We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
AI is being added to classroom tools faster than schools can check whether it works. EdReports reviewed 10 K–12 curriculum and edtech providers, including four of the country's top publishers. Only one offered third-party evidence that its embedded AI features improved educational outcomes . Several other developments this week involve the same gap between using AI and proving its value.
The evidence gap is now a buying problem
EdReports' brief points to a practical issue: AI can be added to products without the vendor telling anyone. One elementary principal said his school found out that some tools it already used had AI built in that it "simply hadn't been told about" . Digital Promise's CEO described the worst case: "quality assurance at adoption no longer guarantees quality at instruction" . Existing review frameworks mostly check privacy and safety. They don't have a consistent way to judge whether AI strengthens learning or keeps a curriculum coherent .
For buyers:
- Ask vendors whether AI updates are opt-in. One expert argued that every new AI feature is an instructional decision .
- Ask how the product is supposed to be implemented, not only whether it works. The Center for Outcomes Based Contracting warns that the study did not account for what students actually do with the materials. Without good implementation, the field stays stuck in a 20-year cycle of "shiny new object[s]" that never build evidence .
A survey shows the guidance gap
Microsoft surveyed 3,345 people in the U.S., U.K., Australia, Brazil, Japan, and Saudi Arabia. 92% of students, 88% of educators, and 92% of education leaders had used AI for school at least once or twice. Daily use was 27%, 33%, and 53% respectively .
The groups see the situation very differently:
- Guidance: four-fifths of leaders said their institution's AI guidance is clear. Half of students and teachers said it is neutral or missing .
- Training: 77% of students and 53% of educators said they've had no AI training. Seven in ten leaders believe at least half their users have been trained .
- Worries: 42% of educators ranked more cheating among their top three concerns, and 41% of students ranked being accused of cheating among theirs. Leaders' top concern was data privacy, at 23% .
Put together, the people who write AI policy and the people who have to follow it are worried about different things.
Khan: a more proactive tutor, still unproven
Sal Khan described a Khanmigo that steps in on its own instead of waiting for students to ask:
- After a wrong answer, it offers credit if the student can explain their reasoning.
- After a right answer, a strong explanation can replace the four more correct answers normally needed to show proficiency.
- Teachers can assign a Socratic AI tutoring session to students struggling with a specific topic and then get a report on it .
Khan was careful about the evidence. He said the team thinks it sees "a signal" but doesn't want to overstate it, given confounding variables. This coming year will test whether the proactive supports add anything beyond traditional Khan Academy .
On safeguards, Khan said the AI model providers agreed not to use teacher or student data, and that grounding the tutor in Khan Academy's human-written content pushed hallucinations close to zero. By his estimate, about one in 40–50 full tutoring sessions contains some error, and the AI sometimes corrects it . He said explicitly that the tutor is not a replacement for a human tutor .
His evidence for pre-AI Khan Academy is about how it's implemented. In a University of Toronto study in Indian government schools, students got 30 minutes a day of Khan Academy with a paraprofessional keeping them engaged. They saw an effect of almost half a standard deviation over seven months . A Tennessee trial with students behind grade level found a 0.14 effect over a year . He also warned that bans like New York City's mostly cut off students whose only access to AI is at school, while affluent peers keep using it at home. He argued that proctoring assessments would address cheating more precisely .
Justin Reich added an equity point of his own. In MIT interviews with 120 teachers and students, AI was the most urgent issue in affluent schools. In low-income communities, teachers called it "the 19th issue on our list," and some couldn't attend AI training because no substitutes were available .
How districts are choosing tools
Denver Public Schools approved MagicSchool as its only student-facing AI tool. It got there after a year-long pilot, a formal RFP, and input from teachers, students, and families. This is according to MagicSchool's own team, who also cited an independent Common Sense Privacy review .
Third Space Learning offers districts a fully funded AI math tutoring program backed by the Gates Foundation. It comes with a Stanford and Cornell study of whether AI tutoring can match high-dosage human tutoring. A sales advisor reported more than 30 pilot "yeses" from 30+ district meetings. He credited the funding and research backing, not the product, which he said made saying yes "the safe choice" . That is interest in pilots, not learning results.
New tools, and how teachers are using them
- Microsoft Teams for Education can now create flashcards and fill-in-the-blank activities from a teacher's attached notes, which teachers can edit or "enhance with AI" . Students can also generate quizzes and flashcards from their own assignment materials . This saves time, but practice built from one document only covers what that document contains.
