You may have seen recent reports of OpenAI’s 99.9% accurate AI text detector, and subsequently been wondering if this is the end of the arms race building between universities and students. After all, if we had a tool that could reliably tell us when students are cheating, we could just ignore this whole AI-thing and get back to business as usual.
But, it’s a lot more complicated than this single variable suggests.
First of all, OpenAI’s tool is aimed at identifying ChatGPT-generated text, rather than AI-generated text. So even if the 99.9% figure is accurate, all this will do is drive students to another model, of which there are more and more every day. Leading us back to the arms race dynamics I’m keen to avoid.
Secondly, it may be the case that AI detectors are ‘simply’ measuring statistical features in a piece of text (i.e. burstiness and perplexity), which can be modified as part of the output. I’m not sure if you’d need to dial these features up or down in order to confuse an AI detector, but simply knowing which variables are being targeted by the system is a good start for anyone wanting to avoid detection. And of course, you can also just pass AI-generated output through other filters, add your own errors, or use any of a range of other adversarial techniques, and it will likely avoid detection.
Finally, ‘accuracy’ as a concept is not as simple as it seems. Philip Dawson makes the point clearly in his short post, providing an example of an AI-detector that is 100% accurate, but effectively useless. He goes on to explain how the specific details around calculating the percentage of accuracy says a lot about how much we can trust it (and also, what decisions we can reasonably make as a result).
So, even if OpenAI really does have an AI detector that’s 99.9% accurate, it’s still not the silver bullet I know many people are hoping for.
For myself, I’d prefer to see us working towards a world where:
- Students don’t feel the need to cheat at all (I recognise that this is complicated but I don’t think we should shy away from problems just because they’re hard). And, related to this…
- Any discussion of cheating includes the recognition that it’s a social issue, and not a technology issue. In other words, we’re not going to solve the ‘cheating’ problem with a new technology.
- Using AI is a required part of the curriculum, rather than something we try to regulate.