Should artificial intelligence AI be used in schools?

AI Is Already in Every Classroom — But Most Schools Have No Plan for It

Walk into almost any classroom in America and AI is already there — whether the school approved it or not. During the 2024–25 school year, 85% of teachers and 86% of students used AI, with student adoption jumping 26% and educator adoption 21% in a single year. The question is no longer whether AI belongs in schools; it's who is actually steering it.

The answer, disturbingly often, is nobody. As of December 2024, only 31% of U.S. public schools had a written AI policy, and 56% of students and educators say their institution is unprepared to manage AI responsibly. Adoption has sprinted ahead of governance, leaving teachers and administrators to improvise rules on the fly — one classroom at a time.

That improvisation gap has consequences. When policy is absent, vendors fill the vacuum. USC researcher Stephen Aguilar describes the current moment bluntly: it's "a gold rush" in which edtech companies sell districts products "without having any data to support their claims." The cautionary tale is Los Angeles Unified's multi-million-dollar "Ed" chatbot, which collapsed within months when its vendor's finances did — leaving a major district scrambling.

The pattern is familiar to anyone who has watched organizations adopt AI tools without vetting them, and it shows up in three predictable ways:

• Tools get purchased on marketing promises rather than evidence of learning outcomes
• Teachers receive no training, so usage varies wildly between classrooms
• Students set their own informal rules, from homework help to outright misuse
• Districts have no exit criteria, so failed tools linger on contracts long after they've stopped helping

The scale of the market makes this urgency worse, not better. The global AI-in-education market is projected to grow from $7.05 billion in 2025 to nearly $137 billion by 2035, and Microsoft alone committed over $4 billion to AI education initiatives in July 2025. When that much money is chasing school budgets, "trust us" is not an evaluation strategy.

Schools that succeed will be the ones that borrow a page from rigorous software evaluation: demand evidence before adoption, pilot with clear exit criteria, and treat every tool as something to verify, not something to believe in. That verification-first mindset is exactly what we're building at SkillzBench — free benchmarking, security scanning, and trust scores for AI agent skills, so organizations can test what actually works before it touches their workflow. The same discipline belongs in the principal's office as much as the server room.

Governance, not technology, is now the bottleneck. The schools that write real policies — with vendor vetting, bias audits, and annual tool reviews — will capture AI's genuine benefits. The rest will keep improvising, one collapsed chatbot at a time.

The Case For AI in Schools: Real, Measured Benefits

Teachers using AI weekly save an average of 5.9 hours per week — the equivalent of six full school weeks reclaimed every year, according to aggregated education research. That is not a marginal efficiency gain; it is time returned to lesson planning, one-on-one support, and the human work that only educators can do.

The strongest evidence comes from a peer-reviewed randomized controlled trial published in Scientific Reports in June 2025. Students working with AI tutors outperformed peers in traditional in-class learning with effect sizes of 0.73 to 1.3 standard deviations — and they did it in less time, with a median session length of 49 minutes versus 60 minutes for conventional instruction (Engageli's research roundup). In education research, effects of that magnitude are rare.

Practitioners back the numbers. Kim Lepre, a seventh-grade English teacher, put it plainly: "It has just taken a load off of the little minute things that I have to do so that I can just focus on teaching the kids." And 69% of teachers say AI has improved their teaching methods, while 59% report it enabled more personalized instruction.

The broader evidence base reinforces the pattern:

• Systematic reviews covering 80+ causal studies found positive effects of student-facing AI tools on engagement, learning behaviors, and cognitive outcomes (IES)
• In an Indiana pilot, just over half of surveyed teachers reported a positive or very positive impact from an AI tutoring platform on both teaching practice and student learning
• A Massachusetts AI curriculum pilot reached 1,600 students across 30 school districts

But the data contains a critical nuance that separates real gains from hype. The OECD's Digital Education Outlook 2026 found that students using general-purpose chatbots produced higher-quality work — yet that advantage "disappeared and sometimes reversed in exams when access was removed." Only tools designed with intentional pedagogical purpose showed sustained improvements.

This distinction matters enormously for schools weighing implementation. The question is not simply "AI or no AI" but which AI, designed for what purpose, with what evidence. As the Institute of Education Sciences concludes, AI works best as a complement to — not a replacement for — the essential work of educators.

That verification-first logic extends beyond the classroom. SkillzBench is building free benchmarking, security scanning, and trust scoring for AI agent skills so that any organization — schools included — can test what actually works before deploying it, rather than trusting vendor claims.

The case for AI in schools, then, is real and measurable — but conditional on choosing tools that teach, not just answer.

The Case Against (Or At Least For Caution): Overreliance, Bias, and Privacy Risks

None of this means AI doesn't belong in schools — but it does mean the risks are real, documented, and unevenly distributed. The faculty concerns alone are striking: surveys of college faculty found 95% expect generative AI to increase student overreliance, 90% expect it to diminish critical thinking, and 83% predict shorter attention spans. Meanwhile, 73% report personally handling AI-related academic integrity cases.

The most troubling evidence concerns fairness. APA reporting found that Black teenagers were about twice as likely as their peers to have schoolwork incorrectly flagged as AI-generated — a serious problem when detection tools are fallible and consequences fall on students. Privacy raises its own alarms: teens are sharing sensitive topics with chatbots operated by companies that can collect that information.

