, or bulleted lists.
Opaque Methods: Most detectors do not explain how they reach their conclusions, making results difficult to verify.
Equity Concerns: Research has found that some detection tools are more likely to flag writing by multilingual students or writers with less conventional styles, potentially reinforcing existing biases.
Automated Proctoring
Monitors behaviors, not cheating: Automated proctoring systems may interpret behaviors such as eye movement, head position, or pauses as indicators of academic dishonesty. These behaviors are not conclusive evidence of cheating.
Accessibility and equity concerns: Automated proctoring can create barriers for students with disabilities and students taking exams in shared, crowded, or otherwise challenging home environments.
Surveillance can impede performance: Being monitored during an exam can increase stress and anxiety, potentially affecting students' ability to concentrate. Behaviors associated with anxiety or disability may also be misinterpreted as suspicious.
Lockdown Browsers
Create technological barriers: Students need compatible, up-to-date computers and reliable internet connections. Software compatibility problems, outdated devices, or connectivity issues can interrupt an assessment and place additional stress on students.
Create equity concerns: Students do not have equivalent access to computers, operating systems, hardware, or internet connections. Students using older, borrowed, or less reliable devices may face greater barriers to completing an assessment. Lockdown software may conflict with accessibility tools, alternative input devices, multiple-monitor configurations, or other customized computer settings.
Provide only partial security: Lockdown browsers restrict activity on the device running the software but cannot prevent students from using phones, tablets, second computers, printed notes, or other off-device resources..
Require significant institutional support: Ensuring reliable access may require technical troubleshooting, loaner devices, testing, and contingency plans. Providing a truly controlled and equitable testing environment requires institution-managed computers and infrastructure, which can be costly and difficult to implement at scale.
Practical Recommendations
Use Scores as Conversation Starters: Detection results should not be treated as definitive evidence. If a concern arises, speak with the student. Ask how they approached the assignment and what tools, if any, they used.
Make it a Teaching Moment: Students may not understand your AI policy or expectations. Use these moments to offer guidance and clarify what responsible AI use looks like in your course.
Reach Out to Administration: See this page for more information about submitting an academic misconduct allegation.
Useful Resources
Overview: Learning Technologies and Academic Integrity (DLD Faculty Guides)
Clifton, M. (2024, September 17). Black teenagers twice as likely to be falsely accused of using AI tools in homework. Semafor.
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & James, Z. (2023). GPT detectors are biased against non-native English writers.
Perkins, M., Roe, J., Vu, B. H., Postma, D., Hickerson, D., McGaughran, J., & Khuat, H. Q. (2024). Simple techniques to bypass GenAI text detectors: implications for inclusive education. International Journal of Educational Technology in Higher Education, 21(1), 53.
Salem, L., Fiore, S., Kelly, S., & Brock, B. (2023). Evaluating the effectiveness of Turnitin’s AI writing indicator model.Temple University Center for the Advancement of Teaching