We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
AI in education is separating into two paths: tools embedded in teacher-led learning sequences and general assistants left to learners to use on demand. The clearest contrast this week is between Google’s reported Guided Learning trial, where educators remained central, and Tennessee evidence that optional Khanmigo use was thin.
Embedded tutoring is showing more promise than optional access
Google DeepMind and Fab AI report an eight-week randomized trial involving 1,763 junior-secondary students across 12 schools in Sierra Leone’s Port Loko District. Guided Learning students gained 0.258 standard deviations in mathematics against a control group—translated by the report into roughly 1.2–1.7 years of typical learning progress achieved during the eight-week trial.
The interaction data fit the product’s intended pedagogy: across more than 113,000 exchanges, the company says students were pursuing conceptual understanding in 91.4% of conversations; Gemini asked scaffolding questions in 76% of messages and gave direct solutions in 2%. Those are process indicators, not proof of durable learning. The result also needs qualification: students who entered with stronger mathematics skills benefited most, and Google says further preregistered trials are underway in other countries. The post discloses support from Google.org and the Gates Foundation.
The Tennessee evidence supplies the counterweight. In a randomized experiment across 18 middle schools, students using Khan Academy made faster mathematics gains by year two, but they used Khanmigo on only about one-third of Khan Academy days; many interactions were off-topic or attempts to obtain answers. In a separate experiment, an AI chatbot appeared without being requested and required a correct answer before students could progress. Those students moved more slowly, answered more accurately, and were slightly more likely to solve similar problems a week later without AI—a preliminary result, not a general license for chatbot use.
The implication for product design is narrow but important: access is not the intervention. Prompt timing, required effort, teacher integration, and an unaided transfer check are part of the intervention. Google says its Classroom integration, planned for this fall where administrators enable Gemini, will let teachers assign course-grounded Guided Learning conversations and review class- and student-level insights, misconceptions, and secure transcripts.
Sal Khan’s own postmortem makes the adoption problem plain: only about 10–15% of students were willing to experiment deeply with Khanmigo, while many others needed more proactive support. The planned response is to require students to explain their reasoning, not merely select an answer, and to integrate the tutor more deeply into lessons.
Distribution is scaling through different governance models
Google is putting a study environment directly in front of eligible students. Its Gemini student hub offers study notebooks, flashcards, and practice quizzes; notebooks can use uploaded class materials to create a study plan, diagnose knowledge gaps, generate short lessons and quizzes, and show progress on a dashboard. The offer is free for one year through Dec. 31, 2026. These are useful workflow capabilities, but the posts describe a product rollout rather than a learning outcome.
OpenAI is taking a more district-mediated route. It says ChatGPT for Teachers is expanding to 55 additional school systems across 20 states, bringing the program to more than 100 K–12 organizations in 30 states and more than 300,000 educators and staff. Access remains limited to verified U.S. K–12 educators and is free through June 2028. OpenAI also describes a 16-state privacy framework, managed workspaces with role-based controls, and hands-on implementation support.
The distinction matters. A direct-to-student hub maximizes reach and convenience; a managed teacher workspace adds institutional control and training. Neither substitutes for evidence that students retain and transfer what the tool helped them produce—the problem exposed by the Khanmigo study.
Procurement is becoming an evidence-and-transition problem
The U.S. Department of Education’s original guidance rejects a simple technology-versus-no-technology debate. It asks states and districts to distinguish recreational from instructional use and to make technology purposeful, instructionally sound, and tied to measurable outcomes. Its operational advice is more demanding: evaluate demonstrated learning outcomes rather than screen time, use evidence—including randomized trials—in procurement and renewal, review implementation continuously, and support professional learning.
New York City shows what happens when governance moves faster than operational planning. After backlash to its initial AI policy, the Education Department froze software purchases while it developed new guidance. With the school year approaching, principals reported being unable to renew even non-AI tools such as electronic gradebooks, family-communication platforms, digital hall passes, and reading and math software; one former department technology expert warned that schools would have to scramble in ways that undermine quality pedagogy.
The city is also canceling the T-Mobile internet component of a four-year student-connectivity contract. The service cost roughly $3.5 million per month and was projected at about $200 million over four years; officials say existing resources can cover students who need access more cheaply. Schools were told to identify those students and move to a Verizon program costing $10 per student per month. Usage was uneven: one Manhattan principal said about one in five students used the devices at home, while a Brooklyn school had dozens unused and had reduced Chromebook use in class because technology was getting in the way of learning.
