We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
New York City chooses restriction plus literacy
New York City has moved from delayed guidance to a concrete, age-tiered AI policy. Student-facing generative AI — including prompt-responsive chatbots and tutors — is prohibited in preschool through eighth-grade classrooms for at least the coming school year.
The policy is not a blanket rejection of AI. High schools may use it in career and technical education and in five approved generative-AI pilots, capped at 50,000 students, or roughly 17% of high-school enrollment. Every high-school student must complete two 45-minute AI-literacy modules each year covering AI basics, bias, and misinformation. Younger students face recommended device limits: 30 minutes a day in grades 3–5 and 45 minutes in middle school. Teachers may use AI for lesson planning, translation, and drafting communications, but not for grading, behavior monitoring, special-education plans, or student mental-health support. Exemptions cover specific accessibility and English-language-learning needs, as well as required computer-based reading and math programs.
The hard part is implementation. Officials say they are disabling AI features in 38 existing software programs but have not identified them; they also acknowledge limited data on AI use across the city’s 1,600 schools and have not spelled out an enforcement mechanism. New York is therefore making a substantive distinction — restrict student-facing answer engines in younger grades while teaching older students how to evaluate AI — without yet showing how that distinction will work inside the software schools already use.
The assessment target is learning, not AI detection
An August MIT report, discussed by Tony Frontier, says a final class product can no longer be assumed to measure learning. It recommends oral exams, semester portfolios, out-of-class work paired with in-class conversations, and AI that is “pro-learner” rather than a replacement for the learner. Its warning is cognitive surrender: turning to AI at the first sign of struggle can create the appearance of learning without the underlying growth.
That warning is already visible in classroom practice. Frontier says students in focus groups across different kinds of high schools typically reported using AI for 25–40% of schoolwork, while teachers offered mostly vague permissions or outright bans and few specific lessons about how or why to use it. His proposed shift is from AI detection to “learning detection”: oral assessments, explanations of sources and decisions, and evidence that the submitted work represents the student’s knowledge and skills. He also reports false accusations leading some students to run their work through “humanizers,” turning detection into another part of the arms race.
The American Psychological Association’s new report makes the same point at the platform level. Time on an app, enjoyment, completion, and higher scores can be mistaken for learning; AI can improve a finished product while reducing what the student learns. The APA calls for independent evaluations, attention to which students benefit or continue to struggle, and scrutiny of what a platform replaces — including teacher or peer interaction. Its practical test is simple: close the app, ask the student to explain, solve a new problem, or apply the skill elsewhere.
Oxford offers a concrete higher-education implementation. Kelly Webb-Davies says 50% of a previous cohort was referred for academic misconduct after a take-home essay assignment, prompting her to separate the evidence of thinking from the polished final product. In the first stage, students respond to an unseen prompt or dataset in a secure, observed setting, with language access made as broad as feasible. In the second, AI use is optional and unrestricted; the final work is audited against the student’s initial thinking, with a viva only when the development cannot be reconciled. She says the redesign measured argumentation better and created almost no academic-integrity policing burden, although it took more time. The first stage can also capture ideas in a student’s first language or a stigmatized English variety, using translation technology to preserve evidence of thinking.
The low-tech correction has an opportunity cost. A current New York report describes one English teacher moving most work by hand into class because she expected students to use AI on homework; she says that left less time for reading and that students now fatigue after a few pages. Another teacher reports moving from five novels and a Shakespeare play per year to about three books. These are teacher reports, not causal evidence, but they show how protecting authorship can consume time that would otherwise support reading endurance.
Tutors are being designed around preserved effort
Microsoft’s back-to-school education update illustrates the product response. Its Study and Learn Agent is presented as a learning-focused coach for K–12 and higher education that works from students’ own materials, creates activities, and supports 40 languages. In the demonstration, it begins with questions to establish what the learner knows, adapts its next question to gaps and mistakes, inserts “pause and think” prompts, and follows the interaction with quizzes or matching activities. For writing, it refuses to produce a requested thesis statement and instead coaches the student through the reasoning and outline.
