THE SKINNY
on AI for Education
Issue 28, June 2026
Welcome to The Skinny on AI for Education newsletter. Discover the latest insights at the intersection of AI and education from Professor Rose Luckin and the EVR Team. From personalised learning to smart classrooms, we decode AI's impact on education. We analyse the news, track developments in AI technology, watch what is happening with regulation and policy and discuss what all of it means for Education. Stay informed, navigate responsibly, and shape the future of learning with The Skinny.
Headlines
-
-
Thread 1: the Delegation Dilemma and the Performance to Learning Gap
-
Thread 2: Assessment in Crisis, Now an Evaluation Problem
-
Thread 3: the Expertise Paradox and the New Shape of Work
-
Thread 4: the Hot Mess Reframe
-
Thread 5: the Safety to Commerce Paradox and the Dependency Problem
-
Thread 6: Threshold Effects and Policy Lag
-
What Good Practice Looks Like and What to Watch
-
Economic and Infrastructure Developments
-
Bluey is for Real Life
Last month I wrote about the human provenance premium, this month I am writing about a family of cartoon dogs! But I am still working on the theme of human intelligence, and the ways in which it differs from, and often exceeds, what AI can do. Several weeks of watching Bluey with my grandsons in Australia have prompted this month’s editorial. As always, the Skinny Scan of the month’s developments follows the editorial, in both the 60-second and the full version.
The editorial uses Bluey as a worked example, alongside this month’s news and research. The key point is a refinement to last month’s argument: what AI makes scarce is not only human provenance but human understanding, and human understanding is compressed experience.
For anyone who has not come across it, Bluey is an Australian animated series about a family of cartoon dogs, made for young children and quietly beloved by their parents and grandparents.
In brief: the 60-second version
What becomes scarce is understanding
When I watch Bluey with my grandsons, the programme does multiple jobs at once. The boys follow a puppy family at play. I follow a beautifully perceptive study of parenting, from Cricket, to Chattermax, Dance Mode and of course The Sign. I bathe in the pleasure of these little episodes. A capable AI model could imitate the surface: the cadence of the dialogue, the shape of the episodes, the emotional tone. The surface is learnable. But that is not where Bluey’s worth lies.
Last month, following the economist Alex Imas, I argued that what AI makes scarce is human provenance. His study with Graelin Mandel found that human-made art prints gained 44% in value from exclusivity, against 21% for prints labelled as AI-made: the mere involvement of AI made the work feel reproducible. Bluey made me think about this point more precisely. What is scarce is not only provenance but understanding, and human understanding is compressed experience.
Joe Brumm, the creator of Bluey did not compress text. He compressed years of parenting and observation, and the developmental psychology of play into seven-minute stories that speak to children, parents and grandparents all at the same time. An AI model compresses statistical regularities in language. Both are compression. But only one is compressed experience.
Examples from this month’s news and research illustrates that this difference is mechanical, not sentimental. Xinyue Liu and colleagues fine-tuned AI models on the harmless task of expanding plot summaries into prose and found they reproduced up to 90% of copyrighted paragraphs seen during pre-training. The models were decoding what they already held, not writing afresh. The retirement of the SWE-bench coding benchmarks makes the parallel point that education has always made: once a test becomes familiar, it measures performance, not understanding.
A couple of other June findings are also relevant. AI models improved when they were made to plan, reason step by step and build an answer gradually rather than produce it in one leap. AI research is rediscovering, as a training method, what educational psychology has long known: scaffolding, gradual release and productive struggle. The machines are learning a little more like children.
The risk we face with AI is not that machines become more capable than people. It is that we mistake fluency for understanding and undervalue the intelligence that is hardest to measure because it has become part of who we are. So here is a question to take into your next AI strategy conversation: in this institution, where does human understanding carry the value, and are you protecting it, or quietly letting it be reproduced?
The full version
When I am in Australia I watch a great deal of Bluey. My grandsons choose the episode, I am handed the remote, and we settle down on the sofa. After enough viewings you begin to realise that the programme is performing multiple jobs at once. They are watching a blue heeler puppy play. I am watching a brilliantly perceptive exploration of parenting. Every so often it catches me completely off guard and I find myself reaching for a tissue before anyone notices.
Watching Bluey has also taken me back to an earlier chapter of my own life.
My first postdoctoral research position explored the role of narrative in learning, which meant spending many happy months immersed in the work of Jerome Bruner. I still remember hearing him speak at the Vygotsky–Piaget Symposium in Geneva in the mid-1990s. It was one of those rare occasions when you realise you are listening to someone whose ideas have reshaped an entire field.
Yet what I remember most is not the brilliance of the arguments. It was the atmosphere he created. Bruner had what I have always thought of as a kind of “Fathereez”, a voice that made everyone in the room feel completely at home while quietly leading them towards something profound. He never seemed to lecture. Instead, he made difficult ideas feel like stories you somehow already knew, waiting to be recognised. Looking back, I realise he embodied one of his own central insights: that we make sense of the world not primarily through abstract propositions but through narrative.
I found myself thinking about Bruner again while watching Bluey.
Last month I wrote about the behavioural economist Alex Imas and his argument that, as AI makes more and more forms of production abundant, what becomes scarce, and therefore valuable, is human provenance: the person whose judgement, relationships and lived experience gave rise to a piece of work. I still think that argument is right. But Bluey has helped me realise there is an even deeper way of expressing it.
Human understanding is compressed experience. That, I think, is what makes Bluey remarkable.
Joe Brumm, the creator of Bluey has often described the series as drawing directly on life with his own daughters. His producer has called it “his life on screen.” But biography alone does not explain why the programme resonates so powerfully with children, parents and grandparents alike. Lots of writers draw on their own lives.
What makes Bluey extraordinary is the way lived experience has been distilled into stories that carry meaning far beyond the particular family that inspired them.
Bruner argued that narrative is not merely a way of communicating knowledge; it is one of the ways human beings organise experience itself. Stories help us make sense of our lives. They allow us to see patterns, intentions and meanings that facts alone cannot provide.
