How to Use AI in Schools: A Practical Walkthrough

The Adoption Gap: Students and Teachers Are Using AI While Schools Scramble for Rules

By May 2025, 84% of U.S. high school students were using generative AI for schoolwork, up from 79% just four months earlier, according to College Board research. Teachers have kept pace: 60% of K-12 educators used AI tools during the 2024-25 school year. Yet the institutions around them have not.

Only 18% of principals reported that their school or district had provided any AI guidance, per TeachAI, and just 26 states had issued AI guidance as of April 2025. The problem is not a shortage of tools — it's that governance is lagging far behind adoption.

The equity implications are stark. Only 13% of principals in high-poverty schools reported receiving AI guidance, compared with 25% in more affluent areas. Where formal rules are missing, policy fragments in predictable ways:

• Roughly 2 in 5 schools and districts don't allow student GenAI use at all
• About 1 in 5 allow use with no policy whatsoever
• 16% delegate AI policy to individual teachers, 12% to departments
• Only 13% actively encourage use across all classes

That last gap creates a real contradiction: students are using AI almost universally while many institutions officially prohibit it. Meanwhile, detection tool use among K-12 teachers jumped from 38% to 68% in a single school year — even though accuracy drops when students paraphrase AI-generated content. Schools are policing behavior they haven't defined.

TeachAI's central premise is that the lack of structured guidance, not the lack of tools, is the primary obstacle. As AI becomes embedded as a built-in feature in everyday software — including agentic tools that act autonomously, like scheduling assistants — that gap widens. Common Sense Media warns that without a clear, structured plan, AI integration becomes "fragmented, inequitable, and even misaligned with your district's priorities."

For administrators building that plan, the same verification discipline matters at the tool level. When evaluating AI agents and the skills that extend them, it helps to know which components are actually safe to deploy — research has found roughly 26% of published agent skills carry at least one vulnerability. That's the problem SkillzBench is building for: a free directory that verifies, security-scans, and benchmarks agent skills, with a free MCP connector to equip agents only with what you've chosen. It's currently in pre-launch, with a waitlist open for educators and administrators who want verified tools before students ever touch them.

The takeaway for school leaders: adoption has already happened. The question is whether governance catches up before the equity gap does.

Start With the Problem, Not the Tool: A Needs-First AI Vision for Your Classroom

Most schools don't have an AI problem — they have a sequencing problem. With only 18% of U.S. principals reporting that their schools or districts provide AI guidance, educators are improvising while students race ahead: 84% of high schoolers already use generative AI for schoolwork. The fix isn't picking a better tool. It's defining the problem first.

That's exactly what UNESCO recommends: "Start with defining the problem, not the algorithm." Rather than asking "which chatbot should we buy," ask "where do teachers lose hours each week?" or "where do students struggle to get feedback?" The problem statement, not the product demo, should drive every decision that follows.

Once the problem is clear, three frameworks converge on a practical sequence. Common Sense Media's district toolkit warns that "without a clear, structured plan, AI integration can become fragmented, inequitable, and even misaligned with your district's priorities" — so it starts with a readiness assessment. TeachAI's roadmap (a coalition led by Code.org, ETS, ISTE, and Khan Academy) lays out the full path:

• Define a vision — what learning problems AI should (and shouldn't) touch
• Establish principles — transparency, privacy, equity, human oversight
• Review existing policy — academic integrity, data protection, accessibility
• Fund professional development so staff can actually use what's approved
• Build a feedback loop to revise as tools and evidence evolve

TeachAI also draws a distinction most districts miss: guidance is flexible, frequently updated advice; policy is durable and long-lasting. Guidance can change monthly as tools shift; policy shouldn't. Schools that skip this distinction end up either frozen in outdated rules or improvising with no rules at all — the current reality, where roughly 2 in 5 schools don't allow student AI use while 1 in 5 allow it with no policy at all.

Stakeholder input comes before approval, not after. TeachAI is explicit that teachers, parents, students, and legal counsel should all shape the plan. College Board's research lead Jessica Howell echoes this: meaningful impact "requires listening to students, parents, and educators, and grounding solutions in evidence."

For districts eyeing agentic tools — autonomous assistants that schedule, draft, and act — the same needs-first logic applies. SkillzBench is building a free directory that verifies, security-scans, and benchmarks agent skills before deployment, so schools can evaluate what actually helps rather than trusting marketing claims. Define the problem, then test the solution.

Deploy AI Where It Pays Off First: Teacher Workloads

Teachers spend hours on tasks that AI handles well today — and the data shows they're already doing it. Among U.S. K-12 teachers who used AI tools during the 2024-25 school year, the most common applications were preparing to teach (20%) and administrative work (18%). That's your roadmap: start where the workload is heaviest and the stakes for students are lowest.

Michigan Virtual's educator guide maps AI across ten educator roles, which makes it easy to pick first tasks. The highest-value starting points for most teachers:

• Creating and curating materials — lesson plans, worksheets, reading passages, and rubrics drafted in minutes, then edited by the teacher
• Differentiation — generating the same content at multiple reading levels, plus special education support scaffolds
• Assessment and data analysis — question banks, draft feedback, and pattern-spotting in student performance
• Parent communication and tutoring support — drafting newsletters and producing practice problems on demand

Before any of this reaches students, apply MIT Sloan's rule: try before you teach. Test every tool thoroughly yourself, treat any AI-generated content as needing validation, and never enter confidential student data into public tools — once data goes in, it's often impossible to remove. Michigan Virtual echoes this with its guardrails of transparency, FERPA/COPPA privacy compliance, and bias awareness.