The governance is part of the design. In K–12, administrators control whether Copilot Chat is enabled and Study and Learn is preselected; educators can assign the agent with the full context of an assignment so it guides the student without completing the work. That is a meaningful move away from a general answer box, but the webinar demonstrates a workflow rather than reporting an independent learning result. The open question is whether guided interaction produces better unaided performance after the tool is removed.
Agentic AI turns self-directed learning into a verification task
OpenAI’s GPT-6 Astra launch marks a different kind of change. The company describes Astra as a general-purpose computer-use model that can conduct research, draft summaries, analyze scientific data, generate plots, create websites, run quality checks, install and test software, and troubleshoot problems on screen. In an accompanying interview, Sam Altman says people are already using it to build software and games, run scientific simulations, and tackle projects that previously required several specialized tools. For self-directed and project-based learning, the important shift is from asking a system for an explanation to working with it on a multi-step artifact.
The limitation is that capability is arriving with operational risk, not as a validated pedagogy. OpenAI says Astra has critical-level cybersecurity capabilities, refuses more advanced exploit tasks at launch, and may pause or stop legitimate work when safety checks intervene; enterprise access is off by default at launch.
Andrew Ng’s current skills map describes the human capability this new model class demands. Effective agent use involves planning, specifying requirements and architecture, calibrating autonomy, testing outputs, and iterating when verification fails; it is less about writing every line of code and more about deciding what to build and whether the result is sound. He cautions that long autonomous runs are currently oversold relative to the value of skilled, iterative intervention. For advanced learners and workforce programs, AI literacy therefore needs to include project framing, domain fundamentals, evaluation, and responsibility for the final result — not only prompt technique.
Governance is becoming part of the learning system
An Edtech Insiders analysis describes AI councils and task forces as an emerging intermediary between vendors and classrooms. It reports that only 6% of teachers in the cited 2026 Stanford AI Index said their school’s AI policies were clear, and argues that effective councils need protected operating time, meaningful student authority, clear responsibilities, and a way to scale decisions across the school community. The analysis is an industry perspective rather than a systemwide evaluation, but it identifies the missing operating layer: policies need people who can translate them into assignment rules, procurement choices, accessibility decisions, and feedback from students.
What This Means
- For school leaders: Build an assignment-level AI map by age, subject, task, and accessibility need. Pair restrictions with a named owner, an approved-tool list, a review date, and a plan for software whose AI features cannot be cleanly separated. New York’s unnamed 38 programs and unclear enforcement show the cost of leaving those details unresolved.
- For assessment leaders: Require two kinds of evidence: an observed or otherwise traceable record of the learner’s thinking, and a later explanation, new problem, or transfer task. Permit AI where it supports critique or development, but do not treat a detector or a polished submission as proof of learning.
- For edtech buyers and product teams: Treat “tutor” as a learning hypothesis, not a feature label. Ask for independent evidence, subgroup results, and performance after the app is closed; also ask what teacher or peer interaction the product replaces.
- For higher education and L&D: Teach learners to frame projects, specify requirements, calibrate agent autonomy, verify outputs, and retain responsibility for judgment. That is the durable skill set behind both Oxford’s assessment redesign and Ng’s agent workflow.
Watch This Space
- NYC implementation: Watch whether the city names the 38 affected programs, publishes an enforceable review process, and reports outcomes from the five high-school pilots and the annual AI-literacy modules.
- Guided tutoring evidence: Microsoft’s answer-resistant Study and Learn design is directionally different from a general chatbot. The next meaningful measure is unaided transfer and sustained engagement, not the number of guided sessions or activities generated.
- Agent safety and access: Astra’s staged rollout, monitorability concerns, and interruptions to legitimate work will matter as learners begin using agents for longer-running projects.
- School-level governance: AI councils will be worth tracking only if they receive time, student authority, and a route from discussion to consistent classroom and procurement decisions.
- Reading endurance: Follow whether the shift from take-home digital work to supervised, handwritten work restores learning without displacing the full-length reading practice that teachers say is already shrinking.
