The Shift We're In

The problem we're actually solving

Who this is for: Both Lanes (Teachers and Leaders)

Time to implement: 1 week

Big idea: The product is no longer proof. The thinking is the proof.

What you’ll be able to do after this chapter:

• Articulate the difference between AI as assistance and AI as substitution, and explain why that distinction matters more than any policy.

• Teach and use the ACE routine (Ask, Choose, Explain) to build verification habits that transfer across every subject and grade level.

• Establish classroom and school norms that allow responsible AI use with clear boundaries without relying on detection tools or blanket bans.

───────────────────────────────────────────────────

The Problem We’re Actually Solving

I was sitting at my desk on a Tuesday in late October, reading a student’s research paper and I knew something was off before I finished the first paragraph. The prose was cleaner than usual, the argument was more structured, and it sounded nothing like the kid who wrote it.

This wasn’t a student who struggled with writing. He was capable, but his usual work looked nothing like the essay I was reading. His voice was gone. The work was missing that show a teacher that a seventeen year old is wrestling with an idea. What I was reading was competent and correct, but also hollow and empty.

What I thought about as generative AI started to hit high school wasn’t the dishonesty, but rather the question the moment forced me to sit with, what am I actually assessing here?

In all honesty, that question had been on my mind long before anyone was talking about ChatGPT. I’d taught at schools where the unspoken bargain was simple. Students produce polished work, teachers grade the product, everyone moves forward. A well-written essay, a clean lab report, a properly formatted research paper were proof of understanding. If the product looked good, the thinking was good and the student was good.

But I had issues with this model before AI arrived. I remember assigning research papers and receiving work that was technically competent but intellectually vacant before ever hearing about generative AI. I remember the feedback loop that wasn’t a loop at all. Students submitted work, I spent hours grading it, they briefly glanced at the grade, and the paper went into a backpack never to be seen again. The grade was the only thing that seemed to matter. I get embarrassed thinking about assessments I created that tested recall under time pressure and calling it rigor. These weren’t AI problems. These were education problems. AI just made them impossible to ignore.

When generative AI went mainstream, schools responded in roughly two ways. Some banned everything. They blocked the tools on school networks, updated the honor code, invested in detection software, and treated every suspicious submission as a potential violation. Others chose to wait for clearer guidance, which is understandable, but waiting became its own decision as students kept using the tools without any adult modeling. Neither approach solved the underlying problem because neither approach named it. The problem was never the tool. The problem was that we had built an assessment culture where the product was the proof, and now the product could be generated in thirty seconds by anyone with a phone.

Nationally, over half of students and teachers are using AI for school, but only a third of districts are offering students any training on how to use it responsibly (Doss et al., 2025). In high-poverty schools, teachers are significantly less likely to receive AI training or use AI tools at all (Kaufman et al., 2025). Students aren’t waiting for us to figure this out. They’re using the tools now. I believe we have a responsibility to prepare our students for the world they’ll live and work in. As nostalgic as I am for marble notebooks and the way I learned, we have to teach the students who are in front of us about their world. My goal is to protect the learning that matters while embracing the tools that exist.

AI didn’t invent the gap between impressive looking work and actual understanding. It just exposed it. I think that is one of the harder truths some of my colleagues are still grappling with today.

Exposure, while uncomfortable, can be a tremendous gift.

First We Model: A Live Demo

Every chapter includes at least one live demonstration where I model the thinking I’m asking teachers to model for students and this is the first example.

I’m going to ask ChatGPT a question that a student in a Literature class might ask. Then I’m going to do exactly what I want students to learn to do. Test the output, decide what to trust, and explain my reasoning. I’m going to think out loud as I do it as part of the modeling process; making the thought process visible.

---

Setup

Subject: English Literature. Topic: Macbeth. A student might type this into ChatGPT the night before a seminar discussion.

Prompt: “Why does Macbeth decide to kill King Duncan?”

The Output

ChatGPT responds with a four-paragraph answer. It names ambition as the primary driver, describes Lady Macbeth’s persuasion, references the witches’ prophecy as the catalyst, and concludes that Macbeth’s fatal flaw is unchecked ambition. It reads like a well organized study guide.

The Critique (Thinking Out Loud)

I project this output in front of the class and I start talking through what I notice.

“Okay. First thing, this is organized and it’s not wrong. Ambition matters. Lady Macbeth matters. The witches matter. But look at how the AI frames it. Macbeth decides to kill Duncan because of ambition, period. That is a little too simplistic for us. Shakespeare spends an entire soliloquy showing us a man who is arguing against killing Duncan. He lists reasons not to do it and nearly talks himself out of it. The play is interesting in part because the decision is complex. The AI gave us what, but skipped the how, the hesitation, the humanity, the fact that Macbeth knows what he’s about to do is wrong and does it anyway. That’s where Shakespeare lives, and the AI missed some of that.”