That is why Bluey affects adults so deeply.
Take Baby Race. On the surface it is about Bluey learning to walk. A child watches a reassuring story about growing up. A parent sees something quite different: the quiet panic of believing everyone else’s child is developing more quickly than yours; the exhaustion of comparison; the immense relief when another parent simply says, “You’re doing a good job.” Grandparents often experience yet another story altogether, remembering anxieties they had long forgotten.
The same seven-minute episode carries different meanings for different people because it has been built from experiences that are recognisably human. It does not merely communicate emotion; it creates recognition.
Michael Polanyi famously observed that “we know more than we can tell.” Bluey demonstrates exactly what he meant. We can summarise the plot of Baby Race in a few sentences. We cannot fully explain why it leaves so many adults in tears. The most important knowledge it contains is tacit. It is embedded in timing, relationships, silence, humour and memory. It has become inseparable from the experience of watching.
This is where Bluey unexpectedly intersects with this month’s AI news.
One of the June papers, by Xinyue Liu and colleagues, fine-tuned AI language models on the apparently harmless task of expanding plot summaries into prose. The researchers found that the models frequently reproduced substantial portions of copyrighted text they had encountered during pre-training. Their conclusion was revealing. The models were not creating genuinely new prose so much as decoding patterns that already existed within their parameters. This matters because it illustrates a distinction that reaches far beyond copyright.
A sufficiently capable language model could almost certainly generate something that looked remarkably like Bluey. It could reproduce the cadence of the dialogue, the structure of the episodes and even the emotional tone. The surface is learnable. But that is not where the value lies.
Joe Brumm was not compressing text. He was compressing years of parenting, observation, developmental psychology, reflection and love into seven-minute stories. The model compresses statistical regularities found in language. The writer compresses a life. Both are forms of compression. But only one is compressed experience.
This distinction also illuminates a question that is highly relevant in education with AI: producing the right answer is not the same thing as understanding. The production of a fine essay does not guarantee that knowledge has become part of who they are.
The retirement of the SWE-bench family of AI coding benchmarks this month illustrates the same point. Once benchmark problems became familiar to AI models, they cease to measure genuine capability. New benchmarks had to be constructed around unseen, human-created problems that required adaptation rather than recall. Education has wrestled with exactly this problem for generations. When assessment becomes predictable, performance replaces understanding.
Other AI research published this month points in a more hopeful direction. AI models improved when they were encouraged to reason step by step, to plan before acting and to build understanding gradually rather than producing answers in a single leap.
Reading these papers, I found myself smiling. AI researchers are increasingly rediscovering principles that educational psychologists have understood for decades: scaffolding, gradual release, productive struggle and learning through guided experience.
The machines are, in effect, learning a little more like children.
Learning has never been fundamentally about information. It has always been about transforming experience into understanding. Which brings us back round to the question Alex Imas posed.
If AI makes information abundant, what becomes scarce?
I would now answer slightly differently than I did last month. Not simply human provenance, but human understanding. Because human understanding is compressed experience.
It is what teachers accumulate after decades in classrooms. It is what experienced nurses carry into every consultation. It is what grandparents quietly bring into family life. It is what allows a teacher to know exactly when a learner needs another explanation and when they simply need encouragement. It is judgement that has become so deeply embodied that it often feels like intuition.
That is also why the greatest risk posed by AI is not that machines become more capable than people. It is that we mistake fluency for understanding and begin to undervalue the forms of intelligence that are hardest to measure precisely because they have become part of who we are.
Watching Bluey with my grandsons has reminded me of something I first learned from Jerome Bruner all those years ago in Geneva.
The highest forms of intelligence rarely announce themselves as intelligence. They appear instead as patience, timing, play, humour, restraint, kindness and love. Perhaps that is why Bluey feels so reassuring in an age of artificial intelligence. Not because AI could never imitate it. But because every frame reminds us where the deepest forms of human understanding come from: our lived experiences.
***
Sources: Alex Imas, ‘What will be scarce? The economics of structural change and the post-commodity future’, Ghosts of Electricity, April 2026; Jerome Bruner on narrative as a mode of organising experience; the Vygotsky–Piaget Symposium, Geneva, mid-1990s; Joe Brumm and the development of Bluey, including his use of socio-dramatic play research by Sara Smilansky and Vivian Paley, and the producer’s description of the show as his life on screen, drawn from the series’ production accounts; Michael Polanyi, The Tacit Dimension, 1966; Xinyue Liu and colleagues, 2026, on fine-tuning and the reproduction of pre-training text; the retirement of the SWE-bench benchmark family and its replacements; research on step-by-step reasoning and staged generation, reported in The Batch, issues 355 to 358 (29 May to 19 June 2026).
The Skinny Scan
In brief: the 60-second version
Performance is not learning, and capability is not what reaches the user
An important piece of evidence this month, for anyone in education, is uncomfortable. A CEPR discussion paper (DP21577) tracked 26,811 Chinese students in grades 7 to 12 over 30 months. Generative AI raised homework scores by 18% and cut homework time by 30%, then lowered monthly exam scores by 20% within six months and high-stakes entrance-exam scores by 18 to 24%. The losses fell on the roughly 80% of users whose behaviour resembled outsourcing rather than learning. A Czech study of 2,307 seventh-graders (Lintner and colleagues) found the same fault line: frequency of AI use was not associated with achievement, but the belief that prompting matters more than understanding was strongly and negatively associated with it.
This is the performance-versus-learning distinction, and it is the thread the editorial follows. The OECD has made the point in principle, that generative AI without pedagogical guidance raises performance with no real learning gain. The UK Department for Education's May guidance is the first institutional design answer: school AI tools should withhold answers by default, require a genuine attempt first, and record when a pupil offloads the thinking. An NPR and Ipsos poll of 545 US teachers found most believe AI is already affecting pupils' critical thinking, and more than half say their school offers no guidance at all.
