The Major Obstacle to Detecting AI Writing

Generative artificial intelligence (AI), is undoubtedly the biggest ongoing story, both in the ethical conversation around plagiarism and the legal conversation around copyright.

It is, without a doubt, the hottest topic in this space.

However, there’s a key problem with it. Determining what is or is not AI writing, is nearly impossible. That is, at least with any degree of certainty.

This issue has been especially acute in academia, where the past few months have seen a groundswell of activity in this space.

First, nearly every major company that does plagiarism detection in this space has launched some form of AI detection tool. This includes Turnitin, Copyleaks and Plagiarism Check have all introduced tools to detect AI writing. 

They’ve been joined by a collection of start-ups including (but certainly not limited to) GPTZero, Originality.AI and ZeroGPT. However, the quality of these services has varied wildly, with Zero GPT recently flagging an article of mine as AI writing, though it was entirely human-written. 

And it’s there that the problem emerges. The space of detecting AI writing is still very new and very uncertain. It’s unclear which products are effective and which are essentially snake oil.

Research is underway in that space, but this is a very dynamic field, both in terms of AIs available and the tools to detect them. Research and testing can only provide a snapshot of the environment as it exists at that moment.

However, even if we set aside the issue of quality, there’s a bigger one looming: Even with the best information available, it can be impossible to act on it. In short, how is one sure that a work was written by AI. 

The Problem of Probability

Anyone who has performed an AI check likely noticed that such checks don’t work the same as a traditional plagiarism checker. 

With a plagiarism checker, one gets highlighted passages where the examiner then clicks through to the results. There, they do a human comparison of the passages to see if the match is actually an indication of plagiarism. Humans, ultimately, have to draw the conclusion of whether a work is plagiarized or not.

AI detection doesn’t work that way. There’s no source to check, and the process of making the decision is completely opaque. 

As a result of this, AI detection tools usually present some indication of the confidence in their findings. For example, Copyleaks and ZeroGPT use a percentage-based system where the closer the number is to 100%, the more sure the findings.

Others, such as GPTZero and Plagiarism Check, label such as “likely” or “highly likely” to indicate its certainty. 

Turnitin, on the other hand, actually does say something is or is not AI generated and claims a false positive rate of less than 1%. However, they further claim that they err on the side of missing AI-generated work and may miss up to 15% of AI-generated content.

But even if Turnitin’s claim is accurate, there is still a probability issue. Even if they are 99% accurate, that means that 1% are still false positives and that there is no way to separate the those from the actual AI works.

In short, since there’s no way for humans to verify the results, false positives are a very real threat. No matter how good the systems get, it’s always a game of probabilities, and those probabilities are never fully known due to the shifting nature of the technology involved.

As a result, though we can detect AI writing fairly well, acting on that detection is difficult, at best.

Some Reasons for Hope

While this seems like a pretty dire situation, there are some reasons to hope.

First, the technology is going to improve. While AI writing is a major evolution, if one looks at the past 25 years in this space, plagiarism detection has always been a game of cat and mouse. The pendulum always swings both ways.

What this new technology or approach will look like is difficult to say, but it will likely involve a combination of both technological and human-driven approaches. For example, an authorship program, such as Turnitin’s can be paired with AI detection to provide greater certainty of the results. Also, teachers can combine technological solutions with different assignment types and in-class elements to virtually eliminate any chance of content being generated by AI.

However, a recent article on Plagiarism Check recently highlighted another issue. According to their research, their plagiarism detection tool was able to detect AI-generated content a significant percentage of the time without the aid of AI detection. 

This is because AI systems often times don’t generate wholly original text and, instead, modify existing text. An advanced plagiarism detection tool that is designed to detect rewritten text can likely detect these overlaps pretty easily and provide either another layer of proof or another way to detect AI-generated text.

To be clear, this isn’t consistent enough to rely upon, but it is consistent enough that it should give pause to any students hoping to use AI as a shortcut.

While these might not seem like major hopes, they are reasons to not believe that the sky is falling. Things are going to change and there are challenges ahead, but there are always opportunities.

Bottom Line

In the end, things are going to change, especially in academia. We will need to have conversations about when and how it is acceptable to use AI and how to incorporate generated text into human-written material.

However, it’s difficult to have those conversations when there is no reliable way to determine what is or is not AI generated. How does one set a boundary on AI writing when there is no way to enforce it?

Hopefully, the technology and processes will grow and improve. As they do, I expect we’ll see some nuance and balance return to these conversations as the panic begins to subside.

There’s not much doubt that generative AI writing will have very significant impacts on academia, but it probably won’t be quite the upheaval some seem to fear it will become.