“The second thing I notice is how absolute this sounds. There’s no “but” or “another perspective,” no ‘scholars interpret this differently’ or ‘one reading suggests.’ It reads like the final word. That’s part of how these tools can work. They generate text that sounds authoritative and correct whether or not it is authoritative. We can’t see exactly how the tool arrived at its answer. We can look back at some of the “thinking,” but usually we can only see the answer. That means we need to do some additional verification and ask follow up questions.”

“The AI also says Lady Macbeth persuades Macbeth. That’s one reading. But it isn’t the only reading. We talked about this a little in class when we read aloud. Another reading is that Lady Macbeth doesn’t persuade him so much as shames him by attacking his masculinity. ‘When you durst do it, then you were a man.’ Is that persuasion or manipulation? The distinction matters because it changes how we understand both characters and what Shakespeare is saying about power and relationships. The AI flattened a complicated human dynamic into a sentence. This is especially important for reading Shakespeare because of how important dynamic human characters are in Shakespeare’s plays.”

The Verification Step

I go back to the text. I pull up Act 1, Scene 7 and read aloud Macbeth’s soliloquy. He says Duncan ‘hath borne his faculties so meek’ and that killing him would be an act against a guest, a kinsman, and a king. MacBeth isn’t just hesitating. He’s building a case against committing the crime. A student raises his hand and references the ‘dagger of the mind’ in Act 2. In a seminar, we discussed how the hallucination tells the reader something about guilt and self-awareness that a simple ‘ambition drove him’ summary over-simplifies and the AI left out of its report. This is the kind of gap students need to learn to spot.

Debrief: What Students Should Notice

After a demonstration like this, I ask students three things. What did the AI get right? Where did it flatten or oversimplify? And what’s missing? Those three questions are the seeds of the types of routines and assessment adaptations that structure this entire book.

What to Do with Your Students on Monday

“I’m going to show you something I asked ChatGPT. I want you to watch how I evaluate what it gives me because this is the skill that matters. The AI can generate answers. Your job is to decide how useful those answers are, when they can improve or expand, and to explain why.”

The Core Idea, Made Simple

The shift isn’t entirely about AI. It’s also about thinking.

Every classroom interaction, every assignment, every discussion, every assessment, is a decision point.

Whose mind is doing the work? If a student can complete a task by copying an AI output without understanding, interrogating, or building on it, then the task isn’t assessing thinking. It’s assessing access to a tool and every student has access to the tool now.

For too long, the polished product was the evidence. A well-written essay, a clean lab report, or a properly formatted research paper were reliable proxies for understanding. Teachers graded the product and inferred the thinking.

Fortunately, that model has broken. The product is no longer proof of anything. A student can generate a polished essay in thirty seconds without understanding the first thing about the topic. The essay that looked like thinking in 2021 doesn’t prove anything in 2026.

The response to that broken model isn’t catching students using an app. It’s teaching them routines and processes to think more deeply and not outsource their learning.

One such routine is something I call ACE.

The ACE Routine

This routine works the same way whether a student is checking a ChatGPT summary, evaluating a Google result, or reviewing a classmate’s claim in a seminar discussion. AI tools are not required.

Ask. Before you trust anything, interrogate it. What was produced? Does it match what you already know? Where can you verify this and have you? Asking isn’t a single question, but a mindset. You’re training students to approach output the way a good editor approaches a draft, ie, with curiosity, skepticism, and a red pen.

Choose. Once you’ve interrogated the output, make a decision. What do you keep? What do you revise? What do you reject entirely? This is the step most students skip when left to their own devices. They copy, paste, and submit. The Choose step is where the thinking happens; it’s where the student stops being a user and starts being a thinker.

Explain. Show your reasoning. Why did you keep that claim and cut that one? What evidence informed your decision? This is the step that makes the invisible visible and gives teachers a window into student thinking. If you can’t explain why you trusted something, you didn’t actually evaluate it.

Notice that these are only three steps and nothing fancy. There is no technology required and you can use these steps in a myriad of ways and formats. A student can ACE a library source, a peer’s argument, or a ChatGPT response. The tool changes. The habit doesn’t.

If You Only Remember One Thing

The question isn’t Did the student use AI? The question is Can the student explain what they did with it and why? If the answer is yes, we’re moving the student in a positive learning direction. If the answer is no, we’re not. ACE makes helps make that distinction visible.

Classroom Moves (Teacher Lane)

Here are some practices that are doable tomorrow and they can change what students produce in your class.

Move 1: Build in micro-decision moments

What you do: Add 3-5 minute choice checkpoints throughout a task, not just at the end.

When: During any multi-step assignment (essay, lab report, project, problem set)

Why: You want to see decisions while they’re being made, not just after everything’s polished.