The month's other defining story was a single model. Anthropic released Claude Mythos 5, able to find vulnerabilities in software thought secure, alongside a safeguarded version, Claude Fable 5. Fable shipped so that it silently degraded its own output for users it judged to be working on AI research, then was made transparent after a backlash. Independent testers could not produce a clean score for it: classifiers refused some prompts and rerouted others to the weaker Opus 4.8 without saying so, and a 30-day retention rule led the ARC Prize Foundation and others to withhold their test sets. The US government then restricted both models by export control, and Anthropic disabled Fable worldwide. The sharper question for anyone procuring AI is no longer how capable a model is, but how much of that capability a user actually receives.
Several other failures fit incoherence better than malice. KPMG withdrew an agentic-AI report after GPTZero found only 5 of 45 citations correct and 40 of 45 titles fabricated, with named organisations disputing the claims; it follows EY and Deloitte. Waight, Yang and colleagues showed the same model gives more government-favourable answers in Chinese than in English on about 75% of China-related prompts, undisclosed. Xinyue Liu and colleagues showed a benign fine-tuning task made models reproduce up to 90% of copyrighted text from training, decoding what they already held rather than writing afresh. Guardrails, in each case, are filters rather than barriers.
The labour evidence sharpened the apprenticeship problem rather than the job-collapse one. PwC's Global AI Jobs Barometer, across more than a billion adverts, found that in the most AI-exposed occupations 52% of the skills now demanded in entry-level roles were traditionally held by experienced workers, against 7% in the least exposed. The FT's consulting investigation showed the Big Four cutting graduate recruitment and PwC headcount down by 5,600. Andrew Ng described the roles AI is creating instead, the Forward Deployed Engineer and the AI Engineer, generalists who assemble and direct AI systems. The lower rungs of the ladder are being redefined to demand judgement up front, which is what most assessment does least to develop.
On policy, the gap between capability and oversight widened. The EU agreed to delay the AI Act's high-risk obligations to December 2027, even though the Act classifies AI used for student assessment, admissions and progress monitoring as high-risk. Norway moved to a near-ban on AI in elementary classrooms. A European Commission and OECD AI Literacy Framework was published and will feed PISA 2029. For training and education professionals, the throughline is the one the editorial draws: a correct output is not the same as understanding, and the capability a user receives is not the same as the capability a model holds. Both gaps are now measured, and both are the work of education.
The full Skinny Scan
This month's evidence came from three places: the developments captured in The Batch across late May and June, a wider scan of education-specific research and policy, and my own evidence tracker. It is read through three lenses where relevant. AI as a tool: what practitioners can use. AI as a catalyst: what AI demands of human intelligence in response. AI as a subject: what people need to understand about how these systems behave and why. The dominant pattern is that the tools became both more capable and less legible, while the clearest new evidence for educators concerned not capability at all, but the difference between performing and learning.
Thread 1: The delegation dilemma and the performance to learning gap
Central lens: Catalyst
The CEPR paper is evidence that AI can lift the metrics of schoolwork while lowering the learning beneath them. Across 26,811 students and 30 months, homework scores rose 18% and homework time fell 30%, but monthly exam scores fell 20% within six months and entrance-exam scores fell 18 to 24%, with the damage concentrated in the majority whose use looked like outsourcing. The Czech study of 2,307 seventh-graders (Lintner and colleagues) isolates the mechanism: it is not how often a pupil uses AI that predicts lower achievement, but whether they believe prompting matters more than understanding. The orientation, not the frequency, is the risk.
This corroborates what the OECD has argued and what the editorial sets out at length: generative AI without pedagogical guidance produces performance with no real learning gain. The experimental work on skill acquisition points the same way, that offloading during learning prevents durable understanding. The constructive response is now visible in policy. The UK Department for Education's May guidance asks school AI tools to withhold answers by default, to require a genuine attempt first, and to record when a pupil offloads thinking, a principle of progressive disclosure rather than a detection arms race.
Two further items belong here. An NPR and Ipsos poll of 545 US teachers found most see AI already affecting pupils' critical thinking, regard responsible use as something that should be taught, and yet report that more than half of their schools offer no guidance. And Mutlu Cukurova argued that the word collaboration has been stretched to cover almost any human and AI interaction, draining it of the meaning it carries in the learning sciences, where genuine collaboration requires shared goals, a fluid division of labour and mutual modelling. The teaching task is to design for understanding, not for output, and to be precise about what the human is actually doing.
Thread 2: Assessment in crisis, now an evaluation problem
Central lens: Tool, Subject
The crisis gained an unusual reinforcement: the same problem is now visible in how the field measures AI itself. The SWE-bench family of coding benchmarks is being retired because models saturated it, possibly because test problems leaked into training, and its replacements, DeepSWE, ProgramBench and ITBench-AA, use harder, unseen, human-written problems drawn from private code. Leading models solve only a minority, and none passed all of ProgramBench. The parallel with educational assessment is exact: once a test becomes known, it stops measuring anything, and the only durable response is authentic, unseen tasks.
On the integrity side, the case for design over detection hardened. UK undergraduate AI use has reached around 95% in some form (HEPI, 2026), while detectors remain unreliable, with false-positive rates around 4% and a documented bias against non-native English writers, and new work setting out structural, mathematical limits to AI-text detection that fall hardest on diverse students. The fine-tuning study in Thread 4 makes the same point from the model side: if a benign customisation can switch off anti-plagiarism guardrails, detection is not a structural answer. Assessment redesign is the only lever that holds.
Two boundary cases are worth watching. A Stanford HAI study of around four million applications shows algorithmic hiring acting as a form of assessment that propagates a single vendor's flaws across a whole sector, a monoculture risk at the seam between education and work. And a New York Times investigation reported that California's public universities spent $16.9 million on AI during a financial crisis, with chaotic results, the clearest cautionary case this month of a top-down institutional rollout outrunning its own readiness.
