We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
This week, three widely followed voices described the same failure in how students use AI. Andrew Ng said grades on AI-assisted homework have risen while memory and long-term results have gotten worse, because the AI does the cognitive work for the student. He went as far as saying that large language models, as most people use them today, are not suited to learning at all . Dan Meyer, who spends time in K–12 classrooms every week, pointed to research in which students using frontier models completed more homework and scored higher on it, but did worse on tests . Justin Reich interviewed 120 teachers and students across the country and found one theme running through almost every conversation: students using AI to bypass learning on homework, in ways teachers find hard to detect .
The debate is no longer about whether students use AI. It is about what kind of use still produces learning, and who decides.
Learning outcomes: the evidence points to guided use, not open access
Reich's main argument is that nobody yet has an evidence-based playbook for teaching with generative AI. Many people promoting "best practices" are, in his words, "making stuff up." Until research catches up, he says schools need "local science": educators testing clear hypotheses, being open with students and families that they are experimenting, and checking whether each approach actually works .
He offered one anecdote that shows the problem. Computer-science professors in Buenos Aires end their course with an oral exam in which students must modify their own code and explain how it works. The professors told him performance "plummet[ed]" after ChatGPT arrived, with students turning in messy code they could not explain . This is a reported experience, not a controlled study. Reich also argues that what makes someone good at judging AI output is expertise in the subject, not general AI literacy. That would make ordinary subject teaching more important, not less .
Meyer's view from classrooms is that there is little student-facing AI use in the K–5 and middle-school rooms he visits. He told teachers they should not feel they are missing out: there is no "two sigma gain" out there that other schools are getting and they are not . He sees the clearer near-term value in back-office and data work. He also criticized many "pedagogical" AI products as "a thin dusting" of Socratic prompting on top of a general-purpose model, built without a real idea of the teacher's role . OpenAI's representative on the same panel confirmed that its K–12 product is for teachers only, with no student product. As an example of the time-savings pitch, she described a principal who produced a chronic-absenteeism report in minutes instead of four hours a week .
Sal Khan made a related point. Looking back at how Khan Academy started, he now says the real power of one-on-one tutoring was "the accountability and the pushing," which video or a chatbot cannot easily provide. He said this in discussing a study that found students were not engaging much with Khanmigo . Ng says his new organization, Learn Vector, is building more personalized one-on-one AI learning and expects to show results by early next year .
Common Sense Media's new course for educators adds a well-being angle. Its survey of more than 1,000 children aged 9–17 found that nearly 9 in 10 use or interact with AI and about one in four do so daily. Among daily users, 42% said going a month without AI would be hard . Loneliness was more common among frequent users (54% of daily users, compared with 33% of those who rarely use AI) . That is a correlation, not proof that AI causes loneliness. The course acknowledges AI can offer low-pressure support, but warns it may keep students from seeking help from people .
Products are shifting from answer engines to guided study
Microsoft Copilot Study and Learn. This agent is now available globally. Microsoft says it is designed to guide students step by step and not simply hand over answers. It includes flashcards, personalized quizzes, matching games, and fill-in-the-blank practice . It is off by default for K–12 students; all higher-ed students have access . There are limits for now. Attaching the agent to Teams assignments is "coming soon," and sending students' usage transcripts and insights back to teachers has no release date . Until that ships, teachers cannot see how students actually used it.
For evidence, Microsoft cited a Fresno study of its Reading Progress tool involving 1,400 students. It reported gains roughly equal to one grade level for English language learners. The study is Microsoft's own and was run two years ago .
Learning Commons. This nonprofit, backed by the Chan Zuckerberg Initiative, launched an open platform with:
- academic standards from all 50 states
- learning progressions and curricula
- evaluation tools that check AI output
It also announced integrations with Canva Education, MagicSchool, TalkingPoints, and three others . It reports more than 70 partners, including OpenAI and Anthropic .
One district shows both the promise and the cost. Lisle District 202 in Illinois built a formative math-assessment app on the platform, and its knowledge graph flags prerequisite standards students may need to revisit. But the district needed to hire computer-science interns and bring in an outside firm to get the tool ready to launch. Most districts have neither . Critics also warn that whoever controls this shared infrastructure can shape what counts as authoritative. They are asking for transparency about which curricula and research get included .
Outcomes-based contracts. At an Edtech Insiders panel, speakers described contracts in which 40% to half of a vendor's payment depends on agreed student outcomes, with required data sharing. They noted that most procurement offices are not set up to buy this way . Teaching Lab Studio proposed a simple two-question test for any tool: does it clearly improve learning over the status quo, and will teachers actually use it ? Its teacher tool reports 40,000 users. As screen-time concerns grow, teachers are increasingly uploading paper student work to get AI feedback, which uses AI without adding student screen time .
Duolingo. Duolingo is one of the clearer cases where AI changes behavior. Luis von Ahn said the company could not get learners to practice conversation when that meant talking to another person. Most were too shy to speak a language they were not confident in. With AI video calls, "you don't feel like it's judging you" . The feature is behind a paywall because, he said, running the AI is expensive .
Other launches aimed at teacher productivity included MagicSchool's Presentation Generator, which builds slides from a learning goal and exports to Google Slides or PowerPoint . OpenAI released GPT-6 Sol and Luna to Edu users, among others .
Policy: the backlash reaches infrastructure, and states set their own rules
The biggest policy development is at the federal level. The FCC is reviewing E-Rate, which spends roughly $3 billion a year and covers 20% to 90% of schools' internet costs. The review could end the program, or require schools to limit screen time or offer parental opt-outs to receive funding . No draft rules have been published .