Best for: Any class where students complete work outside of your direct observation

Prep time: 5 minutes to adjust an existing assignment

What students produce:

• A “draft zero” claim before they write anything formal

• A circled strategy choice with a one-sentence “why”

• A two-minute voice memo explaining their thinking

• A revision note: “I changed X to Y because...”

Common failure mode: You add the checkpoint but don’t actually use it; it becomes one more thing to grade instead of formative data you respond to.

Quick fix: Pick one decision point per unit and conference with 3-5 students about it. That’s enough to change the culture. The most important thing is that it is an aspect of the learning process that you are passionate about and do not want your students to lose sight of.

Move 2: Require source grounding (not just citations)

What you do: Students must explicitly connect claims to texts/data/sources we used together in class.

When: Any time students are building an argument, explanation, or interpretation

Why: AI can generate generic claims. It can’t reference the specific discussion we had Thursday or the data from our lab.

Best for: ELA, social studies, science, any claim, evidence, research writing

Prep time: 10 minutes to revise a rubric or add a reflection prompt

What students produce:

• “Two sources minimum + one sentence on why each is trustworthy”

• Quote + paraphrase + “So what?” annotation

• “What the source says / what it doesn’t say”

Common failure mode: Students cite sources but don’t actually engage with them; they paste a quote and move on. They can use AI to generate this work if they are not with you.

Quick fix: Add one oral check-in question: “Which source did you trust more, and what made it trustworthy?” or “What did the source not say?” If learning is the thing, it’s ok that they are using these tools, but we need to teach them speed bumps so they are not using them mindlessly.

Move 3: Teach the “unreliable classmate” protocol

What you do: Explicitly teach students to treat AI like a smart but unreliable peer.

When: The first time AI is allowed in your class

Why: Students need a mental model for how to use AI responsibly. Without it, they might either avoid it entirely or trust it blindly.

Best for: Any class where AI use is permitted

Prep time: 15 minutes to model it once (then students own it)

What students produce:

• ACE check note: “What I asked / What I checked / What I chose / Why I can explain it”

• Highlighted (I love asking students to annotate and highlight) sections: useful (green), suspicious (yellow), wrong (red)

• A one-sentence decision statement: “I kept this because...” or “I changed this because...”

Protocol (student-friendly):

1. Ask AI for an explanation, draft, or approach

2. Check it and highlight what seems useful, circle what seems vague or suspicious

3. Choose what to keep using class resources (text, notes, data, primary sources)

4. Explain your choices and rewrite a version you can defend

Common failure mode: Students complete the protocol as a compliance task without actually thinking.

Quick fix: Randomly select 2-3 check notes to discuss with the whole class: “What did you catch? What did you change? What made you suspicious?”

---

Leader Moves (Leader Lane)

These moves are for principals, APs, instructional coaches, department chairs and anyone responsible for creating conditions where teachers can actually do this work. I’ve been in your positions and I understand the challenges.

Move 1: Name the shared aim across the building

What you do: Establish one schoolwide principle that grounds all AI conversations.

Owner: Principal and instructional leadership team (every needs to be unified)

Timeline: This week (seriously—don’t wait for a task force to draft a 12-page policy)

Why: If every teacher invents a different standard, students will treat AI expectations like a guessing game. A shared aim reduces chaos without requiring legal language.

The shared aim (use this or adapt it): “At [School Name], we look for decision points, the moments where students make choices that reveal their thinking. We assess those choices, not just polished final work.”

Evidence it’s working:

• Teachers can articulate what decision points look like in their discipline

• Students can name how decision points are used in various classes

• Conversations shift from “Is this cheating?” to “Show me where you chose”

Move 2: Protect time for task redesign

What you do: Give departments protected time to redesign common assessments for visible thinking.

Owner: Principals and department chairs

Timeline: One PLC cycle per quarter (2-3 meetings minimum)

Why: Teachers know their current tasks are vulnerable, but redesigning takes time. If you ask them to “just figure it out,” they’ll revert to what they know and they will do the work they need to get off their plates immediately, like grading.

Evidence it’s working:

• At least one common assessment per department includes multiple visible decision points (not just a final product)

• Teachers report feeling less anxious about AI because tasks are harder to outsource

• PLC time is focused on instruction, not detection

How to structure the time:

1. Meeting 1: Identify “protected thinking” for your discipline (What choices can’t be outsourced?)

2. Meeting 2: Redesign one task using the Decision Points Menu (see Copy/Paste Pack)

3. Meeting 3: Pilot, debrief, adjust

Move 3: Build a culture where verification is normal

What you do: Model “verify before you trust” in leadership communication and decision-making.