The lesson for buyers is not to stop buying technology. It is to pair outcomes-based scrutiny with a transition plan: a district needs a way to test, renew, replace, or remove a tool without disabling the basic systems teachers and students rely on.
Assessment is moving from the submitted artifact to the learner’s thinking
A New York City English and special-education teacher describes a widening gap between polished take-home writing and students’ handwritten, in-class ability to organize ideas, support claims, and sustain analysis. The concern is not only cheating; it is that the assignment’s cognitive practice has been outsourced while the finished product still looks successful.
The teacher’s redesign is sequential rather than simply anti-AI. Students annotate independently, discuss ideas, gather evidence, and form an original claim before opening a computer. They then use AI to generate counterarguments, identify gaps, suggest evidence, or offer alternative interpretations, comparing those outputs with their own and making the final decision themselves. Handwritten planning, printed-text annotation, technology-free discussion, and occasional device-free writing are used to rebuild cognitive endurance. The teacher reports deeper questions, stronger defenses of reasoning, and more thoughtful revision.
A parallel response is emerging around academic-integrity enforcement. Christian Miller says even a detector reporting 99.99% reliability would not, by itself, meet the evidentiary bar for punishment or preserve classroom trust. He favors a low-confrontation conversation in which students explain their process, followed where appropriate by a reduced grade, revision, and a second chance. He also describes oral defense of code as a practical check, while acknowledging the tension with students who need technology for dysgraphia, dyslexia, or written expression.
Higher education and workforce learning are shifting toward judgment
Andrew Ng argues that university curriculum processes move too slowly for AI-driven changes in work, leaving some programs preparing students for outdated roles. He advises students to supplement degrees with current online learning, while also warning that AI can raise homework scores and reduce long-term retention when it performs the work. His new Learn Vector organization is aimed at more individualized, one-to-one learning experiences.
The skill response is not “learn less because the agent can code.” Ng’s AI Engineering Skills Map says developers still need fundamentals to steer agents and recognize tradeoffs in latency, reliability, security, maintainability, and cost. A higher-education workforce analysis makes the parallel institutional case: AI may make learning more adaptive, but programs must structure the experience, evaluate capability, work directly with employers, and teach judgment, communication, ethical reasoning, and general AI literacy alongside technical expertise.
This is also a continuing-education problem. More than half of postsecondary students reportedly enroll in at least some distance education, and one current model calls for disciplinary units to retain academic ownership while shared infrastructure supports online, adult, workforce, and lifelong learners through flexible scheduling, stackable credentials, employer partnerships, and advising.
What This Means
- For K–12 leaders: Treat an AI tutor as a teacher-led intervention, not a software license. Pilot it with defined learning goals, required student effort, unaided transfer checks, and subgroup analysis; the current evidence shows both thin voluntary engagement and a meaningful equity qualification.
- For assessment leaders: Design assignments so the student’s reasoning exists before AI enters the workflow, then use AI for critique or extension rather than invisible completion. Preserve documented accessibility accommodations and use process evidence, explanation, or oral defense instead of treating handwriting or a detector as a universal solution.
- For edtech buyers and vendors: Put the Department’s questions into contracts: the problem being solved, intended users, evidence of impact, implementation support, data protections, review points, and a credible exit path. New York City’s freeze shows that a policy pause without a transition plan can disrupt non-AI services as well as AI tools.
- For higher-ed and L&D teams: Pair AI fluency with domain fundamentals, judgment, and communication; build shorter, stackable learning that can be refreshed as work changes, while keeping employers involved in defining competencies.
Watch This Space
- Replication and distributional effects: Google says additional preregistered trials are underway, but its first report says stronger mathematics students benefited most. The next evidence should show whether guided tutoring travels across contexts and closes—or widens—gaps.
- Outcome-based procurement in practice: Watch whether districts turn the Department’s evidence language into renewal and removal decisions, and whether New York City can publish clear AI guidance without repeating the disruption caused by its software freeze.
- Student AI at scale: Gemini’s study hub and district-managed ChatGPT workspaces are expanding different access models. The meaningful measure will be active engagement and unaided performance, not the number of accounts provisioned.
