AI’s potential in Special Ed? A whole new front?

Article link

Hype rating

AI usage

Who benefits?

There are a several audiences that are depicted as benefitting but I’m not sure I agree based on what’s in the article.

• students – spend more time evaluating and organizing thoughts, avoid getting stuck generating text from scratch, help with executive function
• teachers – avoid stuck students becoming behavior problems, generate custom material quickly
• administrators – increase the speed of data analysis and report creation

The one clear (unmentioned) beneficiary would be the company selling access to the AI.

What concerns does this bring up?

 

Students

With the student examples, I end up with lots of questions. I don’t know if there’s cognitive benefit in that struggle to shift context, to move beyond writer’s block (whatever the cause). AI allows us to get a very quick and very specific, first draft. It might be the “bad first draft” that’s been encouraged as part of writing  processes for a long time. I couldn’t find any actual research on whether this is a evidence-based pattern or not. I also wonder what happens when the “bad first draft” is actually pretty good and if AI keeps improving, what happens when it’s really good?

This doesn’t have to be AI-focused research. It’s really about the impact of struggle in writing and whether different levels of first draft material have any impact on learning. AI makes it possible to generate customized starter content in ways we couldn’t easily do before but the analysis is on the writing process’s impact on learning (regardless of where the starter content comes from). Maybe this kind of research has been done.

If you want shift that analysis to AI, looking at a co-writing process where the student and AI co-create  a final work would involve both student AI competencies (How do you use AI effectively to do this?) and the functional capabilities of the AI (How well can the AI do what it’s supposed to do?). You could then evaluate the student based on an external assessment and compare it to the paper as the assessment of learning. Does the paper indicate they know more than they know on the external assessment? Now do the same with a non-AI paper. How does it compare?

It’s difficult to know where productive struggle lies in the educational context and that’s made more complex in a special education situation. In a scenario where both the teacher and the student are using new technology in new ways there is lots of room for mistakes.

Teachers

I was probably most disappointed in the example used in the article where AI was used to create a infographic comparing animal and plant cell organelles. The claim is that it would have taken 10-15 minutes to find an example on the internet while AI generated it in only 15 seconds (net savings between 9 mins 45 seconds and 14 minutes 45 seconds). I easily found any number of Venn diagrams on this topic on my first search (Venn diagram animal plant cell organelles). I also know that text generation in images is a place where AI is going to do poorly. Since the exact prompt was quoted in the article,  I used it in a paid version of ChatGPT and also saw multiple examples of gibberish created by a paid version of MidJourney. It’s possible there was some other AI involved here that did a good job but it’s highly suspect. What’s frustrating is that I agree with the premise. AI can create some really useful content and save teachers a chunk of time. I’d focus more on differentiated reading levels, content to analyze, images for attention/visual metaphors, etc.

Administration

This example was given for assessment (kind of evading the usual way educators would talk about assessment in my opinion). The focus was on a district using spreadsheets to collect data and create budget and staffing projects related to special ed. They use a GPT called “data analyst” to create a “a compelling narrative with graphics to support staffing decisions” compressing 3 months of work into 3 days.

“Data analyst” is likely this custom GPT. It’s one of the built-in options from OpenAI. The only description is “Drop in any files and I can help analyze and visualize your data.”  It has “Code Interpreter & Data Analysis” capabilities. You can see more about ChatGPT’s data analysis capabilities here. Much of it is using Python to create visualizations based on text prompts.

I really struggle to see how this can save 27 days of work. I know it’s overselling the capability of AI to create “compelling narrative with graphics to support staffing decisions.” At best, you’re generating the two things separately and combining manually.  In many ways I also question if the productive struggle with data gets lost in this process. What gets missed given AI’s desire to please? What kinds of misrepresentation might occur even with the best of intentions? You’ll often hear “person in the loop” as a key aspect of AI verification. With something like this, I find it highly unlikely that the data will be verified, especially if it’s complex and difficult to process.

Additional comments

With AI, educators could do quite a bit without it really mattering if it’s AI (at least on a practical level). We already know quite a bit about the effectiveness of multimodal representation, differentiated content, etc. I’d spend a lot more time there before worrying about using AI with students, using AI to make staffing predictions, etc.