- Google Gemini will replace Gems with Skills. For Workspace education customers that happens in June 2027, and Gems will be migrated automatically .
- In Japan, educators in a Google showcase describe Gemini's Guided Learning as giving hints toward an answer instead of the answer itself. One reports a workload of about 60% of what it was before Gemini. A lecturer has students submit their full AI dialogue logs, saying that grading only the final output has to end .
- In Seoul, a teacher of adult professionals turns rough class notes into feedback documents and next week's lessons using an AI agent. Feedback time per student dropped from about 30 minutes to 2–5 . The teacher's caveats: the output is only as good as the notes, everything needs review, and writing down "my rules for good feedback" took the most effort .
Student opinion is split
Teacher reports from forums don't agree. One teacher describes half of students using AI daily and the other half despising it. A "good chunk" of students walked out of an administrator's fully AI-generated presentation . A computer science teacher's survey of 28 students gave a more mixed result: 5 anti-AI, 4 worried, 3 pro-AI, 5 hopeful, and 11 who didn't know or didn't care. As that teacher put it, the loudest voices get over-represented . A teacher piloting an AI literacy curriculum says skeptical students still accept lessons on how AI learns, bias, and checking AI Overviews . These are anecdotes, not survey data.
Higher ed: AI is now part of the college search
EAB's 2026 survey of more than 5,000 high schoolers found 46% using AI in their college search, up from 26% in spring 2025. 18% had dropped a college from consideration because of what AI told them . AI is being added to existing research habits, not replacing them: 84% of prospective adult learners still use search engines . One analysis warns that AI may favor prominent institutions, rely on outdated information, or mix up similar programs. That could leave a regional university that is a better fit less visible .
Other developments
- Alpha School: Michael Horn and co-authors argue the $65,000 school itself is not disruptive. They see its Timeback software as the more plausible path, potentially powering a wave of low-cost private schools .
- Pearson is acquiring Workera, whose AI tools measure employees' skills for businesses .
Watch this space
- Whether districts start requiring opt-in AI updates and evidence in contracts, following EdReports' findings.
- Khan Academy's results for its proactive tutor this school year.
- The Stanford/Cornell results on AI versus high-dosage human tutoring.
- Whether schools close the training gap for teachers and students that Microsoft's survey found.
- Khan Academy is making Khanmigo proactive: it prompts students to explain their reasoning, can accept a strong explanation instead of requiring four more correct answers, and lets teachers assign AI tutoring to students struggling with identified topics and review reports. Khan said the added efficacy of these newer AI supports was not yet established; further evaluation was planned.
- Khan described privacy and accuracy safeguards: model providers agreed not to use teacher and student education data, and Khan Academy tries to anonymize it; grounding responses in Khan Academy content reduced hallucinations to almost zero in its experience. He estimated an error in roughly one of every 40–50 full tutoring sessions, sometimes corrected by the AI, and said AI should not replace a human tutor.
- Khan cited implementation-dependent learning gains from Khan Academy: an independent University of Toronto study in Indian government schools paired 30 minutes of daily use with a paraprofessional supporting student engagement and saw nearly a 0.5 standard-deviation effect over seven months; researchers described the gains as comparable to two or three years of typical progress. A separate Tennessee randomized trial of students below grade level reported a 0.14 effect size over a year, but students were removed from the intervention when it was working, which Khan said may have understated the effect.
- Khan warned that school AI bans may disproportionately disadvantage students whose main access is through school, while affluent students retain access at home. He argued for proctoring assessments and designing assignments that explicitly permit AI use while requiring students to explain their use.
- Khan said the Conted Institute planned to start with certificate and master’s programs rather than its announced bachelor’s, with an interest list of more than 30,000 people, mostly already holding bachelor’s degrees. Its proposed online program uses team projects on real-world tasks and portfolios of created work, with peer and AI feedback and emphasis on durable skills such as collaboration and communication.
- After ChatGPT’s release, MIT’s Teaching Systems Lab interviewed 120 teachers and students across the country. Reich said AI was often treated as an urgent problem in affluent schools where students were expected to pass state tests, but teachers in lower-income Gateway communities facing homelessness, poverty, and chronic absenteeism often ranked it far lower; some said a lack of substitute teachers prevented them from getting AI professional development.
- Reich argued that schools should review technology as infrastructure, not just as individual apps: in K–6, its benefits often go unrealized, while parents reported poor instructional apps and excessive in-school screen use. Cambridge’s screen-time audit identified practices to reconsider, including using screens for indoor recess, and gave the community data for discussion.