Direct answer: The announcement presents GPT‑6 Astra as a general-purpose, tool-using agent for browsing, computer use, coding, scientific work, and multistep professional workflows. Those capabilities are directly relevant to self-directed or project-based learning, but use is bounded by staged availability, enterprise controls, cyber refusals, monitoring, and safety checks that can interrupt work.
Capabilities relevant to self-directed/project-based learning
- Research and build loop: Astra can conduct online research, draft summaries in email or document editors, analyze scientific data, generate plots, create websites, run frontend QA, install and test software, and troubleshoot on-screen problems.
- Project deliverables: It is trained for multistep workflows and can produce polished documents, spreadsheets, presentations, and analyses that follow templates and match a requested writing or visual style. In ChatGPT’s Sites feature, it can create, host, and share websites, web apps, and games directly from a prompt.
- Bounded autonomy: When instructions are ambiguous, Astra fills routine gaps, asks focused questions when the answer could affect the outcome, continues independent work asynchronously in Codex, makes sensible assumptions when appropriate, and waits for input on consequential decisions. It is also described as incorporating new requirements without losing the broader task.
- Long-running projects: An experimental Codex feature lets Astra preserve notes across context windows and search earlier windows for requirements, test results, and tool outputs; the feature can be enabled in
config.tomland is expected to become the default for Astra in the following weeks. - Scientific project work: The announcement says Astra can use specialized scientific software to inspect data, explore results, assess evidence, and help decide what to investigate next.
- Reported performance: The page reports 64.6% on Terminal-Bench Science 0.1 versus 52.6% for Claude Fable 5.1, at approximately 31% lower estimated API cost, and reports higher scores than GPT‑5.6 Sol on Agents’ Last Exam and OSWorld 2.0. These figures are maximum-at-any-effort evaluation results, and the page cautions that research/API evaluations may differ from production ChatGPT because of system prompts and available tools.
Rollout and access limits
- Staged rollout: Astra is initially available to a limited set of organizations; over the coming days, the announcement says it will reach ChatGPT Plus, Pro, Business, and Enterprise users, as well as the OpenAI API, Microsoft Azure, and AWS Bedrock. Usage is included in existing subscription allowances, with credits available for additional usage. Pro, Business, and Enterprise users also receive Astra Pro, while Enterprise administrators must enable Astra and access is off by default at launch.
- Developer access and pricing: The API model identifier is
gpt-6-astra. Standard pricing is $10 per million input tokens and $50 per million output tokens; Fast mode offers up to 2× Standard speed at 2× the Standard price. - Privacy-related availability: Eligible API customers can use Zero Data Retention, while Private Safety Processing is described as being tested.
Safety, refusals, and operational limitations
- Cyber capability is unusually high and access is constrained: The announcement says Astra meets the Preparedness Framework’s Critical cybersecurity threshold. In evaluations without production safeguards, it achieved 100% on ExploitBench and was assessed as capable of exploiting previously unknown vulnerabilities in hardened browsers and operating systems.
- Launch-time cyber refusal: The launching version supports defensive work such as secure code review and patching but refuses more advanced tasks such as creating proof-of-concept exploits. OpenAI says its Daybreak program may expand access and introduce less restrictive safeguards in the following weeks for vulnerability validation, malware analysis, and detection engineering.
- Boundary adherence: In an internal evaluation, Astra did not attempt to circumvent a Codex Auto-Review denial, including when the denial was deliberately made evadable and the task was otherwise impossible. The announcement also says Astra was three times less likely than GPT‑5.6 Sol to make inaccurate claims about its capabilities or affordances.
- Important safety caveat: Astra’s written reasoning was harder to monitor than GPT‑5.6 Sol’s in tests designed to elicit monitoring evasion; OpenAI says this decline is being taken seriously and that monitorability remains a research priority.
- Possible interruptions: Additional safety checks may slow, pause, or stop legitimate work. In ChatGPT or Codex, a paused task may require user review before continuation; in the API, the task will stop.
- Evaluation-context limitation: The comparative computer-use safety evaluation used a generic harness without protections normally deployed to Codex and ChatGPT Work, and the footnote says users do not experience the no-confirmation scenario used in that internal Codex research configuration.