Direct answer: The supplied August 26, 2026 update reports 55 additional school systems across 20 states, adding access for more than 100,000 educators and staff. With the new cohort, OpenAI says it is working with more than 100 K–12 organizations across 30 states and providing free access and training to over 300,000 educators and staff; the cohort includes one in five of the 20 largest U.S. public school districts.
- Duration and eligibility: ChatGPT for Teachers remains free through June 2028, but is available only to administrators, faculty, and educators—specifically verified U.S. K–12 educators. The update does not state student eligibility or access.
- Privacy and controls: Data shared in a ChatGPT for Teachers workspace is not used to train OpenAI’s models by default. School and district leaders can place educators in a managed workspace with role-based controls designed to support schools’ FERPA requirements.
- Privacy-agreement limitation: The announced common privacy framework spans 16 states and is intended to give participating districts a path to evaluate and adopt the tool under existing state and local processes, rather than guaranteeing uniform nationwide coverage. OpenAI says signed General Offers cover Illinois, Iowa, Maine, Massachusetts, Missouri, Nebraska, New Hampshire, New Jersey, New York, Ohio, Rhode Island, Tennessee, Texas, Vermont, Virginia, and Washington; California has a separate agreement.
- Training and implementation: OpenAI describes a shared implementation model involving hands-on training, district-leader support, peer learning, and ongoing guidance. It also cites training through OpenAI Academy, district work, and a five-year National Academy for AI Instruction initiative with the American Federation of Teachers intended to equip 400,000 K–12 educators; more than 1,600 teachers, administrators, and district leaders were participating in AI Skills Jams.
The guidance recommends that districts treat education technology as a tool serving clearly defined instructional goals and measurable student outcomes—not as an end in itself. Districts should be intentional about the digital experiences students receive and ensure technology use is purposeful, instructionally sound, and connected to measurable outcomes.
Instructional purpose and screen-time distinctions: Districts should distinguish policies governing recreational technology from those governing instructional technology, and evaluate instructional technology by demonstrated learning outcomes and instructional value rather than screen time alone. The guidance says recreational and educational technology are not synonymous, while acknowledging that unnecessary screen exposure should still be minimized.
Outcome measurement and continuous review: Districts should establish processes to review implementation and outcomes so investments continue to meet instructional goals. Procurement and renewal should rely on demonstrated effectiveness, with shared performance measures and clearly defined outcomes preferred over adoption or usage metrics alone.
Evidence for decisions: Technology investments should be guided by evidence alongside cost, usability, accessibility, interoperability, implementation support, and educator experience. Districts and educators should test vendor claims and consider third-party evidence, randomized controlled trials, implementation reviews, educator and parent feedback, and local outcome data; the guidance specifically tells districts to look for products with RCTs and positive outcomes in procurement and renewal.
Data and privacy: The guidance identifies protection of student data as a foundational responsible-technology principle and points to procurement models that standardize modern student-data privacy protections. It also presents transparent product information and clearly designed vendor contracts as part of responsible evaluation, but the district recommendation list does not provide a more detailed standalone privacy protocol.
Implementation: Districts are described as responsible for creating conditions for successful implementation. They should provide or support professional learning that helps educators integrate technology effectively, and implementation support should be considered alongside product effectiveness; providers are encouraged to supply implementation resources and professional learning.
Accountability: The guidance emphasizes that states, districts, educators, families, researchers, and providers share responsibility for outcomes—not merely implementation—and concludes that decisions should keep student academic outcomes at the center.
Direct answer: The supplied account describes a pre-registered, eight-week randomized controlled trial of Guided Learning in Gemini involving 1,763 junior-secondary students across 12 schools in Port Loko District, Sierra Leone. It reports a +0.258 standard-deviation gain in math scores versus the control group.
- Intervention design: The trial was explicitly teacher-led: educators designed lessons, set objectives, and facilitated classroom discussions, with Gemini used as a supporting tool rather than a teacher replacement.
- Outcome and magnitude: The reported quantitative outcome is math-score performance. The report translates the +0.258-SD difference into roughly 1.2–1.7 years of typical learning progress achieved during the eight-week trial; this “years of progress” figure is the source’s interpretation, not a directly observed follow-up duration.