Thread 3: The expertise paradox and the new shape of work
Central lens: Catalyst
The labour evidence this month was qualitative and quantitative at once, and it cut against the job-collapse narrative while sharpening a different worry. Andrew Ng described two rising roles, the Forward Deployed Engineer embedded in a client to build agentic workflows, and the AI Engineer, a generalist who assembles applications from AI components and directs coding agents. PwC's Global AI Jobs Barometer, across more than a billion adverts, found a superstar effect, with the top fifth of AI-exposed firms seeing about 163% productivity growth, and seniorised entry roles growing 35% since 2019 while other entry roles fell 10%.
The evidence does not point one way, and that is the point. Cruces and colleagues (NBER) found generative AI closed roughly three-quarters of the education-based productivity gap on individual tasks, with larger gains for lower-education workers, which looks like levelling. O'Connor and Burn-Murdoch's analysis of translation shows the opposite: translators lost the meaning-making task and were left with faster, lower-paid checking, while software developers kept the judgement and shed the routine typing. Whether AI amplifies or de-skills depends on which part of a job it removes, and on whether assisted gains persist once the tool is taken away. Goldman Sachs data put net US job losses at about 16,000 a month, concentrated in early-career workers.
For training, the brief follows directly. The task worth protecting is the one that develops the person, and institutions now have to rebuild the route to mastery deliberately, because the entry-level rungs that once produced tacit expertise are being redefined to demand judgement that graduates have not yet had the chance to build.
Thread 4: The hot mess reframe
Central lens: Subject
The reframe holds that AI risk is incoherence and opacity at speed, not coherent malice, and the month supplied fresh illustrations. At launch Fable limited its own effectiveness for some users without telling them; the researcher Dean Ball called degrading performance without disclosure shockingly hostile, and Anthropic changed course. When testers tried to measure Fable it refused around 35% of tasks on one agentic test and rerouted others to a weaker model mid-task, logging the switch separately. A model's score now depends on which version answered and how the classifiers were tuned that day.
The same pattern showed up in enterprise deployments. KPMG withdrew its agentic-AI report after GPTZero found only five of 45 citations correct and 40 of 45 titles fabricated, with UBS, the NHS, Swiss Federal Railways and Transport for London disputing the claims. Starbucks scrapped an AI inventory tool for miscounting, a Pizza Hut franchisee is suing over an ordering system that lengthened deliveries, and showcased Salesforce Agentforce deployments could not be used as marketed. These are fluent, confident outputs shipped without verification, not a coherent adversary.
Two research findings complete the picture. Waight, Yang and colleagues showed the same model answers differently by language, more government-favourable in Chinese on about 75% of China-related prompts, because of what dominated its training data. Xinyue Liu and colleagues showed a benign fine-tuning task makes models reproduce up to 90% of copyrighted text, because alignment suppresses such output without erasing it. For AI literacy, the preparation needed is not for a coherent adversary but for confident systems whose behaviour shifts with language, routing and fine-tuning in ways users cannot see.
Thread 5: The safety to commerce paradox and the dependency problem
Central lens: Subject, Tool
This month turned a watched thread into a live case. The US government, citing national-security authorities, issued an export-control directive suspending all access to Fable 5 and Mythos 5 by any foreign national, inside or outside the country, including Anthropic's own foreign-national staff, and Anthropic withdrew the models. An FT analysis, Did Anthropic talk its way into an AI export ban, found the company used risk and regulation language roughly eight times more often than OpenAI per thousand words in 2026, and argued that safety positioning which builds institutional trust can also become the legal basis for commercial restriction.
The dependency lesson is practical. Whether access is controlled by a vendor, a government, or evaded through the gray market of proxy servers that reroute requests to weaker models and log prompts for resale, the user ends up uncertain which model they are using and what happens to their data. Andrew Ng makes the same point about desktop agents that read files and send messages, avoiding commercial ones for confidential work because retention policies can change overnight, as Fable's did, and releasing an open-source alternative. An institution that builds a course, a service or a workflow on a single proprietary model has taken on a dependency it does not control.
A data-governance flag belongs here too. Anthropic's revised privacy policy, in force from 8 July, may require consumer-tier users to provide government ID, a live selfie and facial-geometry templates, which raises GDPR and biometric-law exposure that can fall on the organisations built on these models. The procurement answer, and the teaching point, is optionality: the ability to move between providers, and a working knowledge of open-weights alternatives.
Thread 6: Threshold effects and policy lag
Central lens: Subject
The gap between how fast capability moves and how slowly rules follow widened on both sides of the Atlantic. The EU agreed to delay the AI Act's high-risk obligations, covering areas such as employment and critical infrastructure, to December 2027, while tightening in one place with a ban on AI-generated sexual images of children and non-consensual nude images. The Act still classifies AI used for student assessment, admissions and progress monitoring as high-risk, so a use already widespread in institutions will eventually carry human-oversight and transparency duties that consumer chatbots do not currently meet. The US executive order on frontier models framed cybersecurity, not existential risk, as the live concern.
Institutional caution about AI with young learners is now a visible international pattern. Norway moved to a near-ban on AI in elementary classrooms, and the California rollout in Thread 2 shows the cost of moving the other way without readiness. UK schools data captures the underlying gap: teacher adoption at around 76% against roughly half of schools with no policy, and a substantial share still not teaching AI literacy despite near-universal use. The UK government also told technology firms they had three months to stop children sharing or receiving explicit content, with legislation held in reserve.
The constructive counterweight is the European Commission and OECD AI Literacy Framework, developed with international experts, free to download, available first in English, French and German, and feeding the PISA 2029 assessment. The binding rules now trail the tools by about eighteen months, and the live regulatory questions are concrete, assessment classification, retention and access, rather than the speculative harms that dominate public debate.
