Analyzing AI Images

The following materials are meant to be used as part of a lesson on how to Analyze and spot AI images. Unfortunately, this is a quickly adapting space and models are getting better at output, so feel free to supplement this with more recent developments. In my experience, it’s taken about two 80 minute periods to get through the lesson and activities.

What this lesson does

This lesson asks students to slow down and look critically at AI-generated images — not just to spot technical glitches, but to ask harder questions about what AI defaults reveal about the data it learned from. The technical and the critical are sequenced intentionally: students learn how LLMs work before they’re asked to analyze what that means. It still includes some tips and tricks for identifying AI generated images.

Before you start

The bias image sets were generated using straightforward prompts with no manipulation “a teacher,” “a mechanic,” “a CEO.” Gender is never specified. What students see is what the model defaulted to. That framing is important to share with students. This isn’t about bad actors or extreme cases. It’s about what “normal” looks like when a system learns from the internet as it actually exists. Images were generated in December 2025, using ChatGPT, Google Gemini, and some random sites I found googling for AI image generation.

Before the lesson: what to prepare

Print in advance:

• “Things to check when you look at a photo” handout; one per student for future reference
• Analysis worksheet; one per pair
• Bias image sets (both pages); one per pair, printed in colour if possible; the detail matters for the analysis

Display:

• The slide presentation — open it in presentation mode before students arrive and stay on the title slide until you’re ready to begin

Optional:

• If you’re running the practice tools as a class activity rather than independent extension, test both links ahead of time and have them open in separate tabs

Sequence

Hook (5–10 min)

Before the presentation, give students 60–90 seconds to look at the image sets and write down what they notice. No framing, no prompts from you. Just: what do you see? Collect those observations and set them aside — you’ll return to them at the end.

Presentation (15–20 min)

The slides cover what an LLM is, how it works, and how bias shows up in its outputs. Really reinforce that LLMs are pattern learners, not thinkers. This is worth spending time on, especially the distinction between predicting language and understanding it. The bias slides connect directly to what students are about to analyze, so don’t rush past them.

If your students already have some AI exposure, consider running the LLM slides as a Socratic check rather than direct instruction. What do they already know? What are they less sure about?

Handout (5 min)

Distribute the “Things to check” sheet and walk through it briefly. Students don’t need to memorize it, they need to have it in hand and know it exists. It’s worth pointing out that the checklist covers both technical markers (hands, shadows, text rendering) and contextual ones (gut feeling, reverse image search). But you’ll know if your class is ready for this.

Worksheet (25–30 min)

Students work in pairs. The worksheet asks them to guess the prompt for each image before analyzing specific ones for bias. The built-in stop instruction — “Do not answer any more questions until you correct #1” — is one of the most important moves in this lesson. It requires students to commit to a guess before they know the answer, which makes the bias conversation much more honest. Let that moment land. The discomfort when students recognize what they assumed is the lesson.

Assign specific images for each pair to analyze. If you want to debrief as a class, consider splitting the six images across pairs so every image gets at least one pair’s attention — then share out before discussion.

Discussion (15–20 min)

Return to what students wrote in the hook. What do they see now that they have language for it? Some questions to open the room:

• The AI generated these images because it learned from human-made images. What does that tell us about those images?
• Is this the AI’s fault? The programmer’s? The internet’s? Does it matter who we assign it to?
• If you were going to prompt an AI for “a doctor” and wanted to push back against what it defaults to, what would you have to do and why should that work be on you?
• What’s the difference between an image that reflects bias and one that spreads it?

Optional extension (10 min)

Two linked practice tools are included in the presentation for pairs who want to keep testing their eye independently.

Differentiation

The worksheet’s open observational structure works across levels — pairing strategically here can help, as students who struggle with reading can contribute meaningfully through observation and discussion while their partner handles the written response. For pairs ready for more, push question 5b further: the move from surface observation (“the teacher is a man”) to structural critique (“why does the model default to that, and who is affected when it does?”) is where the real thinking lives.