Institutional data confirms these aren't hypothetical worries:

• One in four respondents to a UNESCO survey of 400 institutions across 90 countries reported encountering AI-related ethical issues, including student overreliance, authorship disputes, and bias in research.
• OECD research found that students using general-purpose chatbots produced higher-quality work, but the advantage "disappeared and sometimes reversed in exams when access was removed" — a warning sign for genuine learning.
• Equity gaps are widening: low-income and rural students are least likely to attend schools that allow or integrate AI, and generative AI use stands at 24.7% of the working-age population in the Global North versus 14.1% in the Global South.

There's also a marketplace problem. USC researcher Stephen Aguilar describes an edtech "gold rush" in which companies sell districts products "without having any data to support their claims." The cautionary tale is LA Unified's multi-million-dollar "Ed" chatbot, which collapsed with its vendor after only a few months.

Yet the answer to these risks is design and policy, not prohibition. Initial district bans in Baltimore, Los Angeles, New York City, and Seattle have given way to adaptation strategies — Australian science students now present findings orally in real time, and the NEA's model policy calls for piloting tools with a commitment to discontinue them if benefits aren't demonstrated. That verification-first mindset matters well beyond the classroom: SkillzBench is building free benchmarking and security scanning for agent skills on the same principle — test what actually works before you deploy it.

The bottom line: the documented risks of overreliance, bias, and privacy demand intentional pedagogical design, formal governance, and evidence-based vetting. Schools that skip those steps aren't avoiding AI's downsides — they're simply absorbing them unprepared.

How Schools Should Implement AI: A Five-Step, Evidence-First Playbook

The evidence is clear: 85% of teachers and 86% of students already use AI, yet only 31% of U.S. public schools had a written AI policy as of December 2024, according to adoption research. The gap between usage and governance is where most problems begin. Here's how schools can close it.

Step 1: Adopt a formal policy before scaling anything. The NEA's model school-board policy offers a ready-made framework covering equitable access, algorithmic bias audits, data privacy compliance, vendor vetting, and an AI oversight committee that evaluates tools annually. With 56% of students and educators saying their institutions are unprepared to manage AI, policy is the foundation, not an afterthought.

Step 2: Demand evidence and pilot with exit criteria. USC's Stephen Aguilar warns of an edtech "gold rush" selling districts products "without having any data to support their claims" — LA Unified's failed "Ed" chatbot being the cautionary tale. The NEA's guidance is blunt: adopt tools only when data supports their efficacy, and pilot unproven ones with a commitment to discontinue if benefits don't materialize.

Step 3: Choose pedagogically designed AI over general-purpose chatbots. The OECD found students using general-purpose chatbots produced better work, but the advantage "disappeared and sometimes reversed in exams when access was removed" — while tools built with intentional pedagogical purpose showed sustained gains. IES's systematic reviews of 80+ causal studies point to AI tutoring tools as the strongest-evidence category.

Step 4: Redesign assessment and teach AI literacy — don't rely on detection. Detection tools are fallible and disproportionately flag Black teenagers, per APA reporting. Proven alternatives include:

• Oral presentations in real time, as Australian senior science classes now use
• Assignments where students fact-check and edit AI-generated output
• Teaching responsible use, analogized to how schools adopted calculators

This matters beyond grades: 66% of education and business leaders won't hire someone without AI literacy skills. Verification-first thinking beats detection every time — the same logic SkillzBench applies to agent skills, benchmarking what actually works before deployment.

Step 5: Invest in teacher training and treat AI as a complement. Teachers report needing better guidance on responsible use, and over half of higher-ed professionals feel uncertain about effective AI application, UNESCO found. The payoff is real: teachers using AI weekly save an average of 5.9 hours — roughly six full weeks per school year. As IES puts it, AI works best as a complement to — never a replacement for — the essential work of educators.

The Verification Mindset: Why Testing AI Before You Trust It Applies Everywhere

The most important lesson from AI in schools isn't about education at all — it's about verification. When USC's Stephen Aguilar described the edtech market as a "gold rush" of companies selling districts products "without having any data to support their claims," he could have been describing almost any corner of the AI ecosystem.

The evidence from schools shows why a test-first mindset matters. The NEA's model policy insists AI tools "shall only be adopted once there is data supporting a tool's appropriateness and efficacy," with pilot programs that include a commitment to discontinue anything that doesn't deliver. Meanwhile, the LA Unified "Ed" chatbot collapsed within months after its vendor failed, leaving a district with a multi-million-dollar lesson in buying before proving.

The same logic extends well beyond classrooms — and it matters most where the risk is least visible. Consider AI agent skills, the instruction packages that extend coding agents like Claude Code, Codex, and Cursor. Published research found that roughly 26% of agent skills carry at least one vulnerability — a number that would fail any serious vetting process, yet most adopters have no practical way to check.

That's the gap SkillzBench is being built to close. The platform is creating free security scanning, benchmarking, and trust scores so anyone can test what actually works before deploying it. The approach mirrors what schools should already be doing:

• Verify before adopting — every skill version pinned to a commit SHA, then validated, license-checked, and security-scanned
• Benchmark in a sandbox with the same model, with and without the skill, so the measured Lift in performance is visible in percentage points
• Assign a 0–100 trust score built from security, quality, license, maintenance, and community signals — with Lift always shown separately, and money never touching scores
• Stay current as upstream changes are tracked, re-scanned, and re-benchmarked automatically

Everything on the platform is free — the directory, the benchmarks, the alerts, and the MCP connector that links any agent to the skills you choose. Whether you're a school district vetting an edtech vendor or a developer equipping an agent, the principle is identical: adopt only what's proven, and keep proving it after deployment.

SkillzBench is currently in pre-launch, and the waitlist is open. Join it to browse and benchmark verified skills via the free MCP connector as soon as the platform goes live — skills, proven.