- Usage qualifier: Classrooms in which teachers incorporated Gemini into roughly half of lessons, targeting 12 hours during the trial, reportedly saw larger gains—roughly 1.8–2.5 years of progress—and 69% of students met or exceeded usage targets. Because the excerpt does not say that usage intensity was randomly assigned or adjusted for baseline differences, this should be presented as a reported higher-use subgroup result, not as a clean causal dose-response estimate.
- What the interaction data do—and do not—show: Analysis of more than 113,000 interactions found that students used the tool for conceptual understanding in 91.4% of conversations; Gemini posed scaffolding questions in 76% of messages and gave direct solutions in 2%. These are process or usage indicators, not the reported math-score endpoint.
- Distributional limitation: The account says that although most students benefited, students who entered with stronger math skills benefited most. This is an explicit achievement-gap qualification: the headline average should not be read as evidence that the intervention benefited lower-performing students equally.
- Scope and evidence gaps: The evidence covers eight weeks in 12 schools in one Sierra Leone district, and the source says additional pre-registered RCTs in other countries are still being conducted. The supplied excerpt does not specify the randomization unit or allocation, control condition, assessment instrument, confidence intervals or other uncertainty measures, attrition, baseline balance, or longer-term persistence of gains; those omissions should temper claims about causality, durability, and generalizability.
- Disclosure: The source reports support from Google.org and the Gates Foundation and collaboration with EducAid, Laterite, and Oxford MeasurEd. This should be retained as provenance in a careful brief without inferring a bias that the supplied material does not establish.
- Justin Reich reported that MIT’s Teaching Systems Lab interviewed 120 teachers and students across the U.S. soon after ChatGPT’s release. Affluent schools where students were expected to pass state tests often treated AI as their most urgent issue, while teachers in lower-income Gateway communities sometimes ranked it as the 19th issue behind homelessness, poverty, absenteeism, and other challenges—and lacked substitute coverage for AI professional development.
- The broadcast characterized a Brookings Center for Universal Education report, based on international focus groups and interviews with K–12 students, parents, educators, and technology experts, as concluding that generative AI’s current risks outweigh its benefits, including risks to cognitive and social development, relationships, recovery from setbacks, and mental health. Malden, Massachusetts, responded by disabling AI chatbots and creating a parent-teacher working group to advise on products that are safe, improve learning outcomes, and are equitable; the district uses an opt-out-by-default approach. Shalin Title also said AI features and chatbots can be enabled on school Chromebooks unless administrators disable them, allowing Google rather than local communities to determine whether students use AI.
- Reich said judging instructional technology by its effect on learning is sensible, but argued that this understates the scale of technology infrastructure: devices, software, management, and major investment. He said K–6 schools often fail to realize the benefits and that families and teachers report poor instruction and excessive non-learning screen time; he cited Cambridge’s screen-use audit, which identified practices such as using screens for indoor recess while giving the community data for deciding what to change.
- Alpha School was presented as a national private-school chain where each student receives an AI chatbot tutor and adults serve as “guides” rather than teachers; a planned Boston campus was described as serving 25 students at $55,000 per year, while Springfield’s Kylie Prep sought to use the model. Reich said Alpha’s roughly two-hour online test-prep model followed by projects and field trips may fit private-school choice, but appears to omit subjects such as history, social studies, and much biology, making it difficult to reconcile with public schools’ broader curriculum. MTA president Matt Bock alleged that the model collects screenshots, keyboard and mouse activity, webcam and microphone recordings, and biometric data from headband monitors, with an opt-out that may not prevent collection during student interactions; Title warned that public-school pilots are experimental products with unknown risks.
- Reich argued that schools should give students opportunities to learn how to use AI with caring educators, since graduating without such experience could leave them less prepared; because LLMs appeared instantly on students’ existing devices, keeping them out requires deliberate changes to policy, classroom operations, and curriculum. For his own course, he planned to discuss AI with students and jointly decide what uses make sense and work, building on earlier models such as Burlington High’s student tech team, which evolved from technical support into lesson co-planning with teachers.
- Alpha’s “Time Back Learning” model is scaling as a time-compressed school format: the company says students complete academics in two hours a day and spend most of the remaining day developing life skills, with online and in-person options at different price points; it reports schools in more than 40 cities. The founder projects about 40 schools opening in fall 2026 and nearly 50 by year-end.