- Reich said generative AI’s rapid arrival on students’ devices means excluding it requires changes to school policy, classroom routines, and curriculum; he favored setting limits while giving students some teacher-guided experience with the tools. He also questioned Alpha School’s two-hour test-prep model, saying it appeared to leave out substantial curriculum, including history and social studies and much biology, which conflicts with public schools’ broader curricular expectations.
- Infinity Academy, an Ontario project-based microschool, uses AI/digital apps for personalized math and literacy alongside explicit instruction. Its academic time is split into one-hour morning and post-lunch blocks, with project/workshop time and outdoor free play as well.
- The apps assess students’ math and literacy against Ontario curriculum standards and show parents their standing. The educators say the apps should complement explicit instruction to avoid learning gaps; they report seeing growth and performance benefits but give no quantified results.
- To improve engagement with online lessons, the school added external rewards for effort and progress rather than grades, and reports a large engagement increase.
At a White House AI-accord event, Trump said data-center companies would help local schools and teachers, including with teacher pay, and described partnerships with communities as part of the expansion. A speaker cited Louisiana’s Richland Parish, tentatively identifying Meta, and said a company was paying every teacher over $50,000 apiece.
- A Microsoft survey of 3,345 people across the U.S., U.K., Australia, Brazil, Japan, and Saudi Arabia found AI use for school-related purposes was widespread: 92% of students, 88% of educators, and 92% of education leaders had used it at least once or twice; daily use was 27%, 33%, and 53%, respectively. Optimism varied by country: U.S. respondents were least optimistic and Saudi respondents most optimistic, while U.S. student optimism fell 21 percentage points in a year.
- Institutional readiness lags adoption: four-fifths of leaders said their institution’s AI guidance was clear, while half of students and teachers said guidance was neutral or absent; 77% of students and 53% of educators reported no AI training, despite seven in ten leaders thinking at least half of users had been trained. Concerns differ across groups: 42% of educators cited increased cheating or plagiarism, 41% of students feared being accused, and leaders’ top concern was student and staff data privacy and security (23%).
- Most respondents considered effective and responsible AI use important for students’ futures, but fewer than half of educators called AI literacy essential to basic education. The proposed scope goes beyond prompt-writing to understanding AI’s limits, evaluating its outputs, ethical judgment, and knowing when not to use it; Microsoft’s Minecraft Education AI Ready Skills is an example of game-based foundational instruction for younger learners.
- The newsletter also reports practical offerings: Articulate introduced AI-powered workplace-training tools, including an AI-native platform for frontline teams and integrations with ChatGPT, Claude, and Microsoft Copilot; Flint’s free, personalized teacher AI-literacy course offers a certificate and had more than 8,500 teachers enrolled; Renaissance launched a beta evidence-resource hub designed to help learning-engineering AI agents find and cite trusted resources.
Altman described stronger education and helping teachers teach students as potential benefits of AI, and said AI should expand people's ability to learn while keeping people—not automation—at the center. He cited OpenAI models' math progress—from being “okay” at grade-school math three summers earlier, to “pretty good” at high-school math competitions two summers earlier, to gold-level performance at a prestigious international math competition the previous summer—as evidence of growing model capability.
A founder-school pitch described a risk that AI-assisted code pushes could accidentally expose student data; the presenter claimed existing scanners did not detect personally identifiable information, especially student data, and proposed internet scanning for $20 per coder per month. The mentor recommended pairing detection with remediation—fixing exposed data rather than only flagging it—as a more complete solution for edtech vendors.
At the summit, Trump said data-center companies would partner with host communities to support schools and teacher pay. As an example, he said a company he believed was Meta was paying every teacher in Richland Parish more than $50,000; the transcript leaves the company attribution uncertain and does not independently verify the claim.
On what schools should teach children, Ian Rogers said he had moved from uncertainty to seeing AI-agent operation as a future work role: people will operate agents and help move work forward, a shift he likened to knowledge workers adopting computers.
Ethan Mollick says the durable skill behind “using AI better” is unclear as AI becomes easier to use, organizational systems increasingly bottleneck AI gains, and job effects remain uncertain but may reach entire job categories. He adds that individual AI use can still create substantial workplace advantage, but recommends building systems for human augmentation rather than automation, since individual performance is not the only factor.