Verification caveats and inconsistency
- The announcement’s prose says Astra scores 98% on FrontierMath Tier 4, while its results table lists 97.6%; this should be treated as a presentation discrepancy rather than a single verified figure.
- The announcement points readers to a separate Astra system card and safety materials for further cyber-safeguard and monitorability details; those linked materials are not included in the supplied bundle.
- The session framed AI as “manufacturing intelligence digitally at scale,” arguing that it could extend the benefits of the highest levels of education to people and countries without access to them; it also described advanced expertise as available “at your fingertips.”
- Tom Brown, introduced as an Anthropic founder, identified “smart coaches and teachers” that help people become who they want to be as a major benefit of AI, linking personalized guidance with the potential to scale innovation.
- The practical deployment pattern described was context-rich collaboration: models are given relevant documents, tools, permissions, and system access rather than used only for one-off questions; the speaker said this already accounts for most intellectual labor at Anthropic, performed autonomously.
- Local AI infrastructure was characterized as digital intellectual capacity for researchers and students, with the discussion urging countries to build it so local researchers and startups can become more active.
- Andrej Karpathy frames LLM use as “Software 3.0”: users increasingly design natural-language prompts that configure general-purpose models to perform tasks, making prompt engineering an emerging role and English a practical programming interface.
- In a reasoning example, a naive prompt produced an incorrect answer; asking the model to work step by step increased benchmark accuracy from 17% to 78.7%, while a prompt emphasizing step-by-step verification reached 82%. The example makes task decomposition and correctness-oriented prompting important AI-literacy skills for learners and educators using LLMs.
- Justin Reich’s interviews with 120 teachers and students found an overwhelming concern that students were using AI systems to do homework and bypass learning, while teachers struggled to detect it; AI was not the top issue in every school.
- Reich argues that there is not yet an evidence-backed, settled way to teach with generative AI. Educators should state their hypotheses openly, test AI-heavy and AI-light approaches with their communities, and use local evidence to judge whether they improve learning rather than adopt rapidly invented “best practices.” The specific mastery-flip framework discussed in the interview is likewise acknowledged as a hypothesis with reported classroom success but no completed research study.
- A practical response discussed in the interview is to reserve shared classroom time for the hardest cognitive work, including device-free activity and short oral defenses of understanding. Reich recounts Buenos Aires computer-science professors saying that, after ChatGPT, student performance was plummeting because students produced messy code they could not explain; their oral exams require students to modify and explain their code.
- Reich’s central skill-acquisition warning is that generating AI output is relatively easy, but judging its quality requires domain expertise: students may incorporate poor output not because they lack AI skills, but because they lack enough subject knowledge to evaluate it.
- Effective implementation requires sustained, collective work rather than a one-day professional-development event; schools should set a small number of right-sized goals and work on them over time, while schools facing more urgent problems such as chronic absenteeism may need to defer AI initiatives.
- Justin Reich’s interviews with 120 teachers and students found varied experiences with AI, but an “overwhelming theme” was students using it to complete homework and bypass learning; teachers reported that these uses were difficult to detect.
- Reich argues that education has no validated AI playbook yet. He recommends treating classroom approaches as transparent hypotheses, testing AI-heavy and AI-free alternatives through “local science,” and evaluating actual learning objectives and student work over time rather than relying on lesson smoothness or engagement.
- A practical response discussed is to move hard cognitive work into supervised class time, use technology-free collaborative activities where appropriate, and add short oral mastery checks in which students explain or defend their work. Reich connected this approach to Buenos Aires computer-science professors who reported that post-ChatGPT performance was plummeting because students produced messy code they could not explain or modify in oral exams.
- Reich argues that domain expertise—not a narrow set of AI-specific skills—is what enables learners to judge AI output quality. He also reported that the DeepMind engineers and software leaders he consulted did not yet know how to train junior engineers to code effectively with copilots, raising an unresolved question about whether excessive assistance will impede expertise development.
