A corgi crossed with a Highland cow. A sheep crossed with a dog. A pig-sheep hybrid. A giraffe with stripes instead of spots. My head of science showed all of these to his students last week, and every one of them was fake. AI-generated. Completely impossible.
The students did not know that. Not yet.
The lesson that almost didn't work
The idea started as a simple conversation. My head of science was planning a genetics and selective breeding lesson, and he wanted to push the concept further than usual. The first half of the lesson stayed grounded in real biology. Real animals that exist because humans have pushed breeding in very particular directions. The absurdly muscular Belgian Blue cow. The domestic dog breeds that look nothing like their wolf ancestors. Examples where you can clearly see selective pressure at work, and where everything is still biologically possible.
Then the question shifted. What happens when you push past what biology actually allows? Where does the logic break down?
That's where AI came in. We generated images of animals that could not exist naturally. The students debated them. Some were obviously absurd. Others were surprisingly plausible. One, a leopon (male leopard, female lioness), complicated things further because it is rare but genuinely real. That slowed the conversation down, because it forced everyone to think harder about what they were actually looking at.
Here is what mattered: we did not tell the students the images were AI-generated. Not at the start.
Why the delay was deliberate
If we had opened with "these are AI-generated images," the lesson would have died immediately. Students would have looked at each image through a single lens: spot the fake. The conversation about genetics, selective pressure, and biological limits would never have happened. They would have been performing scepticism rather than practising it.
Instead, we let them react. Let them argue about what was possible and what was not. Let them bring their existing knowledge of biology into the conversation. The images were a provocation, not a test. And provocations only work when they are not pre-explained.
The reveal came after the discussion had done its work. We told them clearly and explicitly: these images were AI-generated. Not as a gotcha. As the point.
Transparency was always there. It was just timed.
What happened after the reveal
This is where the real learning started. In a science lesson, we suddenly had a conversation about AI literacy. Not bolted on. Not a separate lesson. A natural extension of what they had just experienced.
The students had already committed to opinions. They had already argued that some of these animals could exist. And now they had to reconcile that with the fact that an AI system had generated every image. The emotional investment in the argument is what made the reveal land. They cared about being right. And they had just been shown that their instincts about visual evidence were unreliable.
That led to questions I could not have manufactured with a slide deck.
- How do you tell an AI image from a real one?
- Why did some of these look more convincing than others?
- If AI can fool me with an animal picture, what else can it fool me with?
The students arrived at the critical thinking. I did not have to drag them there.
What timed transparency actually means
Timed transparency is not the same as deception. Deception withholds the truth to manipulate. Timed transparency withholds the reveal to create the conditions for deeper understanding, and then makes the truth explicit.
The distinction matters, because if you get it wrong, you erode trust. Students need to know that when you show them something in a classroom, there is a pedagogical reason. They need to trust that the reveal will come, and that it will be honest.
Three conditions make timed transparency work.
- The delay has a clear educational purpose. You are not withholding information to be clever. You are creating space for genuine inquiry before the answer collapses the question.
- The reveal is explicit, not implied. You say it clearly. "These images were AI-generated." No ambiguity. No hoping they will figure it out.
- The reflection follows the reveal. The reveal is not the end of the lesson. It is the beginning of the important part. What did you assume? Why? What does that tell you about how you evaluate visual information?
If any of those three conditions is missing, you are not using timed transparency. You are just withholding information.
The tools matter less than you think
A side note on tools, because this always comes up. We tried Gemini first. It refused to generate the images, flagging them as "fake." My head of science joked that we should try Grok because "Grok will do anything." We used ChatGPT Teams in the end.
The tool choice is not the lesson. The lesson is that convincing visual content is now trivially easy to produce, and students need to learn how to evaluate what they see. Which tool generates the images is a detail. How students respond to those images is the point.
Why AI literacy works best when embedded
I keep seeing schools treat AI literacy as a standalone topic. A dedicated lesson. A one-off assembly. A PSHE session on deepfakes.
The problem with standalone AI literacy is that it teaches students to be sceptical in the AI literacy lesson, and then they go back to accepting AI output without question in every other subject. The scepticism stays in the box where it was taught.
What worked about this science lesson is that AI literacy was embedded in the subject. The students were not learning about AI. They were learning about genetics, and AI literacy emerged as a necessary part of that learning. The critical thinking was not a separate skill to practise. It was the only way to make sense of what was in front of them.
The best AI literacy lesson does not look like an AI literacy lesson. It looks like a science lesson, an English lesson, or a history lesson where AI made the learning more demanding, not easier.
How to plan a timed transparency lesson
If you want to try this in your own classroom, here is a simple structure that works across subjects.
1. Choose content where AI can produce plausible but flawed output. This could be images (as we used), text, data, or even code. The key is that the output should be good enough to be believable, but contain errors or impossibilities that students can identify once they know what to look for.
2. Present the AI output without labelling it as AI-generated. Let students engage with it on its merits. Ask questions that require them to evaluate the content using their subject knowledge.
3. Reveal the source explicitly. Tell students the content was AI-generated. Do not make them guess. Do not treat it as a game.
4. Facilitate the reflection. This is the most important step. Ask: What did you assume? What made this convincing? How would you check next time? What does this tell you about content you encounter outside this classroom?
5. Connect it back to the subject. The AI literacy is the through-line, but the subject knowledge is the anchor. In our lesson, the reflection circled back to what is and is not biologically possible. The science was stronger because the AI literacy sharpened it.
The real risk is not the deepfake
The striking thing about our lesson was not that AI-generated images fooled some students. It was how quickly students accepted visual evidence that confirmed what they expected to see. The corgi-cow hybrid looked plausible enough because the students had just spent twenty minutes looking at real examples of unusual breeding outcomes. Their brains were primed to accept the next image as another data point.
That is not an AI problem. That is a confirmation bias problem. And it is exactly the kind of thinking that AI literacy, properly embedded, can surface and challenge.
Your students encounter AI-generated content every day, on social media, in search results, in the tools they use for homework. A one-off lesson about deepfakes will not prepare them. What will prepare them is repeated practice, embedded across subjects, at recognising when something looks right but might not be.
Timed transparency is one tool for building that practice. Use it when the delay serves the learning. Be explicit when the reveal comes. And always, always follow up with the question that matters: what made you believe it?
Matthew Wemyss is an AIGP-certified AI in Education consultant and practising school leader. Book a discovery call to discuss AI literacy training for your school.