Research: Can Teachers Identify AI Writing?

Published: 2024-04-25 3:15 PM

Category: Comment | Tags: AI, artificial intelligence, chatgpt, llm, education, assessment, research

An article came across my timeline this afternoon titled "Do teachers spot AI? Evaluating the detectability of AI-generated texts among student essays" (Fleckenstein, J. et. al, 2024) (open access link). I'm in the middle of a large writing project with students where AI has popped up, so I took time to read through it.

The study is split into two parts, evaluating pre-service teachers' (N=89) abilities to spot AI writing as well as experienced teachers (N=200) ability. Both groups also measured their confidence. From the summary,

Preservice teachers were unable to identify the source of the texts correctly, independent of whether they were student-written or AI-generated and independent of the text quality level. Experienced teachers were unable to correctly identify low-quality texts but more successful when it came to high-quality texts. This is at least partly due to the fact that they assigned most of the low-quality texts to being student-written.

Both groups were more confident when they assume texts to be written by students. This finding indicates a certain level of uncertainty when it comes to identifying AI-generated texts that holds true for both novice and experienced teachers.

I know I've been second-guessing a lot more this year and I would consider myself versed in what AI can (and can't) do reliably. I take a little bit of issue in the implication that this disqualifies teachers from being able to identify non-student writing because this is testing the ability of teachers who have no experience with the source material.

I teach the same set of students for 180 days - at this point in the year, I've seen their writing in many different contexts. It isn't difficult to identify "augmented" writing because I know what their authentic work looks like. I would be interested in seeing studies which place writing in front of experienced teachers labelled as student vs AI but for pupils in their classes rather than a general dataset. The context matters.

That said, I'm 100% on board with some of their recommendations based on the results of the study:

Educators may need to rethink their teaching and assessment strategies in light of the availability of AI-based tools. Whenever possible, instead of focusing on reproduction, educators might emphasize skills that AI cannot easily replicate (e.g., critical thinking, literature review).

They also note that AI detection tooling is still really bad at detecting this kind of stuff, so this can't be technologied away. It's going to take education, forethought, and more teaching students about what is - and isn't - okay when it comes to using these tools. Their summary makes the point much better than I can:

In summary, the finding that teachers cannot differentiate between student-written texts and AI-generated texts underscores the need for a thoughtful and ethical integration of AI in education. It calls for a reevaluation of assessment practices, increased awareness of AI's capabilities and limitations, and a focus on student skills that AI cannot easily replace.

It's worth taking time to read.