- AI tutoring is integrated into Alpha’s Montessori variant: Montessorium combines Montessori with Alpha’s academic platform; in its mixed-age classrooms, students can learn through an AI tutor while teachers review students’ academic activity and provide targeted support. The founder characterizes the first-year results as “insane” and says students were “learning like crazy.”
- The platform is being adapted around learner interests and expanded across delivery models: NextGen pairs the two-hour academic program with gaming, virtual-city building, coding, game development, and drone construction, positioning students as creators rather than only game consumers. Texas Sports Academy combines morning academics with afternoon sports and uses Texas education savings accounts to offer access free or for a few hundred dollars per month to families earning under $65,000 or with special needs; Alpha expects a couple thousand students to participate. GT Anywhere provides a virtual option accessible throughout the country. World School plans to bring the same learning platform to schools in Kenya and Ecuador alongside Starlink connectivity and solar, water, and food infrastructure projects.
- BloomTech shifted from its coding-bootcamp model as junior-engineer hiring collapsed; after asking hiring partners what they needed, it moved into AI training. Allred says a research team interviewed hundreds of companies and found that 99% were not using AI as claimed, leading to a $5,000-per-engineer online course. A private-equity customer then sent 100 engineers at a time and reported that measured productivity doubled after the course.
- Gauntlet AI grew from that employer demand into a 10-week immersive program: three weeks remote and seven weeks in Austin, with weekly challenges used to filter participants; applicants are screened at or above the 98th percentile on raw IQ. Employers prepay hiring credits and meet participants during weeks five through seven, with Gauntlet requiring a minimum salary of $200,000. Students pay nothing, while the program covers flights, housing, food, and compute.
- The curriculum deliberately changed the learner workflow: first-cohort engineers were barred from writing code, and some left because they doubted AI could produce useful code; Allred contrasts that early setup—which required roughly a month of data infrastructure—with current tools that he says are free and instant. Gauntlet has also expanded from training new hires to reskilling existing teams, including one- or two-week mixed pods of engineers, product managers, and designers, alongside broader AI-first transformation programs for larger companies.
- AI is reviving earlier edtech hype cycles rather than representing a wholly new break. Audrey Watters argues that AI has a long history intertwined with education and that companies repeatedly present new technologies as revolutionary; she says vendors have sometimes effectively replaced “personalized learning” with “AI” while selling the same software.
- The pandemic operationalized online learning at scale but also exposed its limits. Watters says the pandemic finally enabled the long-promised shift to online classes and forced schools into online settings, despite earlier research suggesting online education was not always best for everyone. She describes a student from a longstanding one-to-one program as exhausted and unfulfilled by spending the entire school day online.
- Technology does not inherently improve pedagogy. Watters says teachers can create rich learning experiences with or without technology, while online environments remove face, gesture, and body-language cues that teachers, students, and peers use to understand one another.
- AI and edtech adoption is also a governance and consent issue. Watters says communities increasingly resist technologies they feel are imposed on schools and children without consultation or control; she also warns that vendor-led media outreach can preselect supportive principals, superintendents, teachers, and students. She recommends grounding decisions in direct conversations with local educators, staff, parents, and students rather than projections from “internet land.”
- Khanmigo’s first wave exposed an adoption and pedagogy gap: Sal Khan said only about 10–15% of students—and a similar share of employees—were willing to experiment deeply enough for the tool to become “a game changer”; for most learners, the assistant was a non-event because they needed more proactive support and often did not know how to formulate the help they needed. Khan Academy planned deeper integration in which students would explain their reasoning to Khanmigo rather than only select a multiple-choice answer, with broader deployment planned for the coming school year.
- Khan Academy’s AI-tutoring design prioritizes guardrails and thinking over answer delivery: Early testing identified cheating, inappropriate conversations, hallucinations, math errors, and safety risks for under-18 users; the team responded by adding oversight, working to reduce errors, and designing a Socratic tutor intended to build critical thinking rather than cognitive offloading. A seven- or eight-person tiger team rapidly prototyped the system, and by January 2023 believed it could deploy with warnings and “extra resource” framing even though errors had not been eliminated.
