We can’t detect our way out of the AI challenge

Not surprisingly, tools for detecting material written by artificial intelligence have created as much confusion as clarity.

Students at several universities say they have been falsely accused of cheating, with accusations delaying graduation for some. Faculty members, chairs, and administrators have said they aren’t sure how to interpret or use the results of AI detectors.

I’ve written previously about using these results as information, not an indictment. Turnitin, the company that created the AI detector KU uses on Canvas, has been especially careful to avoid making claims of perfection in its detection tool. Last month, the company’s chief product officer, Annie Chechitelli, added to that caution.

Chechitelli said Turnitin’s AI detector was producing different results in daily use than it had in lab testing. For instance, work that Turnitin flags as 20% AI-written or less is more likely to have false positives. Introductory and concluding sentences are more likely to be flagged incorrectly, Chechitelli said, as is writing that mixes human and AI-created material.

As a result of its findings, Turnitin said it would now require that a document have at least 300 words (up from 150) before the document can be evaluated. It has added an asterisk when 20% or less of a document’s content is flagged, alerting instructors to potential inaccuracies. It is also adjusting the way it interprets sentences at the beginning and end of a document.

Chechitelli also released statistics about results from the Turnitin AI detector, saying that 9.6% of documents had 20% or more of the text flagged as AI-written, and 3.5% had 80% to 100% flagged. That is based on an analysis of 38.5 million documents.

What does this mean?

Chechitelli estimated that the Turnitin AI detector had incorrectly flagged 1% of overall documents and 4% of sentences. Even with that smaller percentage, that means 38,500 students have been falsely accused of submitting AI-written work.

I don’t know how many writing assignments students at KU submit each semester. Even if each student submitted only one, though, more than 200 could be falsely accused of turning in AI-written work every semester.

That’s unfair and unsustainable. It leads to distrust between students and instructors, and between students and the academic system. That sort of distrust often generates or perpetuates a desire to cheat, further eroding academic integrity.

We most certainly want students to complete the work we assign them, and we want them to do so with integrity. We can’t rely on AI detectors – or plagiarism detectors, for that matter – as a shortcut, though. If we want students to complete their work honestly, we must create meaningful assignments – assignments that students see value in and that we, as instructors, see value in. We must talk more about academic integrity and create a sense of belonging in our classes so that students see themselves as part of a community.

I won’t pretend that is easy, especially as more instructors are being asked to teach larger classes and as many students are struggling with mental health issues and finding class engagement difficult. By criminalizing the use of AI, though, we set ourselves up as enforcers rather than instructors. None of us want that.

To move beyond enforcement, we need to accept generative artificial intelligence as a tool that students will use. I’ve been seeing the term co-create used more frequently when referring to the use of large language models for writing, and that seems like an appropriate way to approach AI. AI will soon be built in to Word, Google Docs, and other writing software, and companies are releasing new AI-infused tools every day. To help students use those tools effectively and ethically, we must guide them in learning how large language models work, how to create effective prompts, how to critically evaluate the writing of AI systems, how to explain how AI is used in their work, and how to reflect on the process of using AI.

At times, instructors may want students to avoid AI use. That’s understandable. All writers have room to improve, and we want students to grapple with the complexities of writing to improve their thinking and their ability to inform, persuade, and entertain with language. None of that happens if they rely solely on machines to do the work for them. Some students may not want to use AI in their writing, and we should respect that.

We have to find a balance in our classes, though. Banning AI outright serves no one and leads to over-reliance on flawed detection systems. As Sarah Elaine Eaton of the University of Calgary said in a recent forum led by the Chronicle of Higher Education: “Nobody wins in an academic-integrity arms race.”

What now?

We at CTE will continue working on a wide range of materials to help faculty with AI. (If you haven’t, check out a guide on our website: Adapting your course to artificial intelligence.) We are also working with partners in the Bay View Alliance to exchange ideas and materials, and to develop additional ways to help faculty in the fall. We will have discussions about AI at the Teaching Summit in August and follow those up with a hands-on AI session on the afternoon of the Summit. We will also have a working group on AI in the fall.

Realistically, we anticipate that most instructors will move into AI slowly, and we plan to create tutorials to help them learn and adapt. We are all in uncharted territory, and we will need to continue to experiment and share experiences and ideas. Students need to learn to use AI tools as they prepare for jobs and as they engage in democracy. AI is already being used to create and spread disinformation. So even as we grapple with the boundaries of ethical use of AI, we must prepare students to see through the malevolent use of new AI tools.

That will require time and effort, adding complexity to teaching and additional burdens on instructors. No matter your feelings about AI, though, you have to assume that students will move more quickly than you.