Written for the people who decide whether AI gets taught deliberately or handled by accident — principals, superintendents, and school boards.
Developed from conversations with public school principals and built the same way as everything CompAI Chronicles publishes: every claim checked against a named, dated source before it's presented as fact.
Draft for discussion — v3, July 2026. Prepared from an initial brainstorm, then reviewed, fact-checked, and expanded against current research and state policy.
Most schools treat AI as either a threat to police or a shortcut to ban. The alternative here: treat it as a foundational literacy every student needs, taught deliberately from kindergarten on — not handed over unsupervised, and not walled off. We teach math so students can reason about the physical world; we should teach AI so they can reason about the digital one they'll work in for the next 40 years.
The sections below verify what the original research got right, correct what it got wrong, and turn the idea into a working framework — five grade bands, each with an exposure level, core learning, a guardrail, and a sample lesson.
The original brainstorm cited Redlands Unified School District (California) as a working precedent. Before building on that claim, here's what checked out and what didn't — modeling the exact audit habit this framework asks students to practice.
| Claim | Verdict | What we found |
|---|---|---|
| Redlands USD is phasing AI in by grade band. | Confirmed | Real and current — presented to the RUSD board July 21, 2026 by Deputy Superintendent & Chief AI Officer Jason Hill. A proposed framework, not yet adopted board policy — worth saying that precisely to a school board. |
| Grade bands: K–2 → 3–5 → 6–8 → 9–12, with a Level 1–4 access rubric. | Partly wrong | RUSD's actual structure is three bands — K–5 (learn about), 6–8 (learn with), 9–12 (work alongside) — not four. No evidence of a Level 1–4 rubric; that appears to be an invented elaboration. Raw ChatGPT, Copilot, and Gemini are blocked district-wide; only two vetted tools are approved, and parents can opt out entirely. |
| The "Day of AI" video shows Redlands students in hands-on AI activities. | Wrong citation | That video is from Redlands School, a private school in Sydney, Australia — an unrelated institution that happens to share a name with the California district. Dropped from this framework. |
| Teaching the mechanics of AI reduces unhealthy attachment, especially for neurodivergent students. | Supported | Current research (APA Monitor, Oct. 2025; multiple 2025–26 studies) documents real dependency and over-personification risk from anthropomorphic chatbots, particularly among socially vulnerable youth. Teaching that these are statistical systems, not conscious agents, is a published mitigation strategy. |
| A formal curriculum reduces blind trust in AI output better than an outright ban or free-for-all. | Supported | A 2026 intervention study ("Teaching Students to Question the Machine") found AI-literacy instruction measurably improved students' self-regulation of LLM use — the strongest evidence for the whole thesis. |
AI4K12 — a national K-12 AI-education initiative run jointly by the Association for the Advancement of Artificial Intelligence and the Computer Science Teachers Association, active since 2018 — already publishes a research-backed content taxonomy (the "Five Big Ideas": Perception, Representation & Reasoning, Learning, Natural Interaction, Societal Impact). Redlands isn't the only precedent, and it isn't the most academically citable one. Ohio now legally requires every public, community, and STEM school to adopt an AI framework by July 1, 2026. This idea is no longer fringe — it's becoming policy.
| Grade Band | Exposure Level | Primary Emphasis | Key Guardrail |
|---|---|---|---|
| K–2 | No direct tool access | Perception & pattern recognition | De-anthropomorphize: machines follow rules, they don't feel |
| 3–5 | Sandboxed, teacher-run demos | Learning & data bias ("garbage in, garbage out") | Every AI claim gets double-checked against a book or adult |
| 6–8 | Guided, teacher-mediated tools | Representation, reasoning & prompting as logic | Fact-finding still requires a primary-source citation |
| 9–10 | Supervised use, integrity tiers | Natural interaction & societal impact — audit and bias | Raw draft required before any AI-assisted revision |
| 11–12 | Full co-pilot integration | Professional workflow & governance | Human-in-the-loop accountability, logged and cited |
Note: this uses five bands, not Redlands' three. The reasoning and tradeoff are addressed in Open Questions below — a deliberate departure worth discussing, not an oversight.
The full five-lesson Unit 1 for every grade band below — 25 lessons total, built to be handed directly to a teacher.
Available on request while this framework is still in active consulting development.
Exposure level: Zero direct chatbot or tool access. Unplugged games and physical sorting activities only.
Core learning: What is a machine? The difference between living things (which feel and think) and machines (which follow rules a human wrote). Pattern games — sorting shapes, colors, and pictures — build intuition for how computers "guess" what comes next. Vocabulary stays literal: "rule," "guess," "pattern" — no anthropomorphic language ("the computer thinks," "the computer wants") modeled or allowed.
Teachers explicitly and consistently avoid language that gives machines feelings or intentions. This is the single highest-leverage guardrail in the whole framework — the foundation the attachment research above says protects young and neurodivergent learners.