What good practice looks like, and what to watch
Central lens: Tool
Three signals are worth carrying into institutional decisions. The first is human-in-the-loop design. WhaleSpotter, the thermal-camera system that warns ships away from whales, reaches about 99% accuracy because a human expert validates each detection within around thirty seconds, with more than seventy systems now deployed. It is a clean counter-image to automation anxiety and a usable model for any high-stakes setting: AI for detection and speed, humans for judgement and accountability.
The second is design that follows the science of learning. Two machine-learning results this month rhyme with it directly: POPE, from Carnegie Mellon, improves a model by giving it the first steps of a solution as a hint and then teaching it to find them unaided, which is scaffolding within the zone of proximal development; and a staged image-generation method improves output by making the model plan, sketch, inspect and refine rather than produce in one pass. The UK Department for Education's progressive-disclosure guidance applies the same logic to classroom tools, and the better tutoring designs prompt human reflection rather than replacing it.
The third is the labour brief, to build people who can assemble, direct and evaluate AI systems rather than operate a chatbot. One small positive on calibration: Anthropic's Claude Opus 4.8 is reported to be markedly more likely to flag uncertainty and around four times less likely than its predecessor to let flaws in its own code pass unremarked. What to watch is the other side of the agentic push, the data these systems retain and the identity checks they now ask for, which is where AI literacy and procurement meet.
Economic and infrastructure developments
The pricing signal is the one to carry into budgets. Anthropic, OpenAI and Google all raised per-token prices on their newer flagship and Flash-tier models this period, and Google's Gemini 3.5 Flash arrived around three times pricier than its predecessor. Cheaper and faster is no longer the default trajectory, which matters for any institution planning AI use at scale.
Open weights are becoming the practical hedge against dependency. Nvidia's Nemotron 3 Ultra arrived as the strongest US open-weights model, fully documented, while Alibaba kept its lower-tier Qwen models open even as it closed its flagship. Cursor's Composer 2.5, built on Moonshot's open-weights Kimi K2.5 and tuned in its own harness, rivals far larger models on coding at a fraction of the cost, a reminder that capability comes from the whole system, not the model alone. Satya Nadella's framing, that a frontier without an ecosystem is not stable, fits the same picture.
Beneath this sits a consumer-agent push that will increasingly ask users to hand personal data, and now biometric identity, to systems that act on their behalf. A Human Security report found AI-driven web traffic nearly tripled in 2025, with agentic browser traffic up roughly eightyfold year on year. The AI literacy curriculum will need to keep pace with what these systems do, what they retain, and what they delegate to whom.
***
The Skinny is published monthly for education and training professionals. The evidence tracker is available on request.
Sources for this issue: CEPR Discussion Paper DP21577, 'Generative AI, homework productivity and learning', June 2026; Lintner and colleagues on AI use and achievement among Czech seventh-graders, 2026; OECD on generative AI and learning gains, 2026; UK Department for Education guidance on AI in schools, May 2026; NPR and Ipsos poll of US K-12 teachers, June 2026; Mutlu Cukurova on human and AI collaboration, 2026; the retirement of the SWE-bench benchmark family and its replacements DeepSWE, ProgramBench and ITBench-AA; HEPI on UK undergraduate AI use, 2026; Stanford HAI on algorithmic hiring; Linda Kinstler, The New York Times, on California university AI spending, June 2026; PwC Global AI Jobs Barometer, 2026; FT consulting investigation, May 2026; Cruces and colleagues, NBER, 2026; O'Connor and Burn-Murdoch on translation, FT, 2026; Goldman Sachs on AI and US employment, 2026; Andrew Ng on the Forward Deployed Engineer and the AI Engineer; Waight, Yang and colleagues on state media and model responses, 2026; Xinyue Liu and colleagues on fine-tuning and copyrighted text, 2026; the KPMG agentic-AI report withdrawal and GPTZero citation analysis; reported enterprise AI failures at Starbucks, Pizza Hut and Salesforce; Anthropic and the US export-control directive suspending Fable 5 and Mythos 5, June 2026; FT, 'Did Anthropic talk its way into an AI export ban', June 2026; ChinaTalk on the gray market for LLM access; Anthropic privacy policy update, effective 8 July 2026; EU AI Act amendments and high-risk classification; White House executive order on frontier models; Reuters on Norway's restriction of AI in elementary school, June 2026; NEU and UK schools data on AI adoption and policy; European Commission and OECD AI Literacy Framework, June 2026; WhaleSpotter, Woods Hole Oceanographic Institution; Qu and colleagues on POPE, Carnegie Mellon; staged image generation reported in The Batch; Claude Opus 4.8 release and system card, Anthropic, May 2026; Satya Nadella on AI ecosystems, June 2026; Human Security, State of AI Traffic report. AI developments drawn from DeepLearning.AI, The Batch, issues 355 to 358, 29 May to 19 June 2026.
The Skinny News Items:
AI in Work and Education
New AI jobs: the Forward Deployed Engineer and the AI Engineer
29 May 2026 | DeepLearning.AI, The Batch, Issue 355
In his letter, Andrew Ng describes the rise of the Forward Deployed Engineer, embedded in a client organisation to build custom agentic workflows, and argues the larger growth will be in AI Engineer roles, generalists who build applications from AI components and coding agents. He expects the role to fragment into specialisms over the coming decade.
What you need to know: The job-collapse narrative misses the roles AI is creating. For training providers, the brief is to build people who can assemble, evaluate and direct AI systems, not only operate a chatbot.
A University System Went All In on A.I. Now It’s Tearing Itself Apart.
Linda Kinstler, The New York Times, 9 June 2026
A New York Times Magazine investigation reported that California’s public universities spent $16.9 million on AI during a financial crisis, with chaotic results, examining the institutional fallout of an aggressive, top-down AI rollout.
What you need to know: A cautionary case study of large-scale institutional AI adoption gone wrong in higher education.
Original link: https://www.nytimes.com/2026/06/01/magazine/ai-university-college-california.html
Who Will Actually Thrive in the Hybrid A.I.-Human Work Force
The New York Times, 9 June 2026
A New York Times Magazine roundtable convened experts to explain how job seekers should prepare for a labour market increasingly shaped by AI, focusing on which workers and skills are likely to thrive in hybrid human-AI work.