- AI is becoming a learning skill in its own right: Khan distinguishes learning about AI, learning other subjects with AI, and learning to use AI generally; he argues that effective use requires asking better questions, evaluating outputs, and prompting iteratively. He also reports that access to Khanmigo, ChatGPT, Gemini, and Claude has accelerated his own self-directed exploration of unfamiliar and “edge” topics.
- AI may improve assignment performance while weakening learning: Andrew Ng says studies and data indicate that college students score higher on homework when using AI, but show much worse long-term retention and performance because the AI does the work for them; he argues that common uses of AI models are therefore poor for learning even though they are effective for getting tasks done.
- Learn Vector is pursuing personalized AI learning: Ng says he is leading Learn Vector to build learning experiences that are substantially more one-to-one and customized than the largely one-to-many online-course model, with more to be shown by early the following year.
- Universities are adapting more slowly than AI-driven jobs: Ng says university faculty and curriculum approval processes can take years, leaving many programs preparing students for outdated roles; he advises current students and graduates to supplement coursework with cutting-edge AI skills through online platforms such as Coursera, DeepLearning.AI, and Udemy.
- AI is expanding the skill requirements of work: Ng reports early signs that AI is helping software developers, marketers, and recruiters take on broader, more end-to-end roles, but says learners need both AI capabilities and deeper disciplinary skills to make that transition.
- K-12 use requires protection against cognitive offloading: Ng worries that children’s reliance on AI could damage long-term learning retention and says adult-supervised use of digital tools is beneficial, while recognizing that many adults lack the time to supervise consistently.
- Enterprise AI adoption is moving from standalone generative-AI chatbots toward agents and autonomous workflows, but the pace of innovation is outstripping many companies’ ability to adapt. Mehta identifies role-specific AI fluency as a critical response, citing an AI center of excellence spanning HR, marketing, finance, engineering, and product, plus a manufacturer seeking to train factory workers to build low-code/no-code agents on iPads.
- Workforce learning is becoming continuously responsive to model improvements: Allred says his training program changes its curriculum every cohort and quarter because tasks that were difficult six months earlier can become free with newer frontier models, increasing the emphasis on deep business expertise alongside AI workflow skills. Mehta likewise says his entire Google Cloud organization was scheduled for intensive training and retraining as the technology evolved rapidly.
- AI is lowering the importance of narrow programming-language experience while increasing the value of domain expertise. Mehta recommends hiring domain experts who are AI-curious and trainable; a retailer’s AI-generated marketing content failed without brand and customer knowledge, illustrating that AI augmented rather than replaced its marketing team. High-stakes AI workflows still require human domain oversight, such as an advertising-budget agent whose recommendations could misallocate substantial spend.
- Use AI as a learning and teaching collaborator: Ethan Mollick recommends broad experimentation—asking questions, drafting, recording activity for performance feedback, summarizing notes, writing code, and analyzing data—to discover where AI helps and where it fails. He suggests giving AI relevant context, such as assigning it the role of a “lesson plan designer.”
- Keep human judgment central: AI should support people and take on tasks they do not want to perform, while humans remain involved in final decisions and focus on their uniquely human strengths rather than treating AI as a replacement.
- Build AI literacy and skepticism into use: Although AI can be prompted conversationally, it is not human; its persuasive responses can reflect what users want to hear and make them overlook bias or overestimate its capabilities. AI is improving rapidly, so educators and institutional leaders retain control over how—and toward what ends—it is deployed.
- Prompt scaffolding can substantially improve an LLM’s performance on multi-step problems: in the video’s juggler riddle, zero-shot prompting often produces the wrong answer (8 instead of 4), while adding “let’s think step by step” raises reported accuracy from about 17% to nearly 79%, and adding “to be sure we have the right answer” pushes it above 82%. The explanation is that generating intermediate reasoning tokens gives the model additional computation before producing the answer.
- The video presents the context window as an LLM’s working memory and describes external scratchpads—writing intermediate computations to a file and reading them back—as a way to overcome context limits, paralleling how learners use paper to offload complex intermediate steps.
- The discussion raises a consequential skills-education question: as natural-language prompting replaces some direct coding and system construction, learners may need to focus less exclusively on deterministic syntax and more on steering, evaluating, and working with statistical AI systems; it also warns that future generations could lose deep coding ability.