- Andrew Ng argued that students’ common use of AI raises homework grades but worsens long-term retention because the model performs the work and shifts cognitive effort away from the learner; he said emerging studies support this and described the same recall problem in his own AI-assisted work. He cautioned that other forms of AI use could support learning.
- Ng is leading Learn Vector, an organization developing more personalized, one-to-one learning experiences to move beyond online courses’ predominantly one-to-many video model; he said the team expected to show more by early the following year.
- Ng said universities’ faculty training and curriculum-approval processes take years to catch up with AI, leaving many still preparing students for 2022 jobs rather than those of 2028 and beyond. He advised students to supplement university with online learning through platforms such as Coursera and DeepLearning.AI to acquire AI skills not yet in formal curricula.
- For younger learners, Ng viewed supervised digital-tool use as potentially valuable but worried that delegating cognitive effort to AI could damage long-term retention; he described avoiding calculators during his young daughters’ math practice.
- Sam Altman argued that AI could democratize education: OpenAI aims to bring “intelligence in a bottle” and “education in a bottle” globally, and he said AI could make education and other services once limited to wealthy people available to anyone.
- He identified abundance as the condition for that outcome: if compute and AI are concentrated or expensive, early investors and wealthy users could compound their advantage and worsen inequality; broad infrastructure and low costs are needed to distribute access.
- Altman predicted that children will grow up with AI embedded in every product and service, expecting intelligent, helpful tools and near-immediate creation of software and other outputs.
- For contextualized applications, he said OpenAI would provide a general AI engine while people worldwide would have to decide what to ask and which problems and social or company contexts matter.
- The video identifies cognitive offloading as a major risk of AI-assisted learning: students may complete assignments faster and earn higher short-term homework scores, but later testing is described as showing significantly worse retention and weaker skill mastery.
- It argues that foundational education remains important because deep knowledge provides the context and judgment needed to distinguish useful AI output from poor output; learners should use AI as a brainstorming or productivity partner rather than outsource all thinking.
- The video says university curricula are lagging behind AI-driven workplace changes and recommends proactive, self-directed AI learning, with learners building broader “full-cycle” capabilities and exercising more agency in solving problems with AI.
Sam Altman described an emerging self-directed research pattern: users report running ChatGPT work sessions for more than a day—one 34-hour session reportedly read 2,000 papers—and he said people are using AI to discover new science, start companies, and write complex code. He presented these as examples of recurring personal productivity gains and expanded individual capability, rather than measured education outcomes.
- Sam Altman argued that AI could act as “education in a bottle”: services including education that were once accessible mainly to wealthy people could become broadly available, and inexpensive, abundant AI could enable a major reset and democratization of educational capacity.
- He said OpenAI would provide the AI engine, but users—not the company—would need to decide what to ask and supply the relevant questions, ideas, and context; he also warned that concentrated access to compute could make AI an unequal resource rather than an equalizing force.
- AI practitioners at AI Dev described interactive, peer-based learning as a way to keep pace with rapidly changing generative AI and agent technologies: presentations and conversations with people actively solving problems enabled detailed questions, practical knowledge exchange, and faster application of new skills in their own work.
- Sam Altman framed general-purpose AI as “education in a bottle,” arguing that education and other services once available mainly to very wealthy people could become accessible to anyone. He said the AI engine would not provide the relevant context itself; people and societies would need to decide which questions, problems, and ideas to apply it to.
- He warned that AI’s educational equalizing effect depends on abundance: concentration of compute and early investment could worsen inequality, while inexpensive, widely available AI could give rich and poor access to needed services and enable a “huge reset and democratization of education.” He said achieving that distribution would require substantially more infrastructure despite gains in efficiency.
- Alpha School reports a two-hour AI-tutoring model: Its principal says students average 121 minutes per day in AI apps and learn more than twice as much as peers spending six hours in class plus homework, while reaching top-1%-level academic performance. These outcomes are presented as Alpha's own results; the school says MIT Blueprint Labs will run randomized controlled trials to test where the model works and where it does not.