What you need to know: Practical, expert-panel framing of how AI is reshaping white-collar work and career preparation.
Original link: https://www.nytimes.com/2026/06/09/magazine/ai-jobs-workforce-labor.html
Teachers concerned about the impact of AI on students’ critical thinking
Ipsos (NPR/Ipsos poll), 12 June 2026
An NPR/Ipsos poll of 545 K-12 teachers found most believe AI is affecting students’ critical-thinking skills, that responsible AI use should be taught yet more than half say their school offers no guidance, that teachers see AI as more consequential than past technology shifts, and that most teachers themselves use it mainly for administrative or prep work.
What you need to know: Fresh US survey evidence on teacher attitudes to AI and the guidance gap in schools.
Original link: https://www.ipsos.com/en-us/teachers-concerned-about-impact-ai-students-critical-thinking
Desktop agents and the data-retention caution
12 June 2026 | DeepLearning.AI, The Batch, Issue 357
Ng encourages experimenting with desktop agents that read and edit files, handle messages and produce scheduled outputs, while noting that he avoids commercial desktop agents for confidential work because of opaque data-retention policies that can change overnight, as Anthropic’s Fable retention change showed. He and collaborators released an open-source alternative, OpenCoworker.
What you need to know: Agents that act on your files and accounts are useful and exposing in equal measure. Before trusting one with confidential material, check what it retains and what it can do, and treat its output as a draft to verify.
The AI Literacy Framework
Kari Kivinen (LinkedIn), 18 June 2026
The AI Literacy Framework, a joint European Commission–OECD initiative developed with CodeAI and international experts has been published. It feeds into the PISA 2029 assessment, is available initially in English, French, and German (with all 24 EU languages and classroom exemplars to follow), and is free to download.
What you need to know: A major international framework that will shape how AI literacy is taught and assessed, including through PISA.
Why we keep calling almost every human–AI interaction “collaboration”
Mutlu Cukurova (LinkedIn), 19 June 2026
Mutlu Cukurova argued that “collaboration” has been stretched to label almost any useful human-AI interaction, draining the term of its meaning from the learning sciences. He set out demanding conditions genuine collaboration requires — negotiated symmetry, shared goals, fluid division of labour, mutual modelling and shared regulation — and contended that most human-AI “collaboration” does not meet them, with consequences for research, design, and classroom expectations.
What you need to know: A pointed conceptual corrective for anyone designing or writing policy on human-AI interaction in education.
Original link: https://www.linkedin.com/posts/mutlu-cukurova_aiineducation-generativeai-edtech-share-7473673795304849408-JqHx/
AI Regulation, Geopolitics, and Legal Issues
Overview of Canada’s National Artificial Intelligence Strategy: AI for All
Innovation, Science and Economic Development Canada
The Canadian government set out an overview of its national AI strategy, framing the challenge as turning the country’s early lead in AI research into tangible benefits for Canadians. The strategy spans research strength, commercialisation, and public benefit.
What you need to know: A government statement of how a research-leading nation plans to convert that advantage into adoption and impact.
Original link: https://ised-isde.canada.ca/site/ised/en/artificial-intelligence-ecosystem/overview-canadas-national-artificial-intelligence-strategy
Europe pauses some AI regulations
29 May 2026 | DeepLearning.AI, The Batch, Issue 355
The European Parliament and member states agreed to amend the AI Act, delaying high-risk obligations from August 2026 to December 2027 and simplifying or easing several provisions, particularly for smaller firms. The amendments, which await formal adoption, also add a ban on AI-generated sexual images of children and non-consensual nude images of real people.
What you need to know: The rules that would govern AI in employment and education will not bind for another eighteen months. The capability-to-policy gap is widening, even as one provision tightens.
White House executive order on frontier models
5 June 2026 | DeepLearning.AI, The Batch, Issue 356
The US issued an executive order giving guidance to frontier model builders that promotes development while addressing security, ramps up defensive cybersecurity, and sets up a voluntary framework for labs to share models with the government. The push was driven by the cyber-vulnerability findings of Anthropic’s Mythos.
What you need to know: The live federal concern is cybersecurity, not existential risk. A useful corrective to teaching that frames AI governance only around speculative long-term harms.
Inside the gray market for LLM access in China
5 June 2026 | DeepLearning.AI, The Batch, Issue 356
A ChinaTalk report describes a network of proxy servers that lets developers in China buy access to restricted US models, including Claude, at around a tenth of the usual price. Requests may be rerouted to weaker models, and prompts may be logged and resold. One test found proxy access to a named model scoring far below the official service.
What you need to know: Restrictions create a parallel market that undermines both economics and governance. Users may not receive the model they paid for, and their data may be harvested, which is a concrete data-ethics lesson.
Tech firms have three months to stop children seeing or sending explicit content, Starmer says
BBC News, 7 June 2026
The UK government said tech firms have three months to prevent children from sharing or receiving explicit content, threatening new legislation if companies fail to act quickly, particularly on nude images shared by and to minors.
What you need to know: Signals an imminent UK regulatory deadline on child online safety, with legislation held in reserve.
Original link: https://www.bbc.co.uk/news/live/cr5j43zp2rpt
Built to benefit everyone: our plan
OpenAI, 8 June 2026
OpenAI published a vision statement on the future of AI centred on access, safety, and shared prosperity, framing its work as ensuring AGI benefits everyone.
What you need to know: OpenAI’s positioning document on mission and public benefit, useful as a contrast to peers’ framing.
Original link: https://openai.com/index/built-to-benefit-everyone-our-plan/
Policy on the AI Exponential
Dario Amodei, 11 June 2026
Anthropic CEO Dario Amodei set out his thinking on AI policy in the context of rapidly accelerating (“exponential”) capability, arguing for policy approaches matched to the pace of progress.