- LLM-assisted reading: Andrej Karpathy says he now routinely reads books and papers with an LLM by uploading or pasting chapters, asking for an initial summary, and posing questions while reading; in his experience, this “dramatically increases” retention and understanding and makes older, technical, or unfamiliar texts more approachable. The workflow remains clunky—mostly manual copy-paste—and he is not aware of a tool that lets readers simply highlight a passage and ask questions.
- Language learning and active practice: Karpathy built custom GPTs that extract Korean vocabulary in dictionary form for copying into Anki, provide whole-sentence and part-by-part translations with follow-up clarification, and OCR, translate, and explain subtitles from screenshots. He attributes better task accuracy to detailed instructions and few-shot examples. Claude can also turn source text into flashcards and generate a browser-based app where learners reveal answers and mark whether they were correct.
- Multimedia source-based learning: Google NotebookLM lets users upload text, web pages, or PDFs, chat with those sources, and generate roughly 30-minute custom podcasts; its interactive mode allows questions during playback, which Karpathy uses for niche topics while walking or driving. He cautions that even research-oriented LLM outputs can hallucinate, so learners should treat them as first drafts and inspect the cited original sources.
Khan Academy created a free, hands-on STEAM activity in which students investigate how a giant balloon stays suspended, test balloon models, explore balanced forces, and design their own parade balloon. The activity includes downloadable teacher and student guides.
- Alpha School uses a hands-on stock-market simulation in which students allocate earned “alphas” between riskier and safer investment bundles, then track weekly gains or losses and connect market movements to economic news.
- The guide suggests using ChatGPT to generate age-appropriate fictional news stories for the exercise—for example, a fifth-grade scenario about how a major technology event might affect markets—and describes the activity as roughly 15 minutes of preparation plus a 20-minute weekly lesson that can be adapted for home or classroom use.
- AI is making self-directed learning more immediate for practical, low-interest questions: AstroForge CEO Matt Gialich described photographing a yellowing plant and receiving actionable advice, bypassing the need to read a book or search for the answer. He said this lets him “shortcut” information he does not care about, while not replacing the effort invested in subjects he does care about.
- A study of nearly 2,000 parents found that 57% of those shown evidence that AI erodes learning preferred a world where no student used AI, yet they would still pay for home access because other children have it.
- Alpha’s practical “BrainLift” model delays AI use until students choose a focused question, assess their existing knowledge and weaknesses, document credible-source facts, explain them plainly, identify takeaways, and form a research-supported point of view. Once the BrainLift is complete, students upload it to an AI tool and use prompts to identify gaps without receiving answers, challenge their position, design a simple experiment, or build an evidence-constrained outline while preserving their own viewpoint; the intended behavior is learner-led thinking rather than cognitive offloading.
- Google Classroom’s Gemini Guided Learning: Google is integrating a mode that guides students through complex problems with step-by-step support, probing questions, and productive struggle instead of simply providing answers, with the goal of building metacognition and independent problem-solving. The Classroom integration is planned for fall, subject to administrator enablement; educators can assign curriculum-grounded conversations and review class- and student-level engagement insights, misconceptions, and secure chat transcripts.
- Reported classroom impact: Google and Fab AI conducted a pre-registered randomized controlled trial with 1,763 Grade 7–8 students in Sierra Leone; after eight weeks, students using Guided Learning reportedly moved from the 50th to the 59th percentile in mathematics, described as roughly 1.2–1.7 years of additional learning progress. Teachers retained discretion over how and how much to use the tool, and Google says the strongest impact depended on active classroom integration by educators.
- Broader deployments: OpenAI is expanding ChatGPT for Teachers to 55 additional U.S. school systems across 20 states, bringing the program to more than 300,000 educators and staff across 100+ K–12 organizations, with free access through June 2028, privacy agreements in 16 states, and hands-on training. Harvard Business School is also using AI-generated faculty video avatars in a $699 entrepreneurship bootcamp, an early higher-education test of scaling instruction through AI representations of professors.
Khan Academy announced that Khanmigo can generate interactive math and science diagrams that respond as students explore, while teachers can create, review, and assign targeted practice. The update is supported by Google.org.
Ethan Mollick flags early evidence that Google AI Overviews may be affecting Wikipedia in a way analogous to how coding agents affected StackExchange, citing an SSRN study.