Khan Academy promotes Khanmigo as support for parents and grandparents helping with homework: it can help them revisit concepts and support learners with more confidence.
- Arkansas launched Arkansas LAUNCH in 2025, a free digital skills marketplace for jobseekers and employers that featured nearly 6,000 job openings from more than 600 employers at launch.
- Maryland’s Longitudinal Data System links education and workforce records and has expanded to include data from other state agencies and external credential providers such as Microsoft; its research has informed the design and continuing implementation of Maryland’s education reforms. Linked administrative data can reveal patterns and associations, but establishing that a program caused an outcome requires an appropriate evaluation design; the paper emphasizes secure, governed sharing to protect privacy and sustain trust.
- Third Space Learning is offering districts a fully funded AI math tutoring program backed by the Gates Foundation, alongside a Stanford and Cornell study testing whether AI tutoring can match high-dosage human tutoring. K–12 sales advisor Jacob Kantor says he secured more than 30 pilot “yeses” across 30+ district meetings by presenting the program; the post reports pilot interest, not learning outcomes.
- A Flint episode preview identifies teacher AI literacy as a prerequisite to student use, data privacy as a reason districts hesitate, and a lack of formal AI guidance for many teachers. It says Flint’s personalized course offers thousands of possible learning paths to address “AI literacy fatigue.”
For AI/EdTech decisions, assess evidence by whether methods are appropriate, transparent and rigorous—not by whether they come from a lab, classroom, company or university—and combine controlled studies of effects under specified conditions with teacher- and classroom-based research on workability, meaning and sustainability.
EdTech’s emphasis on standardized literacy and numeracy outcomes reflects wider funding, procurement and evaluation incentives, not technology alone; product metrics can therefore encode system priorities. Evaluate broader educational aims as well, and apply the same rigor to digital and analogue approaches.
Ethan Mollick observes that AI has made PowerPoint presentations more engaging through better layouts, humorous visuals, and diagrams, while some users also delegate the thinking; he notes that default-template use predated AI.
- Alpha’s $65,000 school is positioned at the high end, and its highly paid learning guides and low student-to-adult ratios may make the model difficult to scale affordably; the authors therefore say the school itself likely does not fit the disruptive-innovation pattern.
- The authors see a more plausible technology-led route in Timeback, which oversees Alpha’s morning two-hour learning cycle and could enable a wave of low-cost private schools. Alpha’s planned low-cost, fully virtual software version may signal such a strategy, but the post presents this as a possibility rather than an established outcome.
Alpha School’s K–12 model uses an AI-powered dashboard for two hours of personalized academics: students follow the same curriculum at their own pace, while the system identifies when they are struggling or progressing and supplies level-appropriate material; staff act as motivational guides rather than lecturers. Laptops then close for lunch and afternoon hands-on group projects and life-skills workshops. The school says personalized plans let students master academics in two hours rather than six, with the intended benefit of freeing time for creativity, communication, collaboration, and critical thinking.
The author cites a recent Harvard study reporting that students learning physics with an AI tutor learned more than twice as much, in less time, as students in an active-learning classroom; this is presented as evidence for the potential of personalized AI tutoring, not as an evaluation of Alpha’s own model.
- The author links AI-driven expansion in higher education to two shifts: AI is “seniorizing” entry-level roles, increasing the value of gaining experience in college, while social capital matters for job searches. The author says at least 50% of jobs are filled through networks rather than blind online applications and predicts that share will rise in the AI era.
- Universities’ AI infrastructure may not reach classrooms without faculty buy-in: the post describes a widening gap between faculty skepticism and administrators’ impatience over AI adoption.
Shake Up Learning shared a teacher-facing Canva tips article titled “10 Canva Secrets Every Teacher Should Know”; the post does not describe the tips or provide evidence of classroom impact.
- An ESL instructor developing AI-literacy curriculum prioritizes explaining how AI works, critically evaluating its outputs, and deciding when its use is appropriate over product-specific practice . Commenters say K–12 AI-literacy courses are already available from Learning.com, BrainPOP, and others, while a former product seller describes the market as “awful” .
- One commenter says they use a game to help children understand how AI training works ; another says a school offering needs a FERPA agreement and a closed environment .
- Views on introducing AI literacy to ages 7–12 differ: a parent and learning technologist opposes it before children develop basic literacy and cognition, claiming technology is proven to stunt cognitive development , while another commenter argues that technology becomes more intuitive and compares early AI-literacy teaching to teaching children about Zip disks .