Objective: Students can sort objects into "living" and "machine" and explain one rule a machine follows. Activity: Unplugged sorting game with picture cards (animals, people, machines). Check for understanding: Exit question — "Can a machine feel happy? How do you know?"
Exposure level: "Learn about AI" via sandboxed, teacher-run demos (e.g., Google's Teachable Machine) — no independent chatbot accounts.
Core learning: Garbage in, garbage out — training data determines what a model can and can't recognize. The hallucination rule: AI can sound completely confident while being completely wrong. Double-check requirement: any factual claim from an AI tool must be verified against a book or trusted adult before it's used.
No claim from an AI demo is accepted at face value in any subject — a schoolwide habit, not just an AI-class rule.
Objective: Students can explain why an AI model trained only on apples would mislabel an orange. Activity: Class trains a simple image classifier on a narrow dataset, then tests it outside that set and documents where it fails. Check for understanding: Students predict a failure before testing, then compare to what happened.
Exposure level: "Learn with AI" — teacher-mediated, age-appropriate platforms; no unsupervised open-web chatbot use.
Core learning: How large language models predict text from patterns in training data, at a level students can actually reason about. Structured prompting as applied logic — precise instructions and constraints produce better output, a skill that sharpens writing and problem-solving generally. Privacy and deepfakes: personally identifiable information, digital consent, and how synthetic media can mislead. Executive-function scaffolding: using AI to break a big assignment into steps, with all facts still requiring primary-source citation.
Every fact-based claim, regardless of how it was generated, still needs a primary source. AI can help organize thinking; it does not replace sourcing.
Objective: Students can revise a vague prompt into one with clear constraints and explain why the revision produced a better result. Activity: Given a deliberately vague prompt and its poor output, students rewrite it with explicit constraints, run it, and compare results. Check for understanding: Students annotate which constraint changed which part of the output.
Exposure level: "Work alongside AI" in designated subjects, under explicit integrity tiers (e.g., AI for exploration vs. AI-assisted production).
Core learning: The "spot the fiction" audit — students grade AI-generated content on a historical or scientific topic for errors, bias, and missing context. AI ethics and data law: intellectual property, copyright, environmental cost of data centers, and algorithmic bias in hiring and lending. Mandatory friction: a raw, unassisted draft is required before any AI-assisted editing, establishing a documented baseline of the student's own thinking.
No AI-assisted work is submitted without an unassisted first draft on file. This is the course's academic-integrity backbone, not an optional add-on.
Objective: Students can identify at least two factual errors or biases embedded in an AI-generated summary of a topic they've studied. Activity: Teacher provides an AI-generated summary seeded with 2–3 planted errors; students audit it against primary sources and produce a corrected, cited version. Check for understanding: Students submit an audit log — what was wrong, how they knew, what source corrected it.
Exposure level: Full co-pilot integration, mirroring real industry workflows in relevant career pathways.
Core learning: Domain-specific applications — how AI is used in fields students are heading toward: drafting, diagnostics, code review, design, logistics. Workflow optimization: using AI for boilerplate while retaining full ownership of the final product. The human-in-the-loop standard: professionals remain accountable for AI-assisted work, including legal and ethical liability.
A capstone project requires students to submit their prompt logs and a documented audit trail alongside the final product — not just the output.
Objective: Students can produce a professional-quality deliverable in their pathway area and document how they verified every AI-assisted claim in it. Activity: Capstone project using AI as a co-pilot, submitting prompt logs, an audit record, and the human-refined final product. Check for understanding: Rubric scores the audit trail as heavily as the final product — a flawless output with no verification record does not pass.
A course outline alone won't survive contact with a real district. These gaps show up consistently across every state policy reviewed (Ohio, Vermont, Missouri, Virginia, Maryland) and in Redlands' own rollout.
Redlands appointed a Deputy Superintendent as Chief AI Officer. Whatever the title, someone needs explicit accountability for vendor vetting, policy updates, and coordinating this across departments — otherwise it's everyone's job and no one's.
Virginia and Maryland now require AI-related professional development for teachers by statute. Teachers can't teach discernment they haven't practiced. This should be a line item and a timeline, not an assumption.
Standalone required course, graduation requirement, or embedded across existing subjects — each has different political and staffing costs. Mapping the choice explicitly to AI4K12's Big Ideas (or your state's framework) gives it standing in an accreditation or curriculum-committee review.
Tool licensing, device access, and a genuine opt-out process for parents who decline all need to be resolved up front — an opt-out that quietly creates a second-tier experience will draw the wrong kind of attention.
FERPA/COPPA-compliant vendor contracts belong in place before any tool reaches a K–5 classroom, not after. Easiest item on this list to overlook, most expensive to fix retroactively.
The original brainstorm proposed four units — Mechanics & Limits, Logic & Prompt Engineering, Verification & Audit, and Ethics & Identity. AI4K12's Five Big Ideas describe close to the same territory from a different angle: one is workforce-facing and easy to explain to a school board or a parent on social media; the other is the research taxonomy a curriculum committee will recognize.
