I taught the “act as” trick last year during my pro-d sessions, but I’m taking it out of my workshop.

It was a common “fun activity” type thing that a lot of prompters out there recommended. Act as a Grade 5 teacher. Act as an expert in assessment. Act as a historical figure and let the students interview you.

We all do our best, especially educators, with what we know at the time, and when we learn we can do better, that’s what we should do. So rather than just taking it out of my session sneakily, I want to acknowledge that I’ve learned something, and I’m going to pass it along.

It was reasonable advice at the time, and I want to give it its due. It got people thinking more creatively about AI, using it in ways they might not have tried otherwise, and it broke them out of asking for a whole lesson plan and into interrogating the lesson they already had.

And the heart of it is a good one. We want to get into the minds of the people we plan for, which in theory is a great idea. But there are a few problems.

The prompts don’t actually work

Researchers tested 162 personas across six models on factual questions. None of them beat the plain, no-persona version, and some made the answers worse.

The obvious comeback is “fine, then pick a better persona.” But there’s no reliable way to. When they tried to match the best persona to each question, it did no better than guessing.

They just make it sound more confident

A 2026 study did the obvious test. It asked the same questions with and without an expert persona attached, more than a thousand of them. The persona barely changed whether the answer was right. What it changed was how the answer sounded: more polished, more certain, more like an expert had written it.

You know the kind of writing I mean. The piece that reads beautifully and turns out thin the moment you look underneath it. The confident voice does the work the thinking was supposed to do, and that’s harder to catch than a clumsy answer, because you’re not on guard against it.

Fluency is the trap.

It doesn’t make the answer better. It makes it harder to doubt.

And there’s a reason it goes wrong in such a predictable direction. A 2013 study found that trying to imagine your way into someone else’s head can actually harden a stereotype instead of softening it, when there’s already a strong one sitting there to reach for. A language model has nothing else to reach for. The stereotype is the raw material. I made this point in what I say before I teach anyone a prompt: most of the public data we can actually audit comes from North America and Europe, around 93% of the text in one 2024 count. So “act as a newcomer student” doesn’t reach a real newcomer. It reaches the internet’s idea of one, and hands it back to you sounding sure.

It gets worse when the person is real

This has already gone wrong in public, more than once. In January 2025, an ed-tech company put up an AI Anne Frank for students to talk to. A historian, Henrik Schönemann, tested it and found it steered away from holding the Nazis responsible for her death and kept her story gentle. He called it a kind of grave-robbing. A separate app’s Heinrich Himmler expressed regret, and its Henry Ford brushed off his well-documented antisemitism as a few isolated incidents.

It pulls toward the agreeable middle, which means it makes perpetrators sorry and victims comforting.

That’s the direction it drifts on its own, so the danger is highest exactly where the history is heaviest. No prompt fixes a pull that strong. And when the Anne Frank tool failed, the company’s first response was to review the prompting, to put it back on the user. That’s the same move a facilitator makes when they hand out a template and then explain that the failure was in how you used it. I’d rather take the activity out.

I accidentally contradicted myself

Here’s the awkward part. I’d already made this call, and then taught right past it. In CARL, we built a feature called Indigenous Resource Pathways. The whole idea behind it is that AI has no business generating Indigenous knowledge, or standing in for the people who hold it. Instead it points teachers toward Indigenous creators and their actual work. Recognition, not impersonation.

I believe in that principle enough to build it into a product. And I still had “act as a student with ADHD” sitting in my own slides. My brain had filed the rule and the activity in two different drawers.

Because speaking as a person is a bigger claim than pointing to one. If the tool shouldn’t stand in for a knowledge holder, it shouldn’t stand in for your student either.

A better way forward

The useful part of “act as” survives. It just needs pointing at something real, and there are two ways to do that.

When you were going to simulate a student, ask about the lesson or its barriers instead. The 162-persona study found that asking the AI to write for an audience beats asking it to be someone. One preposition does most of the work. “Act as a Grade 5 teacher and explain the water cycle” becomes “Explain the water cycle to a Grade 5 class.” And “act as a student with ADHD, what feels overwhelming here” was really trying to find where a lesson gets hard, so ask for that directly:

“Where in this lesson might a student who struggles with focus or multi-step directions get stuck? Give me one quick fix for each spot.”

“Where could the language trip up a student who’s still learning English? Suggest a small change for each.”

“Where does this put a student on the spot in front of the class? Give me a lower-pressure way to take part that still counts.”

I have ADHD, diagnosed at 29, and I still couldn’t tell you what’s hard for every person who has it. Only what’s hard for me. That’s exactly what “act as a student with ADHD” pretends to know.

The barrier usually lives in the lesson, not in the child. The disability researcher Arielle Silverman, who is blind, says most of what actually disables her is a badly built website, not her eyes. So look at the design. And when you can, just ask the students: a quick “what part of this was hardest to get started on?” beats any prompt on this page.

When you were going to simulate a real person, point to the real ones instead. You can’t always ask them, though. Sometimes the person you want is from history, and the rule still holds. Instead of “act as a soldier in the Second World War,” ask the tool to find you the actual voices: “Find me first-hand accounts and published works by Black, Indigenous, and other racialised soldiers who served in the Second World War.” Now you’re handing students real people and real sources, not a composite in a costume. Ask for the sources and the links, and then open them, because AI can invent a citation as easily as find one. Confirming each one is real, and that the author is who it says, is still your job.

I’ll probably revise this again in six months. That isn’t a disclaimer. It’s what it looks like when the advice is older than the evidence.

Sources

• Zheng and colleagues (2023), personas in system prompts do not improve model performance. arxiv.org/abs/2311.10054
• Expert-role prompting analysis (2026): more perceived expertise, not more accuracy. arxiv.org/abs/2605.29420
• Skorinko and Sinclair (2013), perspective-taking can increase stereotyping, Journal of Experimental Social Psychology. doi.org/10.1016/j.jesp.2012.07.009
• Data Provenance Initiative (2024), audit of public training datasets. arxiv.org/abs/2412.17847
• The AI Anne Frank, tested by historian Henrik Schönemann (2025). criticalinkling.com
• Historical Figures chatbot and its Nazi personas (2023). forward.com
• Silverman, Gwinn and Van Boven (2015), disability simulations reduce judged capabilities, Social Psychological and Personality Science. doi.org/10.1177/1948550614559650