Owner: All leaders

Timeline: Ongoing (start Monday)

Why: If your culture becomes “gotcha,” students learn secrecy. When your culture becomes “show your thinking,” students learn honesty. Leaders set that tone.

How to do it:

• In staff meetings, model your own checking process: “I asked AI for data on X, but then I checked it against Y, chose what to trust, and here’s why...”

• When students or teachers are accused of AI misuse, lead with curiosity: “Help me understand your process. What did you ask? What did you check? Can you explain your choices?”

• Share “good catches” publicly: “A teacher noticed a claim that didn’t match our sources…this is what they did.”

Evidence it’s working:

• Students report feeling safe to disclose AI use

• Teachers report fewer adversarial conversations

• The default question shifts from “Did you cheat?” to “Show me your thinking and explain

Assessment and Evidence

What you can assess in this chapter is foundational. You’re not grading the quality of AI use. You’re assessing whether students can interrogate an output, make a decision about it, and explain that decision.

What “Good” Looks Like

ACE Step

Developing

Proficient

Ask

Accepts output at face value or identifies only surface-level issues (“it looks right”).

Identifies specific claims to verify, notes gaps or missing perspectives, checks at least one claim against an independent source.

Choose

Copies output without meaningful revision, or makes only cosmetic edits.

Keeps, revises, or rejects specific elements with a clear rationale. The final product reflects the student’s judgment, not just the AI’s output.

Explain

Cannot articulate why they trusted or rejected specific elements (“I just thought it was good”).

Explains reasoning with reference to evidence: “I kept this because I verified it in ___, but I cut that because ___.

Conferencing Prompt

Pull a student aside and ask: What did the AI give you, what did you do with it, and why?” That’s it. You’re not grading. You’re not investigating. You’re just listening. What you hear will tell you more about that student’s thinking than any written log.

Make It Real

I don’t think we are in a technology shift so much as we are in a formation shift. The tools will keep changing. The companies will keep releasing new models. The detection arms race will keep escalating and keep failing. None of that changes the fundamental work. The fundamental work is the same as it has always been. We are building young people who can sit with complexity and not reach for the first clean answer. Who can think for themselves, not because a rubric demands it, but because the world will demand it every day of their lives.

That was the job before AI and it is still the job.

Students cannot opt out of an AI saturated world. They are going to use these tools at work, at home, in every corner of their adult lives. The question is whether anyone showed them how to think critically about what those tools produce or whether they learned, through years of silence and avoidance, that the adults in their lives didn’t know how to talk about it either. That’s not a technology failure. That’s a formation failure. And formation is what we do.

The Bastani study found that the same tool harms learning without guardrails and supports learning with them. That’s important to remember. The tool didn’t change. The guardrails did. The guardrails are us, the teachers who slow down and think out loud, the leaders who build cultures of authenticity rather than cultures of compliance, the schools that decide their students deserve more than a policy memo and hope.

First we model.

Try It Tomorrow

Generate one AI response in your content area. Project it. Think out loud for five minutes about what’s right, what’s wrong, and what’s missing. That’s it. One live model. Five minutes.

Try It This Week

Introduce an ACE routine that works for you to one class. Post the three steps. Ask students to annotate one assignment with a brief ACE log. Read five of them. Notice what you learn about their thinking.

Try It This Month

Leaders, draft and share a one page guardrails stance with your staff.

Teachers, bring a live model moment to a department meeting and debrief with colleagues.

Everyone, start a conversation about the difference between assistance and substitution in your building.

The question isn’t whether your school has an AI policy. The question is whether your school has a shared understanding of what thinking looks like and whether you’re modeling it.

─────────────────────────────────────────────────────

References

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabiṙ, Ö., & Reis, R. (2025). Generative AI can harm learning. Proceedings of the National Academy of Sciences, 122(2). https://doi.org/10.1073/pnas.2422633122

Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), 1–21. https://doi.org/10.1145/3411764.3445721

Doss, C. J., Steiner, E. D., Woo, A., & Zuo, G. (2025). AI use in schools is quickly increasing but guidance lags behind. RAND Corporation. https://www.rand.org/pubs/research_reports/RRA4180-1.html

Kaufman, J. H., Woo, A., Steiner, E. D., & Doss, C. J. (2025). Uneven adoption of AI tools among U.S. teachers and principals. RAND Corporation. https://www.rand.org/pubs/research_reports/RRA134-25.html

Klein, A. (2024, September). Black students are more likely to be falsely accused of using AI to cheat. Education Week. https://www.edweek.org/technology/black-students-are-more-likely-to-be-falsely-accused-of-using-ai-to-cheat/2024/09

Lee, S., et al. (2025). The impact of generative AI on critical thinking. Microsoft Research. https://www.microsoft.com/en-us/research/wp-content/uploads/2025/01/lee_2025_ai_critical_thinking_survey.pdf