Is Artificial Intelligence (AI) detection software accurate?
How do AI detectors work?
Faculty have numerous AI detection tool options to choose from. Many are quite tempting – especially when they share their claimed accuracy rates of 90% and higher! However, great debate rages around AI detection tool use, and many criticisms are valid.
AI detectors focus on two components of writing: perplexity and burstiness. Perplexity refers to the variety and complexity of words being used, and burstiness refers to the same in relationship to sentences. It has been determined that AI writing, in general, is not very creative in its vocabulary and sentence structure. Therefore, written work scoring low in both perplexity and burstiness is often labeled as “AI-generated.”
Two immediate issues or concerns associated with these criteria are:
ELA learners often display low perplexity and burstiness since they are in the process of learning English vocabulary, grammar, syntax, etc. Studies have shown their work is falsely accused of being AI-generated at much greater rates than that of other students
“Word spinners” and other AI-based technology can rewrite text to improve perplexity and burstiness – and thus avoid detection. Studies have shown that AI detectors are quite vulnerable to even slight upgrades in text quality, thus avoiding detection is quite easy
But there are other issues associated with AI detectors to consider:
Due to bias and vulnerabilities, it is strongly recommended never to use AI detection tool scores as the sole determinant of AI-based plagiarism
With many tools, the submitted work is retained and sold to other parties (so be sure to read terms of use policies)
With many tools, the submitted work is used to train the model (so be sure to read terms of use policies)
If AI detection software is not officially purchased/licensed through an academic institution, it is possible to violate FERPA if personally identifiable information is not removed from submitted work
If AI detection software is not officially purchased/licensed through an academic institution, a student’s copyright/intellectual property rights could be violated
If a tool requests accuracy opinions/feedback, then the user is provided uncompensated labor to train the AI detection model
Research has shown that students’ academic careers have been harmed by false claims of AI-based plagiarism
Academic research has demonstrated that AI detection tools can be very susceptible to trickery and evasion techniques
According to some sources, student work that is heavily edited and/or rewritten by Grammarly can be flagged as AI-generated (there is not universal agreement on this, however). For additional information refer to “Grammarly and AI Detectors: Myth or Fact?” shared by Originality.ai.
How to Use AI Detectors
If you do choose to use AI detectors, how can you best capitalize on the information you gain?
Use the report’s data as a conversation starter = engage with the student whose work is raising red flags
Tell them, frankly, why you are suspicious that their work was AI-generated
Ask them explain why they resorted to using AI when it's not allowed
Ask if there were conditions, circumstances, or directions that could have been changed to help reduce the appeal of AI’s “quick-fix”
Readings on AI Detectors