We Didn't Ban AI. We Started Asking Better Questions.

I’m a high school English teacher, and one of the questions I’ve been wrestling with is what assessment should look like when a student can ask AI to produce a polished answer in seconds. 

One answer for us has been surprisingly old-fashioned: talk to the student.

Our students complete one-on-one verbal assessments with a teacher to demonstrate that they understand their core content standards in our personalized, competency-based learning option called Empower[Ed]. The permeation of AI chatbots created a real friction point for us, especially in our Social Studies courses. How could students use AI to prepare for those conversations without turning AI into a machine that simply tells them what to say?

So I built BrianGPT.

Brian is an AI pre-assessment coach with one important limitation: it is intentionally designed not to give students a polished answer.

A student starts by identifying the U.S. History standard they are preparing to assess. Brian then asks the student to explain what they think the standard means in their own words. If the answer is vague, Brian doesn't immediately rescue them with an explanation. It asks questions, points them toward specific historical evidence they need to investigate, and makes them try again.

From there, Brian conducts a miniature version of the assessment. Students have to answer questions using specific historical evidence, respond to a more challenging inquiry question, and practice explaining their thinking aloud. Brian gives them a strength and a next step, but it won't write the speech for them.

In fact, I gave it an explicit guardrail for students who try the obvious shortcut and ask, "Just give me what to say":

"I can't provide a full script. Your teacher needs to hear your thinking in your own words. Let's build it together."

The part of this experiment that interests me most isn't really the GPT. It's what designing the GPT forced me to think about.

For years, educational technology has often promised to make things easier, faster, and more efficient. Generative AI can certainly check those boxes. But learning sometimes requires exactly the opposite. Students need to struggle with an idea, retrieve evidence, explain a connection, hear themselves say something that doesn't quite make sense, and try again. I started teaching around the time Angela Duckworth’s Grit reminded us to cultivate perseverance in young people. A decade later, we’re teaching in an educational ecosystem where students can outsource the struggle, the very thing we were trying to teach them to persist through, with a single prompt.

So I found myself deliberately building friction into an AI tool.

Brian knows the standards students are expected to demonstrate. It can help them practice. It can ask another question when their evidence is weak. But ultimately, it has to hand the thinking back to the student.

And at the end, Brian doesn't give them a grade. When they've demonstrated the standard, used evidence, tackled an inquiry question, and practiced speaking, it tells them they're ready to schedule the real conversation with a human teacher.

That's probably my AI story right now: I started by wondering how AI could help students prepare for assessments, and I ended up building an AI whose most important role is knowing when not to help.