Stop trying to detect computer-generated writing.

It’s a risky distraction that just encourages an arms race.

Ken Arnold


January 17, 2023

Many people saw the quality of AI-generated text and immediately thought that they need to find ways to detect such “cheating”. I respect those trying to make those tools (including my former colleagues who made GLTR back in 2019, and more recently Edward Tian, who is “trying to save us from ChatGPT before it changes writing forever (NPR)” with GPTZero), and the educators feeling like they’re forced into using them—at the present moment this might be a best-of-bad-options situation. But we need to stop it. Reasons:

  1. It’s going to have false positives. There will be text that’s actually written by a thoughtful human but gets flagged as machine-generated. The author will get accused of cheating and will probably have no recourse. Even if they successfully counter the accusation, it’s discouraging — especially if classmates who used ChatGPT a bit more cleverly were able to sneak actual computer-generated text past the system (a false negative). And although I don’t have specific evidence in this situation, the false positives of other AI risk assessment systems tend to fall disproportionately on those already marginalized.

  2. It’s just going to lead to an arms race. We’ve already seen examples of how simple tweaks to how the system is used (e.g., telling it to introduce mistakes) can sneak past some ways of detection (some examples I found without trying hard: 1, 2, 3). If there’s a clear signal of computer generation, the bad actors will find a way around it. Meanwhile the false positive rate will go up (see #1).

  3. It’s distracting and unproductive for most common purposes of writing. For writing-as-art, it might be productive to detect forgeries. But when writing is used to communicate, think, and learn, this focus on evaluating the output distracts from the emphasis we need to have on the process. Others have written more about this, and I’ll post some more thoughts myself soon.

Post-script 2023-02-09*: OpenAI released their own Classifier for AI-written text. They acknowledge the first two of the above points.