Recommendation: keep the four-unit structure as the public-facing course spine — it's clearer to non-specialists — but cite the AI4K12 Big Idea underneath each unit in any board-facing document. Plain language for the pitch, academic credibility for the paperwork, without having to choose one or reinvent the other.
This is a first draft. A few decisions in it are debatable, and better flagged than buried:
Since the working direction leans standalone, the question changes from "what do we invent" to "what do we adopt and customize." Two free, actively maintained curricula cover most of the K–12 span between them:
Free, Creative Commons-licensed, built by MIT RAISE with the Computer Science Teachers Association, aligned to the UNESCO AI Competency Framework. Spiral curriculum across elementary, middle, and high school in 30-to-60-minute blocks; reached roughly 500,000 students since its 2022 launch; includes free optional educator PD. Note: this is the legitimate "Day of AI" — distinct from the Australian-school video wrongly cited in the original brainstorm (see What We Verified above).
Free, no prior experience required, built specifically to function as a for-credit high school course rather than a unit or activity set — which matters if 11–12 needs to carry real transcript weight. Every unit integrates ethics and bias directly rather than as a bolt-on chapter, and it ships with self-paced teacher PD.
AI4K12 Directory (supplemental): a searchable bank of vetted activities mapped to the Five Big Ideas — useful for filling specific gaps in either backbone curriculum rather than as a course in itself.
Caveat: both backbones are scaffolds, not finished courses. Each still needs to be checked against the grade-band guardrails defined above — no direct chatbot access in K–2, mandatory primary-source citation in 3–8, raw-draft-first in 9–10 — and run through your district's normal instructional-materials review before adoption.
The Wild Robot by Peter Brown — a stranded robot on an island; the strongest entry point for talking about what a machine is and isn't, and it doubles as an ELA text. House of Robots by James Patterson & Chris Grabenstein (ages 9–14) — humor-first, works across the upper-elementary/middle-school seam. STEM Starters for Kids: Artificial Intelligence Activity Book — workbook-style supplement.
A few frequently recommended picture/board books turned up repeatedly in review lists but author/edition couldn't be independently confirmed — verify before ordering rather than take the title alone.
House of Robots (see above) still lands well here for reluctant readers. Future AI Expert: A Journey into the Exciting World of Artificial Intelligence for Kids — straightforward expository nonfiction pitched at this band.
You Look Like a Thing and I Love You by Janelle Shane — how AI actually works and where it breaks; pairs directly with the 9–10 "spot the fiction" audit unit. Hello World: Being Human in the Age of Algorithms by Hannah Fry — accessible to strong 9–10 readers, comfortable fit for 11–12. Weapons of Math Destruction by Cathy O'Neil — algorithmic bias in hiring, lending, and education; written for adult readers, treat as excerpts for 11–12 rather than a full assigned read.
Critical AI in K-12 Classrooms by Stephanie Smith Budhai & Marie K. Heath. Putting AI to Work in Disciplinary Literacy by Rachel Karchmer-Klein — grades 6–12 focus, useful even in a standalone model for teachers who'll field AI questions in their own subjects regardless.
This has to be sequenced before the course launches, not run alongside it — a teacher can't teach discernment they haven't practiced. Three tiers, cheapest and fastest first:
Tier 1 — Free entry point: MIT Day of AI's registration includes optional free educator PD on its own materials. ISTE's AI & STEM Network also publishes free online-course-and-PLN resources. Good starting point for every teacher touching the course.
Tier 2 — Structured, evaluable PD: ISTE's AI Explorations program (~2,000 educators and leaders trained to date) and Digital Promise's competency-based micro-credentials — the latter assessed by human experts against real classroom evidence rather than course completion.
Tier 3 — Formal certification (for the designated owner): ISTE's Edtech Teacher, Edtech Leader, and Instructional Leader certifications — appropriate for whoever holds the governance role above, not necessarily every classroom teacher.
Virginia's S.B. 394 ties AI pilot-program funding to required educator PD on AI literacy and responsible use, effective July 1, 2026. Maryland's S.B. 720 requires the state to fund and provide AI professional development for educators and school leaders, requires a state-built rubric for evaluating AI tools, and requires every district to designate an AI coordinator. If building for a district in either state, confirm what's already funded before commissioning anything new.
The companion lesson plans lean on real, named cases as teaching material — pending litigation (The New York Times v. OpenAI and Microsoft), a regulatory dispute that was never fully resolved (the Apple Card / Goldman Sachs credit-limit controversy), and a still-unsettled area of professional liability (AI-assisted radiology misdiagnosis). Those are teaching strengths precisely because they're real and current — which also means they age. Whoever holds the governance role above should own an annual, dated re-check of every named case and statistic before each school year. This should be a standing line item on that role's calendar, not a one-time step done at course launch.
Unit 1, all five bands, 25 lessons total — ready to hand to a teacher.
Available on request while this framework is still in active consulting development.