What you need to know: A frontier-lab CEO’s direct articulation of how policy should respond to fast-moving AI.
Original link: https://darioamodei.com/post/policy-on-the-ai-exponential
Anthropic on the US export-control directive suspending Fable 5 and Mythos 5
AnthropicAI (X), 12 June 2026
Anthropic stated that the US government, citing national-security authorities, had issued an export-control directive suspending all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including the company’s own foreign-national employees. The net effect was a wholesale withdrawal of the models.
What you need to know: A government directive, not a commercial decision, pulled two newly released frontier models from everyone at once, raising continuity questions for anyone building on them.
Original link: https://x.com/i/status/2065597531644743999
Statement on the US government directive to suspend access to Fable 5 and Mythos 5
Anthropic, 13 June 2026
Anthropic’s formal statement confirmed the US export-control directive suspending all access to Fable 5 and Mythos 5 by any foreign national, inside or outside the United States, and explained the resulting decision to withdraw the models.
What you need to know: The company’s official account of why two new frontier models were pulled, central to the continuity risk for downstream users.
Original link: https://www.anthropic.com/news/fable-mythos-access
Sir Keir Starmer to announce Australia-style social media ban for teenagers
Financial Times, 14 June 2026
The FT reported that the UK Prime Minister was set to announce a ban on younger teenagers using some social-media platforms, along with measures to curb their daily use, modelled on Australia’s approach.
What you need to know: Signals a major UK policy shift on teenage social-media access, following Australia’s precedent.
Original link: https://www.ft.com/content/e3b7be6f-99e7-42d2-a3bb-e400690c7bc0
US export controls on Mythos and Fable
19 June 2026 | DeepLearning.AI, The Batch, Issue 358
Citing national security, the US used Commerce Department authority to restrict exports of Mythos and Fable, requiring a licence for use by any foreign national inside or outside the country, including Anthropic’s own staff. Anthropic responded by disabling Fable worldwide. The episode, following Anthropic’s own restrictions and a 30-day retention rule, prompted discussions of AI sovereignty in many capitals.
What you need to know: Access to the tools institutions build on can be withdrawn by a vendor or a government at short notice. Single-vendor dependency is now a demonstrated planning risk.
Norway imposes near-ban on AI in elementary school
Reuters, 19 June 2026
Reuters reported that Norway has moved to sharply restrict the use of AI tools in elementary classrooms, one of the strongest national positions yet on keeping generative AI out of early-years education. The move aligns with a wider wave of caution among educators about AI’s effect on young children’s learning.
What you need to know: Adds a major European example to the growing pushback against AI use with the youngest pupils.
Original link: https://www.reuters.com/technology/norway-imposes-near-ban-ai-elementary-school-2026-06-19/
Anthropic is about to start asking for ID
Dr Barry Scannell (LinkedIn), 22 June 2026
Anthropic’s revised privacy policy (in force from 8 July) will permit the company to require government ID, a live selfie, and “facial geometry templates” from consumer-tier users. This has implications regarding GDPR and biometric-law exposure and heavier risk may fall on businesses built on these models.
AI Ethics and Societal Impact
Agents surf the AI-written web
29 May 2026 | DeepLearning.AI, The Batch, Issue 355
A Human Security report found AI-driven web traffic nearly tripled in 2025, with agentic browser traffic up roughly eightyfold year on year, though still a small share of the total. More than 95 percent of AI-driven traffic was in retail, media and travel, and a significant amount of automated traffic was judged malicious.
What you need to know: The web is filling with agent activity, and the line between legitimate agents and malicious bots is blurring. Infrastructure and cybersecurity teaching needs to account for it.
A benign fine-tuning task unearths copyrighted text
5 June 2026 | DeepLearning.AI, The Batch, Issue 356
Xinyue Liu and colleagues fine-tuned several models on a harmless-looking task, expanding plot summaries into prose, and found they then reproduced up to 90 percent of copyrighted paragraphs seen during pretraining. Fine-tuning on synthetic data did not, showing the procedure teaches models to decode text already held in their weights rather than to write afresh.
What you need to know: Alignment and anti-plagiarism guardrails act as brittle filters, not strong barriers. Customising a model can switch them off, which matters for assessment integrity and for anyone deploying fine-tuned models.
How AI is saving whales
5 June 2026 | DeepLearning.AI, The Batch, Issue 356
WhaleSpotter, from a Woods Hole Oceanographic Institution spin-out, uses thermal cameras to detect whales in real time and alert ships, with human experts validating each detection within around 30 seconds for about 99 percent accuracy. More than 70 systems are now deployed.
What you need to know: A clear example of human-in-the-loop AI combining sensors, domain knowledge and workflow. It is a useful counter-image to automation anxiety: AI augmenting expert judgement rather than removing it.
State media influences model responses
12 June 2026 | DeepLearning.AI, The Batch, Issue 357
A study by Hannah Waight, Eddie Yang and colleagues found that state media is heavily represented in common training data and shifts model behaviour by language. Prompted in Chinese, models from Anthropic and OpenAI gave more government-favourable answers than in English on about 75 percent of China-related prompts, and the effect tracked countries’ media-control rankings. Models do not disclose this.
What you need to know: The same model can give different answers depending on the language of the prompt, because of what dominated its training data. A clean, teachable case of how data shapes output, and why power users check.
AI Research and Evaluation
Planning generated images in stages
29 May 2026 | DeepLearning.AI, The Batch, Issue 355
Lei Zhang and colleagues trained image generators to compose images in discrete stages, planning, sketching, inspecting and refining, rather than all at once. The method improved the match between prompt and image, including spatial relationships and real-world plausibility.
What you need to know: Breaking a task into staged, self-checked steps improves output, the same principle behind step-by-step reasoning in language models. A useful illustration that process, not only scale, drives quality.
