In late December 2022, I wrote about my first reactions to ChatGPT and urged educators to engage with the technology rather than ban it. Now, roughly ten weeks later, I want to report on what happened when I moved from theory to practice - when I actually brought AI tools into my classroom and asked students to use them as part of their learning.

The short version: some experiments produced genuine pedagogical value. Others failed in ways I did not anticipate. And the experience surfaced a set of implementation questions that I believe every teacher integrating AI will eventually face.

The Experiments

I designed four distinct AI integration activities during January and early February 2023, each targeting a different instructional purpose.

Experiment 1: AI as a brainstorming partner. I asked students to use ChatGPT as the first step in a research project. Before doing any traditional research, students spent 15 minutes conversing with ChatGPT about their topic, asking questions, exploring subtopics, and generating a preliminary outline. They then printed their conversation and annotated it - marking claims that needed verification, identifying interesting angles they had not considered, and noting where the AI's responses were vague or generic.

This worked remarkably well. Students reported that the brainstorming conversation helped them clarify their thinking before confronting the overwhelming volume of a Google search. The annotation exercise - marking what needed verification - turned out to be an excellent critical thinking activity. Students were practicing source evaluation skills applied to AI-generated content, which is arguably more relevant to their future information environment than evaluating traditional web sources.

Experiment 2: AI-generated text as a revision mentor. I had students write a first draft of an argumentative essay, then paste it into ChatGPT with the prompt: "Identify the three weakest points in this argument and suggest how to strengthen them." Students then used the AI's feedback to guide their revision, ultimately writing a reflection on which suggestions they accepted, which they rejected, and why.

This produced mixed results. Approximately half the students engaged meaningfully with the AI's feedback, treating it as one perspective among several (alongside peer feedback and my own comments). These students produced stronger final drafts and, more importantly, demonstrated sophisticated metacognitive reasoning in their reflections. The other half, however, treated the AI's suggestions as authoritative instructions, accepting every recommendation without critical evaluation. This uncritical deference to the AI's judgment concerned me and became a focal point for subsequent class discussions.

Experiment 3: AI-generated content as an object of critical analysis. I asked ChatGPT to write an essay on a topic we had studied extensively in class, then distributed the AI-generated essay to students for critical analysis. Their task was to evaluate it using the same rubric I would use for their own writing: thesis clarity, evidence quality, logical coherence, counterargument engagement, and writing mechanics.

This was the most successful experiment. Students were highly engaged - there is something inherently motivating about being asked to grade a machine's work. More importantly, the exercise produced sophisticated analytical conversations. Students identified that the AI essay was superficially competent but substantively thin - it made claims without specific evidence, relied on generalities rather than concrete examples, and treated complex debates as though they had simple resolutions. One student observed that "it sounds smart but doesn't actually say anything" - a critique that captures the fundamental limitation of large language models with remarkable precision.

Experiment 4: AI as a differentiation tool for teacher preparation. This experiment was not student-facing. I used ChatGPT to generate differentiated versions of reading materials - taking a grade-level text and asking the AI to produce versions at lower and higher reading levels while preserving the core content. I also used it to generate additional practice problems in mathematics at varying difficulty levels.

This was useful but required significant quality control. The AI-generated differentiated texts were a reasonable starting point, but each version needed careful review. In several instances, the "simplified" version omitted conceptually important information rather than expressing it more simply - a distinction that matters enormously for instructional quality. The math problems were more reliable, though I caught occasional errors in the solutions the AI generated. The time savings were real - perhaps 40-50% reduction in differentiation preparation time - but the idea that AI can automate differentiation without teacher oversight proved premature.

What I Learned About Implementation

Beyond the specific experiments, several broader implementation lessons emerged.

Access equity is an immediate problem. ChatGPT requires internet access and a device. While my school provides Chromebooks, students who wanted to use ChatGPT outside of class needed personal devices and home internet. This meant that AI-enhanced learning activities risked widening the digital divide unless they were exclusively conducted during class time. OpenAI's terms of service also technically require users to be 18 or older (or 13 with parental consent), which creates a compliance challenge that most schools have not yet addressed.

Students' prior conceptions about AI vary enormously. Some students arrived with sophisticated understanding of what large language models are and how they work. Others believed ChatGPT was "looking things up on the internet" or "thinking like a person." These misconceptions directly affected how students used the tool and how critically they evaluated its outputs. Any AI integration must include explicit instruction about what the technology actually is - a statistical text generation system trained on patterns in data, not a knowledge retrieval system or a thinking entity.

The speed of AI output disrupts classroom pacing. When a student can generate 500 words of text in five seconds, the temporal structure of a writing lesson changes fundamentally. Students who used AI for brainstorming finished that phase much faster than I anticipated, creating a pacing mismatch with the rest of the lesson. I found myself needing to redesign lesson timelines - allocating more time for the analytical and evaluative phases that follow AI interaction, and less time for the initial generation phase.

Emotional responses from students were complex. Some students were excited and energized by AI tools. Others expressed a form of anxiety I had not anticipated: "Why am I learning to write if a computer can do it better?" This existential question - about the purpose of learning skills that machines can perform - deserves a serious answer, not dismissal. I found myself having philosophical conversations with 14-year-olds about the nature of thinking, the relationship between process and product, and what it means to develop a skill even when automation exists. These were among the most intellectually rich conversations of the semester.

Preliminary Principles

Based on these first experiments, I am developing a set of preliminary principles for AI integration in K-12 classrooms.

Use AI to raise cognitive demand, not lower it. The worst use of AI in education is as a shortcut that reduces the amount of thinking students do. The best use is as a tool that enables higher-order thinking activities that were previously impractical. Having students critically analyze AI-generated text is more cognitively demanding than having them write a generic essay. Having students fact-check an AI conversation is more demanding than having them conduct a standard web search.

Make the AI's limitations a central part of instruction. Every AI-integrated lesson should include explicit attention to what the AI gets wrong, what it cannot do, and why. This is not just good pedagogy; it is essential digital literacy for a generation that will live and work alongside AI systems.

Document everything. We are in the earliest stages of understanding how AI affects learning. Every teacher experimenting with AI integration should be documenting what they try, what happens, and what they learn. We need a shared evidence base, and right now, the only way to build it is through practitioner research.

I am ten weeks into what I expect will be a multi-year journey of experimentation. The technology is evolving rapidly - GPT-4 is rumored to be imminent, and competing systems from Google and others are on the horizon. The only certainty is that standing still is not an option.

Dr. Janette Camacho is a K-12 educator with 28+ years of classroom experience, a Google Certified Educator (Level 1 & 2), Adobe Creative Educator, and Apple Teacher. She is documenting her ongoing experiments with AI integration in K-12 instruction.