Resist the urge to adopt a different tool for every task. MIT Sloan notes that a single platform can support a wide range of tasks — the goal is using AI flexibly toward instructional goals, not accumulating subscriptions. This consolidation matters more as TeachAI's taxonomy of predictive, generative, and agentic AI blurs: agentic tools that act autonomously (scheduling assistants, workflow automation) are increasingly embedded as built-in features, which makes careful vetting harder, not easier.

For teams exploring that agentic layer, the same try-before-you-teach logic applies: verify and test before deployment. SkillzBench is building a free directory of agent skills — with security scanning, license checks, and sandboxed benchmarking — plus a free MCP connector, so educators and administrators can evaluate what actually helps before it touches a classroom. Whatever you deploy, keep Michigan Virtual's framing front and center: AI is an aid, never a substitute for human instruction.

Classroom Safeguards: Privacy, Transparency, Bias, and the Detection Trap

The same year schools finally started talking about AI, teacher use of AI detection tools jumped from 38% to 68% — a sprint toward surveillance that the research suggests solves the wrong problem. Before any AI tool touches a classroom, four safeguards need to be in place.

Michigan Virtual's teacher guide distills these into three ethical pillars: transparency, privacy compliance, and bias awareness. Transparency means telling students exactly how AI is being used — in lesson prep, grading assistance, or feedback — rather than hiding it. Privacy is the hard legal line: educators must comply with FERPA and COPPA, and MIT Sloan's EdTech team is blunt about the stakes, warning that once data is entered into most AI systems, it's challenging—if not impossible—to remove it. That means confidential student records, grades, and IEP details never belong in public AI tools, and students uncomfortable sharing data should always have an alternative.

• Never enter confidential student data into public AI tools — removal is effectively impossible once it's in.
• Be transparent with students about when and how AI shapes their instruction and feedback.
• Check AI-generated content for bias before it reaches students, since models reflect the skew of their training data.
• Validate everything. Treat every AI output as a draft requiring human review — "try before you teach."

On detection, the numbers tell a cautionary story. Turnitin has reviewed more than 200 million papers since 2023 and found about 11% contained at least 20% AI-generated text. But detection accuracy drops sharply when students paraphrase AI content, which means a rising score is a conversation starter, not a verdict. With 84% of high school students now using generative AI for schoolwork, according to College Board research, accusation-first approaches will misfire constantly.

The better investment is integrity training. Research cited in the same surveys shows students already distinguish between writing an entire paper with AI (a major violation) and using it for brainstorming or editing (less severe) — and that integrity training measurably improves their judgment about acceptable use. Teaching students where the lines are works; catching them after the fact mostly doesn't.

As schools begin exploring agentic AI — tools that act autonomously, like scheduling assistants — the same safeguards apply, plus a verification step. SkillzBench is building a free directory where educators can review, security-scan, and benchmark agent skills before anything reaches a classroom, with a waitlist now open. Safeguards first, tools second: that ordering is what separates meaningful integration from fragmented adoption.

From AI Literacy to Agentic AI: Building a Sustainable AI Practice at School

The schools that get lasting value from AI are the ones that treat it as a capability to build, not a tool to buy. With 84% of high school students already using generative AI for schoolwork and 60% of teachers using AI tools during the 2024-25 school year, adoption is no longer the question — sustainability is.

SRI International's research is blunt about what separates surface-level experimentation from meaningful integration: professional development, administrative support, and clear curriculum standards. That means AI literacy budgets, not just AI tool budgets. Teachers need structured training on where AI fits across their roles — from lesson prep and differentiation to parent communication — and students need explicit instruction on acceptable use, since integrity training measurably improves their judgment about what crosses the line.

Equity deserves equal attention. Adoption data shows ELA and science teachers are almost twice as likely to use AI tools as math or general elementary teachers, so any rollout should include targeted support for lagging departments. And as MIT Sloan's guidance warns, assignments should never disproportionately benefit students who can pay for expensive premium tools — design tasks that work with free tiers or school-provided access.

The roadmap, in short:

• Fund AI literacy and ongoing professional development for teachers and students, backed by visible administrative support
• Set clear curriculum standards so individual departments aren't improvising policy alone
• Audit adoption gaps across subjects and close them with role-specific training
• Design assignments that don't favor students with paid AI subscriptions
• Prepare now for agentic AI — autonomous tools like scheduling assistants that act on their own, not just generate text

That last point matters more each semester. TeachAI's taxonomy already treats agentic AI as a distinct category entering schools, and agents introduce a new evaluation problem: the skills and connectors that extend them vary wildly in quality and safety. Before any agent touches a classroom workflow, someone should verify what it's actually being told to do. That's the gap SkillzBench is being built to close — a free, independent directory of agent skills with security scanning, version pinning, and sandbox benchmarking, plus a free MCP connector so administrators can browse, verify, and test skills before deployment. It's in pre-launch now, with the waitlist open for educators and technology leads who want vetted agent tooling from day one.

Sustainable AI practice isn't a finish line. It's a cycle of training, policy, equity checks, and tool vetting that keeps pace with what students and teachers are already doing — whether the institution has caught up or not.