End the AI Detection Arms Race
Pitting students and faculty against each other helps no one.
Special Note: Below is an editorial article I was invited to write for the Patterns journal. The accepted version is below per the journal’s publishing guidelines. Let me know if you want a copy of the final printed version.
While Large Language Models (LLMs), like ChatGPT, have numerous flaws, they are excellent "word calculators" that can easily mimic human use of language and understanding of our world. No group in education was quicker to recognize the potential of LLMs than students, who quickly put the essay writing, question answering, and tutoring abilities of LLMs to use after the Nov 2022 release of ChatGPT by OpenAI. Less than twenty days after ChatGPT's public availability, some Stanford students used it to help with their Fall Quarter Final exams1. Seventeen percent of the 4,497 Stanford students informally surveyed reported using ChatGPT for their exams, and five percent of those used output from ChatGPT without any edits. Rapid adoption at Sanford was not surprising—OpenAI’s leader Sam Altman is a Stanford dropout—but adoption spread like wildfire worldwide, quickly making ChatGPT the fastest-adopted platform in history2.
A discussion immediately started about how AI-generated materials could be detected and flagged. While AI-generated images can be watermarked, either explicitly or in ways that hide the watermark, text generated by AI can't be easily tagged as being AI-generated, and if it is tagged, those tags can be easily removed.
Several companies rapidly deployed tools that purported to recognize patterns indicative of AI writing. Initially, this technology looked promising, and even OpenAI released its own AI detector in early 2023. An early startup developed by a Princeton grad, GPTZero, analyzed text in terms of burstiness (how repetitive or clustered a group of words is) and perplexity (how predictable the word use is) and became very popular.
However, people soon discovered that the generating prompts could be altered, or the output text could be edited to avoid detection3. Numerous startups launched to help students make AI-generated text undetectable. The AI generation vs. AI detection arms race began to escalate, with Turnitin and other established Ed-Tech companies became the arms dealers for faculty headed into battle.
Unfortunately, many faculty misused AI-detection tools during those early days, leading to much "collateral damage" and headlines about professors failing entire classes because the professor did not understand how to use AI detection.4 Hundreds of TikToks showed the anguish of accused students, encouraging the public to adjudicate their case.
To complicate matters further, it quickly became apparent that AI detectors had high false positives and negatives.5 These systems also discriminated against non-native English speakers.6 OpenAI shut down its detector relatively swiftly and warned that such systems were unreliable.7 GPTZero pivoted to producing AI writing tools that instructors can use to monitor students as they write. Yet, that didn't stop other Ed-Tech companies from baking AI detection tools into their systems and selling them to institutions at all levels.
While faculty and administrators searched for new tools to help them maintain the status quo, most students saw clearly what these new tools revealed about higher education.
Writing in The Princetonian in February 2023, student Christopher Lidard concluded, "... if students are being assigned essays that can be written by ChatGPT, perhaps it's not a good assignment in the first place. ChatGPT begs us to rethink the purpose and value of homework. Some would argue that ChatGPT subverts the purpose of a good education by providing students with instant, personalized assistance on a wide range of subjects and topics. Yet if Princeton homework can be completed by a machine, what is its true value?"8
The type of student assignments Lidard refers to are a form of a priori knowledge on the faculty's part. We know what the student's response should be before the student goes through the research and writing experience. For example, suppose I assign a student to write an essay about Harry S Truman's significant achievements during his presidency. In that case, I know what should be in that student's response: something about FDR's death, dropping the bomb, integrating the military, firing McArthur, recognizing Israel, etc.
This type of assignment works well in our Taylorist factory model of undergraduate education, with large classes and armies of teaching assistants trained to process papers. And I hate to say it, but some faculty also use AI-based grading programs to provide initial scoring for these types of assignments.
Of course, these are the types of essays that are super simple for an LLM like ChatGPT to write. And for essay writing companies to churn out in the pre-ChatGPT days, at least we forget that cheating didn't start with LLMs.
A better assignment is one where the faculty and student don't know the answer in advance; the knowledge comes a posteriori or after the student's experience. For example, what was the impact of President Truman's decision on the trajectory of the student's family or the place where they grew up? I don't know the answer to that, and neither does the student or ChatGPT. An LLM could help the student write up their research but can't do it well on its own. This example is also a much more interesting type of assignment for the student and faculty member, one that creates knowledge.
While epistemologists might have issues with my use of a priori and a posteriori, I find these valuable constructs when designing assignments. ChatGPT forces us to discard the use of homework assignments as a tool for evaluation and turn it into an opportunity to expand students' understanding of core concepts through original research and insight. It also allows us to raise our expectations of what we expect from students, especially if we allow some use of LLMs in key areas of our assignments.