The stakes for learning are concrete. In Alaska Gateway, where many schools cannot be reached by road, E-Rate supports virtual career-technical courses and online Athabaskan language resources . District leaders warn that cuts would hit less-affluent districts hardest, and that mandatory opt-outs could single out students with disabilities who rely on technology for accommodations . The review comes while the White House continues to promote AI in education . Research on whether internet access improves learning is mixed .
State rules differ. Education Secretary Linda McMahon praised Florida's new classroom AI rules. By her account, parents must opt in before AI is used in their child's education, and the rules set guardrails so students use AI as a tool instead of becoming dependent on it . South Carolina's ban on minimum grades took effect this year, with 10% of state funding at risk for districts that don't comply . Experts link the law to doubts about what grades mean, fueled partly by AI in schoolwork .
Reich expects most change to happen locally rather than through bans. Cambridge, Massachusetts, stopped letting students use Chromebooks for indoor recess after an audit. Arlington's sixth-grade school now keeps devices in carts instead of sending them home . He cautions that removing devices will not improve learning on its own any more than adding them did. Either way, schools have to build new routines with adults, families, and students .
Classrooms: administrators' AI use is turning into a trust problem
Teacher forums this week were full of complaints about administrators, not students, using AI badly:
- One teacher counted at least five AI-generated posters within twenty feet of their classroom door .
- Another said their principal took down student art to make room for AI-generated posters .
- One class accused administrators of hypocrisy, since students themselves are not allowed to use AI .
- A teacher said district-provided slides included AI maps of their state with two capitals and a city placed in the water .
- A principal ran event photos through AI, which changed staff members' appearances and blurred students' faces, without anyone's consent .
A middle-school teacher's district made a writing-grading tool mandatory after one PD day, without asking teachers. Feedback on how it was working was scheduled for a month later, after the money was spent . These are individual accounts, not survey data, but they fit a request students made in a separate 12-district listening project. Students there asked schools to disclose when AI is used to grade or give feedback, and to set clear rules for each assignment .
Some teachers are pulling back entirely. A first-year history and civics teacher dropped classroom technology "unless absolutely necessary" after problems with AI misuse and students copying answers from Google. The teacher reports more engagement and better scores on the latest test .
Assessment: agents can fake the writing process
Lance Eaton showed that the "grade the process" approach has limits. He had Claude Co-work retype an essay he had already written into a Google Doc over two hours. The version history then looked "precisely like two hours of real human drafting" . He also warns that agentic browsers can log into a learning management system and complete assignments , and that detection tools have not held up under scrutiny . Principal Chris Leman, as Reich tells it, reached a similar conclusion for project-based schools: students can use AI to produce the "process" too .
Eaton's advice is to decide first whether an agent is operational, analytical, or pedagogical, since each needs a different level of human judgment. Then build governance and a sandbox before agents touch real students or data .
Higher ed and workforce: skills are changing faster, and entry-level roles need more experience
A new state-policy paper by Michael Horn and co-authors reports that between 2021 and 2024, a third of the skills in the average job description changed. Turnover was twice as fast in jobs most exposed to AI . The paper also notes that AI is taking over the "grunt work" new hires used to learn from, so employers expect entry-level workers to arrive with experience. Work-based learning before graduation is its proposed response . Separately, a new study found the college wage advantage has dropped sharply. Researchers roughly estimate AI accounts for a third of the drop but say the attribution is uncertain, and the premium is still high .
Alpha School's model is an example of mastery learning at scale. Its AI platform won't let students advance until they show more than 90% mastery. It also uses spaced repetition, and it packs core academics into a two-hour block . The staff are "guides" who motivate students; they don't teach .
What This Means
- Grades and learning are separate measures. Meyer's homework-versus-test finding and Ng's point about retention both mean AI-assisted work can make students look better on paper than they are. Check learning with work done without AI (in class, oral, or on new problems), not only with work where AI may have helped.
- When evaluating AI study tools, look at the design. Guided modes like Copilot Study and Learn are a better starting point than open chatbots. But their key feature for teachers (seeing how students used the tool) is not available yet. Try them for low-stakes practice and judge them on unaided performance.
- Make vendors carry some risk. Outcomes-based contracts and Teaching Lab's two questions (better than the status quo? will teachers use it?) apply to any tool. Most procurement systems will need to change to use them.
- Administrators set the example. Poor-quality AI posters, altered photos of staff and students, and tools imposed without asking teachers undermine the classroom AI rules schools are trying to enforce. Tell students and staff when AI is used, and ask teachers before buying.
- Don't rely on the writing process as proof of authorship. Version histories can now be faked. Check understanding more directly, and plan for how oral and in-person assessments affect students with disabilities.
- Keep connectivity separate from the screen-time debate. Changes to E-Rate would hit rural and high-poverty districts first, whatever position a district takes on classroom devices.
Watch This Space
- FCC E-Rate rulemaking: whether draft rules add screen-time or opt-out requirements, or narrow eligibility to rural schools.
- Microsoft's teacher insights: whether Study and Learn starts sending usage transcripts back to teachers, and whether independent studies follow.
- Learning Commons governance: how it decides which curricula and research go into infrastructure that many products build on.
- Learn Vector: Ng promises early results on personalized AI learning by early next year.
- Florida's opt-in model: whether other states copy parent opt-in for AI, and how schools handle students whose parents don't opt in.
- Assessment redesign: more oral and in-person verification, and whether institutions address the access problems Eaton raises.
