The AI labs smell $6 trillion in the classroom, and they're offering teachers free tools to get a taste.
The Summary
- Anthropic and OpenAI are offering free and discounted AI solutions to educators, making their first serious push into the $6 trillion global education market
- AI cheating detection tools are proving unreliable, while fears about student cognitive atrophy grow
- Universities need to teach students "eval" skills, the ability to constantly test AI models against actual needs, not just how to prompt them
The Signal
The AI labs are executing the classic platform playbook. Free tier for students. Discounted enterprise for institutions. Get them hooked on your interface, your workflow, your model's particular strengths and blindspots. OpenAI and Anthropic aren't doing this out of educational altruism. They're doing it because whoever owns the next generation's AI habits owns the next generation's workflow, purchasing decisions, and mental models of what AI can do.
The education market is worth $6 trillion globally. That's larger than the entire crypto market cap at its peak. It's a market where the customer, the user, and the payer are often three different people (state, student, parent), which makes it complicated but also creates multiple entry points for disruption.
"Whoever owns the next generation's AI habits owns the next generation's workflow."
But here's where it gets messy. Cheating detection tools, the supposed guardrails against AI abuse, are failing. They flag false positives. They miss obvious AI-generated work. They create a cat-and-mouse game that teaches students to game systems rather than think clearly. Meanwhile, educators worry about cognitive atrophy, students outsourcing thinking to the point where they can't think without the tool.
The answer, according to one professor's argument, isn't to ban the tools or rely on faulty detection. It's to teach "eval" skills. Not prompt engineering. Not "how to use ChatGPT for your homework." Eval. The ability to test an AI model constantly against what you actually need. To know when it's bullshitting you. To understand its failure modes. To recognize when you're being handed synthetic slop instead of insight.
This is the skill gap nobody's talking about:
- Students learn to prompt, not evaluate
- They learn to generate, not discern
- They learn to ask questions, not test answers
Organizations will need people who can eval models, not just people who can use them. That's a different skill entirely. It requires understanding statistics, epistemology, the data the model was trained on, and the specific domain you're applying it to. It's closer to being a quality assurance engineer than a power user.
The Implication
If you're building edtech or working in education, the question isn't whether to integrate AI. The labs have already made that decision for you by giving it away. The question is what skills you're actually teaching. Prompt engineering is a temporary advantage. Eval is permanent. The students who learn to distrust and verify their AI outputs will run circles around the ones who just learned to ask nicely.
For the rest of us, watch what happens when an entire generation grows up with AI as their default co-pilot. They'll either be cognitively stronger because they had to learn discernment at scale, or weaker because they never built the muscle. The education decisions being made right now determine which future we get.