Pangram claims to be a highly accurate AI detector with a false positive rate of 1 in 10,000. Let's take this at face value and see what it means.

The claimed false positive rate (the chance of incorrectly detecting human-written text as AI-generated) seems very impressive. Definitely an improvement over the first generation of AI detectors. So how useful is Pangram? Let's take a concrete application: is it a viable solution to the problem of college students using AI in violation of course policies?

Suppose every instructor started using an AI detector on all student submissions. I'd estimate that students submit 500 – 1,000 written works in the course of a 4 year education (!) — 30+ courses X ~5 assessments per course X many independent problems per assessment. If each of these were run through an AI detector with a FPR of 1 / 10,000, you'd have 5–10% of your student body falsely accused of cheating at some point during their time at university.

So now you have three options:

* Continue to treat cheating as the serious violation that it is, and initiate disciplinary proceedings whenever the AI detector flags suspected cheating. I hope it's obvious that this is not really viable. Even if we assume that most innocent students will be exonerated, the anxiety and wasted time is unfathomable.

* Apply a small penalty instead of treating it as a serious violation. This normalizes cheating and is likely to backfire.

* Use AI detection as only one signal and gather additional evidence of integrity violation (Pangram itself recommends this). But the problem is that all of the ways of doing this that I'm aware of either don't work, or can only be done once you've already initiated disciplinary proceedings, which brings you back to option 1.

There are many other downsides to the systematic use of AI detection.

* Students who know what they are doing can easily evade AI detection by paraphrasing their text either manually or using automated tools. If Pangram (or any other specific tool) starts to be adopted on a much bigger scale, the evasion tools will be incentivized to get better as well, specifically by training on the outputs of Pangram.

* While simply offloading an exercise to AI of course fails to achieve learning goals, depending on the course and activity, there may be many healthy ways to use AI. Use of AI detection will make students uncomfortable with using any of these, since they likely increase the risk of false positives.

If instructors treat AI as the problem, I doubt there is any solution. The actual problem is that our testing practices aren't that effective at assessing student mastery and engagement with learning. We should look to alternative assessment practices such as complementing written work with oral exams and assignment sequences where the student builds on their work throughout the semester. Once I started deploying these, I realized that they bring pedagogical benefits far beyond AI detection!

This is not to say that AI detectors are useless. Pangram published a recent analysis of concerning levels of AI use in ICLR reviews. This is a good application of AI detection because it isn't about accusing individuals but about the aggregate. It doesn't require a very low false positive rate in order to be useful.