Showing your students how to use an LLM for their assignments or co-learning it with them offers an opportunity to de-escalate the situation and restore trust in the Faculty-Student relationship. Even more rewarding is that incorporating LLM use in your class is a gateway to teaching how to ask good questions in your field, as getting good results from LLMs requires asking good questions.
However a posteriori project-based assignments are not amenable to the industrial production of education. ChatGPT will not destroy education, but I am hopeful it will finally blow up the large lecture and huge student credit load many faculty have to bear.
Unfortunately, in many institutions this fall, the AI detection arms race will continue, pitting institutional goals (massive classes and easily gradable assignments) against student goals (clearing mundane assignment hurdles so they can move on to more significant challenges). It is a race that can't be won. To borrow a line from the movie Wargames, the only winning move in the AI-Writing vs AI-Detection war is not to play.
📆 Upcoming Talks/Classes
I will be presenting “Managing the Learning Machine” at 8:00 AM on September 10th for the MU Retiree’s Association (In Person and Zoom). More information and Registration will be available on MU Retiree’s Association website.
I will give a talk on Artificial Intelligence and The Election a couple of times in September:
Tuesday, September 10, 6:30pm - 8:00pm at the Missouri River Regional Library in Jefferson City. More information is available on the Missouri River Regional Library website.
September 13th, 3:00pm to 4:15pm at the Trulaske College of Business in the Cornell Auditorium. If you want to attend, let me know and I will get you a parking pass.
Description: Artificial Intelligence (AI) is challenging our ability to distinguish between truth and deception, a critical skill for democracy. This session will explore AI's potential role in spreading disinformation, particularly during election periods. Prof C will also offer strategies for navigating a world where trust is increasingly elusive.
My friend and colleague, Sophia Rivera Hassemer, is teaching “Technology Potpourri” for Osher on Sept 12, 19, 26, and Oct 3 from 9:30 to 11am, and I will be her assistant! It will be in person only at the Moss building, and will be very hands on with technology. More information and Registration will be available on the Osher website.
I will present “Harnessing AI for Nonprofit Growth” from
10:45 - 11:45 a.m., on November 7 via zoom. More information and Registration will be available on the New Chapter Coaching website.I will present “AI: Current Trends and Future Directions” for the Mid-Missouri PMI Chapter on November 12th at 7:30am via zoom. Registration will be available on PMI Mid-MO Chapter's website.
Cu, Mark A., and Sebastian Hochman. "Scores of Stanford Students Used ChatGPT on Final Exams, Survey Suggests." The Stanford Daily, January 22, 2023.
Hu, K. (2023, February 2). ChatGPT sets record for fastest-growing user base - analyst note. Reuters News Service. Retrieved August 8, 2024.
Pan, W. H., Chok, M. J., Wong, J. L. S., Shin, Y. X., Yang, Z., Chong, C. Y., Lo, D., & Lim, M. K. (2024). Assessing AI Detectors in Identifying AI-Generated Code: Implications for Education. Proceedings of the 46th International Conference on Software Engineering: Software Engineering Education and Training, 1-11.
Verma, P. (2023, May 18). A professor accused his class of using ChatGPT, putting diplomas in jeopardy. The Washington Post.
Dalalah, D., & Dalalah, O. M. (2023). The false positives and false negatives of generative AI detection tools in education and academic research: The case of ChatGPT. The International Journal of Management Education, 21(3).
Liang, W., Yuksekgonul, M., Mao, Y., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7).
David, E. (2023, July 25). OpenAI can’t tell if something was written by AI after all. The Verge. Retrieved August 8, 2024.
Lidard, Christopher. "If ChatGPT Can Do Our Homework, AI Isn't the Problem." The Daily Princetonian, February 1, 2023.
Excellent article as usual! I am looking forward to hearing your talk at Trulaske on September 13th.
It really is quite silly to try to stop AI usage; the cat is already out of the bag. At a school i attend, Fall semester 2023 policy was basically "using AI is cheating". But the next semester in Spring 2024 the policy became "AI is here to stay and we're giving you instruction on how to use it properly and the pitfalls to look out for." It was a welcome change driven by great leadership that understands the nature of the world.
100% agree that the arms race is a game not worth playing. Much better to try and get at the root of why students don’t trust faculty to give them meaningful work and faculty don’t trust students to do meaningful work. That feels like a rich space for curiosity to do its magic.
I do wonder if there are places where the type of a priori assignments you mention still have a place in a classroom if correctly motivated. For example, I could imagine that just because an LLM can produce text that would reasonably pass for the product we would want, the cognitive process that a human would go through to get there could nonetheless still be valuable, even in the presence of other ways to get there.