AI in Teaching: Recognize and Respond to Inappropriate AI Use
Generative AI can now complete a wide and growing range of tasks, from writing essays to solving complex problems. This creates a challenge for educators when students attempt to pass off AI-generated content as their own. How can instructors recognize when a student is submitting unauthorized AI-generated content? To what extent can they trust AI detection reports from tools like Copyleaks? What do they do if they suspect unauthorized AI use?Identifying AI-Generated Work
It is very difficult to identify and confirm when a student's assignment submission has been generated, in whole or in part, by AI. Generative AI is incredibly versatile in its ability to craft text in a variety of styles and for a variety of audiences. A little knowledge of prompt engineering and some creative remixing can produce content that evades popular AI detectors. Even when an AI detector flags a submission as a possible AI creation, it is difficult to be completely confident that the finding is not a false positive—especially if the student denies using AI.
Easy solutions to identifying AI-generated work do not exist. The best an educator can do is look for a variety of clues that a student may have used AI. Such clues include:
- Writing style inconsistencies: The writing style or voice of a student's submissions may differ noticeably from one assignment to another, indicating that the different writing samples did not come from the same source.
- Too perfect or too generic: An essay may seem too polished, or it may lack depth, demonstrating only a generic understanding of the topic.
- Overly complex or overly simple language: AI-generated text will often seem a little too ... something. It may use language that seems unusually complex, phrases that seem a little over the top, explanations that seem overly simplified, dialects that seem a little too stereotypical, or arguments and conclusions that seem a little too sterile and saccharine.
- Off-topic content: Often, AI-generated assignment submissions will read well but miss the point of the assignment—lacking personal reflection, strong statements of opinion, course-specific references, or assignment-specific criteria.
- Inaccuracies or inventions: AI-generated content may confidently state things that are not true, include links to sites or references that don't exist, or use made-up quotes.
- Heavily formulaic text: AI-generated content frequently uses abundant headings followed by short explanatory paragraphs, lists, and distinct introduction and conclusion paragraphs. A good way to train yourself to spot AI patterns is to try completing the assignment using AI yourself, using multiple chatbots and a variety of prompts.
- Unusual formatting: Sometimes AI-generated assignment submissions show changes in font midway through the assignment. Egregious examples may even appear in a font common to an AI chatbot. They may also show sudden shifts in text alignment, heading style, bullets and numbering, or spacing.
- A high percentage score on an AI detector: Sometimes the first hint that a student may have used AI to generate an assignment is a high score on an AI detector like Copyleaks. Ideally, the high AI detection score should be accompanied by one or more of the other indicators listed above. Assignments that show high AI detection scores without meeting any of the above criteria should be viewed with skepticism.
About AI Detectors
A number of AI detection tools exist, including Copyleaks (currently used by USU for plagiarism and AI detection), Turnitin, GPTZero, and others. These tools attempt to use algorithms to identify patterns consistent with either human or AI-generated text. Their scores are a measure of probability as to how likely it is the text was generated by a human.
Published research differ s as to the accuracy of AI detectors. As a rule, AI detectors are more likely to report false negatives than false positives, although both are statistically possible. AI detectors are less accurate the less text they have to work with. As of February 2024, Copyleaks documentation suggests that text should be at least 400 words in length for reliable results.
Tips for Reviewing AI Detector Reports
Until more conclusive research exists on the accuracy of AI detection tools, positive AI detection reports should be approached with skepticism. They should not be the only point of evidence or suspicion in an accusation of student cheating. If you run AI detection on a student's submission and get a report that it was likely AI-generated, check for a few things that may have led to a false positive:
- Is the submission shorter than 400 words in length?
- Does the submission list question text followed by question answers (resulting in changes of voice and a more formulaic appearance)?
- Does the submission appear to use a presentation template that results in unusual formatting, such as a PDF with extra spacing in the words or other formatting oddities that may throw off a detector?
- Does the submission mix the work of multiple authors into a single file, resulting in changes in author voice and other inconsistencies?
- Does the student speak English as a second language and consistently write in a particularly formulaic style?
If you don't see any factors that may have led to a false positive, read the text itself.
- Does the text read like AI? (See the list above under "Identifying AI-Generated Work")
- Does the text contain common grammatical or speech idiosyncrasies that AI would be unlikely to use?
- Does the text use inconsistent or incorrect punctuation, typical of unpracticed writers?
- Does the text include personal statements, opinions, or strong statements that are atypical of AI?
It may not hurt to run the same text in multiple AI detectors to see if they also turn up a high AI detection score. You will likely see conflicting results, but consider them as data points in your investigation. Next, if you are still in doubt, approach the student.
- Ask the student to share with you how they completed their assignment.
- If you can have a live conversation, ask the student to retell the main ideas and arguments of their submission in their own words.
- Make sure the student understands the class policy on AI use.
Options if You Suspect AI-Generated Work
Once you have weighed all of the evidence, if you still believe a student used AI in an unauthorized manner to complete their work, then you should respond according to the policies you outlined in your syllabus for unauthorized AI use.
A less aggressive option is to have a discussion with the student about the course's AI policy and the assignment expectations, and then ask the student to resubmit their work with something that better meets these expectations.
Importantly, before you penalize a student with a poor grade on the basis of academic dishonesty, Article VI of the Student Code states that you must submit an Academic Integrity Violation Form. This ensures due process for the student, as described in Article VI, Section 4. If you do not file a form, according to Section 4, "the student may appeal the determination that an academic violation occurred."
Sometimes you may decide there is too much reasonable doubt to charge a student with academic dishonesty—especially considering how impactful such a charge could feel to a student who did not intend to cheat. In these cases, you can still grade the student according to how well they met the assignment requirements without filing any forms. You can also provide the student with advice on how to improve their work.
Recognizing the Varying Degrees of AI Use
As you grade student assignments, you may not see high AI detection reports, but you may see other oddities like students answering open-ended questions in very similar ways and with very similar ideas. Even though students may not be using AI to write their responses, they may still be feeding the prompts to an AI chatbot to get ideas, and then writing those ideas down in their own words. Although this may seem out of line to you, in students' minds, it may not seem different in principle than a Google search. You may need to define in your syllabus the extent to which you are willing to let students use AI, beyond having it write their essays outright. Note, however, that policing some of these policies may prove difficult and may only be possible in a proctored exam scenario.