EdTechnical Why AI Detectors Don't Work for Education
Oct 2, 2025
Explore the challenges of AI detection in education as traditional tools struggle against clever student tactics. Learn how paraphrasing and translation thwart detection efforts, while false positives compromise accuracy. The hosts advocate for process-based assessments like keystroke tracking and oral exams, offering more reliable evaluations. They also examine institutional barriers to innovation and discuss the implications of students' motivations in using AI. This insightful conversation questions how we can enhance academic assessment in the digital age.
AI Snips
Chapters
Transcript
Episode notes
False Flag From Old Group Work
- Owen Henkel recounts his PhD application essay being flagged for plagiarism due to a group assignment uploaded to a repository.
- He resolved it by reviewing side-by-side matches and explaining the provenance to admissions staff.
Copy Matching Breaks With New AI Text
- Traditional plagiarism detection matches text against existing databases and works well for verbatim copying.
- It fails when content is newly generated or heavily paraphrased, as with generative AI outputs.
Three Approaches To AI Detection
- AI detectors use watermarking, stylistic statistics, and process tracking as three main approaches.
- Each approach faces practical limits: watermark secrecy, style evasion, and implementation cost for tracking.