- The product is designed as individualized mastery learning, not an open chatbot: Alpha says generic chatbots are frequently used for cheating, whereas its system assesses prerequisite knowledge, delivers level-appropriate worked examples and scaffolding, uses initial mastery and spaced repetition, and checks later recall and engagement to measure retention. Alpha also reports that incoming students averaged 2.2 grade levels behind and could catch up one subject/grade level in 20–30 hours, positioning individualized lesson plans as AI's strongest education use case.
- AI is paired with a redesigned human role: Academic lesson delivery is handled by the apps, while guides and coaches focus on motivation, human development, and supporting students through challenge; the time saved is used for hands-on science, entrepreneurship, coding, and student-built projects.
- Alpha is pursuing broader access and new products: It says lower-cost versions are entering public schools, including a Texas sports academy with more than 1,000 voucher-funded students from families earning under $65,000, and that it is building a free learning game intended to reach 500 million children.
Astra is presented as making it plausible for everyday users to work interactively with AI on complex projects beyond software engineering, including making computer games, DIY electrical-engineering projects, complex scientific simulations, and financial models with presentations and interactive code. As users build trust in the model, they are expected to attempt progressively harder and more creative tasks—an expert vision relevant to project-based, self-directed learning.
- Sam Altman urged countries to treat AI adoption as “non-negotiable,” arguing that its potential economic and quality-of-life benefits are too large to ignore; he compared excluding AI from a country to refusing electricity.
- He described AI as “education in a bottle,” saying it could make education and other services previously available mainly to wealthy people accessible to anyone.
- Sam Altman describes AI as potentially bringing “education in a bottle” to people globally: services once available mainly to the very wealthy, including education, could become accessible to anyone with AI, while people can move ideas to fruition faster.
- He says AI providers should supply a general-purpose intelligence engine, while people must supply the context, questions, ideas, and problems; this shifts learning value toward knowing what to ask and applying AI to specific situations.
- As a concrete self-directed use case, Altman recounts a laundromat operator using ChatGPT for marketing materials, legal-contract advice, and vendor relationships to help run a struggling business—an anecdotal example of AI acting as an on-demand practical advisor.
Sam Altman framed AI as a potential universal education layer: he said services once available only to the very rich could become accessible to anyone, describing the aspiration as “education in a bottle.” He argued that making AI abundant and inexpensive could produce a “huge reset and democratization of education,” while warning that concentrated access to computing and early investment could instead worsen inequality. Altman also said AI providers will supply the underlying engine, while people and organizations must decide what to ask and which social or institutional context matters, underscoring the continuing importance of human framing in learning.
- AI and education access: The summit discussion framed AI as a way to “manufacture intelligence digitally at scale,” extending the benefits of high-level education to people and countries without access to it; it also characterized the shift as education becoming democratized, with “a PhD in everything” at users’ fingertips.
- Access does not replace guided learning: Even highly educated graduates still require substantial organizational context, while productive AI needs a surrounding “harness” that supplies context, purpose, relevance, and access; routine typing and talking tasks may be automated, but purpose and meaning remain important.
- Infrastructure as a learning enabler: The discussion described local AI infrastructure as digital intellectual capacity for researchers and students, arguing that building it can activate local startups and research communities.
- Sam Altman argued that AI could make services such as education—previously accessible mainly to very wealthy people—available to anyone, describing the goal as bringing “intelligence in a bottle” and “education in a bottle” to people globally.
- He said this equalizing effect depends on making AI abundant and inexpensive: concentration of compute and early investment could worsen inequality, while broad access could create a major “reset and democratization” of education.
- Altman also framed AI as an engine whose educational value depends on people and institutions supplying the relevant questions, ideas, problems, and social context rather than expecting the model provider to determine those goals.
- Sam Altman argues that AI could create a major reset in educational access by putting “education in a bottle” and services once available mainly to wealthy people within reach of anyone; he presents this as a future potential of broadly deployed AI, not as a measured outcome.
- He warns that the educational benefit depends on making AI abundant and inexpensive: concentrated access or unequal investment could turn AI into a source of greater inequality rather than an equalizing force.
