AI Reading Investigation

Some readings are hard because they’re long, dense, or packed with unfamiliar context, and students will reach for AI whether you invite them to or not. So invite them, and shape what happens next. Instead of asking for a summary, students work through a sequence of prompts that treats AI as a smart but unreliable study partner. It helps them get oriented, name the argument, find the passages worth rereading, and try out interpretations. Then they go back to the article and check.

The Setup

Pick an article with a clear question that is longer or denser than it needs to be, the kind students skim and misremember. Frame one or two central questions the reading answers. In the version I ran, the article was Steven Shapin’s “Trusting George Cheyne,” and the question was simple: why did people trust this person, and what does that reveal about how expertise worked?

Before opening AI, students jot down two or three quick notes: what they think the article is doing, and what seems confusing.

Then the prompt sequence, which students can adapt as they go. Each prompt does a different job:

1. Get oriented — the question, the argument, how the author uses evidence. Explicitly not a flat summary of topics.
2. Map the central problem — a table of the factors the author identifies (in my case, sources of Cheyne’s credibility), what evidence in the article supports each, and what the model is uncertain about.
3. Clarify key concepts — plain definitions of the article’s key terms, why they mattered in context, and a question to ask on returning to the text.
4. Find evidence — a checklist of passages to reread and what to look for in each. A reading guide, not a summary.
5. Test interpretations — three possible readings, what supports and complicates each, and which is most interesting.
6. Connect to course themes — specific, non-obvious claims a student could make in discussion.
7. Sharpen your own angle — the student pastes their own rough idea, and AI suggests ways to sharpen or complicate it, without writing it. For each suggestion, what would need verifying in the article?

The post (~400 words, written by the student) is organized around five points: one useful AI insight, one AI claim they verified in the article, one AI claim that was vague, wrong, or needed correction, one passage they understand better because of the exchange, and their own answer to the central question.

The Prompt

The first prompt sets the tone for the sequence. Adapt the article, the course, and the emphasis:

Why It Works

The sequence models a better habit than the one students arrive with. Their default, “summarize this,” flattens the article and gives them no reason to open it. Every prompt here points back into the text (a passage to reread, a claim to check, a piece of evidence to find), so the AI conversation becomes a trail of leads instead of a substitute.

The post makes verification the deliverable. Students name one claim they confirmed and one they had to correct, so they can’t finish without going back to the article. Hunting for a vague or wrong claim teaches them what AI smooths out: tone, hedging, the texture of the evidence.

The last prompt flips the usual roles. The student brings the idea and AI plays critic, which keeps the reflection theirs while still putting AI to work at the writing stage.

What to Grade

The grade rests on the movement between AI and the article: what the student checked, corrected, and understood better.

• Strong: careful engagement with the article, thoughtful use of AI, specific passages cited, and a clear answer to the central question.
• Middling: uses AI productively and refers to the article, but could be more specific, better verified, or more clearly connected to course themes.
• Weak: leans on AI summary, gives little evidence from the article, or doesn’t explain what the student verified for themselves.

What to Watch For

• Students skip the notes-before-AI step, which is what lets them notice how the AI changed their reading. Ask for those notes in the post.
• Posts drift into summary of the AI conversation. The five required points exist to prevent that; grade against them.
• The “claim I had to correct” can be manufactured: a trivial quibble offered to satisfy the requirement. Reward the students who explain why the AI version was too generic, not just that it was.
• The sequence is long. For a shorter version, orientation, evidence-finding, and interpretation-testing (prompts 1, 4, and 5) do most of the work.

Using This Sketch

Print

This sketch is licensed under Creative Commons BY-NC-SA 4.0. You're welcome to share and adapt it for noncommercial purposes, with credit, as long as you share your adaptation under the same license.

Image Rembrandt, A Scholar Seated at a Desk, 1634. National Gallery Prague. The image's own license applies to it, separately from the sketch.

Suggested citation

Fred Gibbs. "AI Reading Investigation." AI Sketchbook, Amaranth, University of New Mexico, September 25, 2026. https://amaranth.unm.edu/ai-sketchbook/teaching/ai-reading-investigation.html. Licensed under CC BY-NC-SA 4.0.

Related Sketches

activity · lightly tested How Else Could This Look? What did this article decide to be about, and what did that decision cost? assignment · tested AI as Second Opinion What does AI see in a primary source, and what does it smooth over?