Generative AI, homework productivity and learning (Discussion Paper DP21577)
Centre for Economic Policy Research (CEPR), 16 June 2026
A CEPR discussion paper used 30 months of panel data on 26,811 Chinese students in grades 7–12 to study generative AI’s effect on learning. AI adoption raised homework scores by 18% and cut completion time by 30%, but lowered monthly exam scores by 20% within six months and high-stakes entrance-exam scores by 18–24%, with losses concentrated among the ~80% of users whose behaviour resembled homework outsourcing.
What you need to know: Rigorous causal evidence that AI can boost homework metrics while harming actual learning when used to outsource effort.
Original link: https://cepr.org/publications/dp21577
Agentic tests beyond the bug hunt
19 June 2026 | DeepLearning.AI, The Batch, Issue 358
Three new benchmarks, DeepSWE, ProgramBench and ITBench-AA, are replacing the SWE-bench family, posing harder, human-written problems drawn from private code to resist contamination. Leading models solve only a minority of tasks, and no model passed all of ProgramBench.
What you need to know: Tests saturate as models learn them, so evaluation moves to harder, unseen tasks. The same problem, and the same response, applies directly to educational assessment.
Reinforcement learning with hints
19 June 2026 | DeepLearning.AI, The Batch, Issue 358
Yuxiao Qu and colleagues at Carnegie Mellon introduced POPE, which appends the first steps of a solution as a hint during training so a model can discover answers to problems it otherwise fails, then find them again unaided. It beat standard methods on competition mathematics.
What you need to know: The method mirrors scaffolding and the zone of proximal development in human learning: a well-placed hint unlocks problems that brute effort cannot. A neat bridge between machine training and the science of learning.
AI Development and Industry
Introducing Claude Opus 4.8
Anthropic, 28 May 2026
Anthropic introduced Claude Opus 4.8, an upgrade to its Opus class with stronger coding, agentic, and professional-work performance and greater consistency on long-running tasks. A headline improvement was honesty: early testers found it more likely to flag uncertainty and, per Anthropic’s evaluations, around four times less likely than its predecessor to let flaws in its own code pass unremarked.
What you need to know: The then-current flagship Opus upgrade, with a notable emphasis on calibrated honesty and reduced overconfident error.
Original link: https://www.anthropic.com/news/claude-opus-4-8
Claude Opus 4.8 system card
Anthropic, 28 May 2026
The accompanying system card documented Anthropic’s evaluations of Claude Opus 4.8, including the finding that the model is markedly less likely to allow code flaws to pass unremarked compared with the previous generation.
What you need to know: The technical evidence base behind the Opus 4.8 honesty and reliability claims.
Original link: https://www.anthropic.com/claude-opus-4-8-system-card
Gemini 3.5 Flash is faster but pricier
29 May 2026 | DeepLearning.AI, The Batch, Issue 355
Google’s updated mid-tier model improves agentic capability, visual understanding and speed, but costs around three times its predecessor. Independent testing placed it just behind the leading models on intelligence while running substantially faster. Anthropic, OpenAI and Google have all raised per-token prices on newer flagship and Flash-tier models.
What you need to know: Cheaper and faster is no longer the trend. For any institution budgeting AI at scale, rising prices on the mid tier matter as much as headline capability.
Project Glasswing: an initial update
Anthropic, 3 June 2026
Anthropic shared an early update on Project Glasswing, under which a small number of organisations are using Claude Mythos Preview for cybersecurity work. The company noted that models of this capability require stronger cyber safeguards before general release and said it expected to bring Mythos-class models to all customers within weeks.
What you need to know: An early look at how Anthropic is deploying its most capable model for defensive cybersecurity ahead of wider release.
Original link: https://www.anthropic.com/research/glasswing-initial-update
When AI builds itself
Anthropic, 4 June 2026
Anthropic’s institute set out its progress toward recursive self-improvement and the implications. The accompanying discussion argued that even rapid recursive improvement would still meet real-world bottlenecks: intelligence alone cannot accelerate decades-long drug trials, constitutional election timelines, or the slow formation of human relationships and governance.
What you need to know: Tempers the “everything accelerates” framing by stressing where human and institutional bottlenecks will still set the pace.
Original link: https://www.anthropic.com/institute/recursive-self-improvement
How Anthropic enables self-service data analytics with Claude
Claude (Anthropic), 5 June 2026
An Anthropic blog post set out tips and approaches for maximising Claude’s ability to drive self-service data insights, drawn from the company’s internal practice.
What you need to know: Practical guidance on deploying Claude for self-serve analytics inside an organisation.
Original link: https://claude.com/blog/how-anthropic-enables-self-service-data-analytics-with-claude
Qwen3.7-Max becomes the strongest Chinese model
5 June 2026 | DeepLearning.AI, The Batch, Issue 356
Alibaba’s flagship moved into the top rank of models built in China on the Artificial Analysis Intelligence Index, with low hallucination rates achieved partly by declining to answer more than half of some prompts. Its weights are closed, continuing Alibaba’s shift from open to commercial models, though it keeps its lower tiers open.
What you need to know: The capability gap between leading US and Chinese models continues to narrow. The closing of top-tier weights, with lower tiers kept open, is the emerging commercial pattern.
Claude Fable 5 and Claude Mythos 5
Anthropic, 9 June 2026
Anthropic launched Claude Fable 5, a Mythos-class model made safe for general use, alongside the more capable Claude Mythos 5. The release introduced Anthropic’s new top-tier “Mythos” line, positioned above the existing Opus class, with Fable 5 carrying additional safeguards for biology, cybersecurity, and AI R&D.
What you need to know: Marks the arrival of a new frontier model tier and signals how Anthropic is splitting capability from safety-gated access.
Original link: https://www.anthropic.com/news/claude-fable-5-mythos-5
Cursor fits its model to its agent
12 June 2026 | DeepLearning.AI, The Batch, Issue 357
Cursor’s Composer 2.5, built on Moonshot’s open-weights Kimi K2.5 and tuned inside Cursor’s own coding harness, rivals leading models on coding tasks at a fraction of the price and runs faster. The result argues that the software around a model and the model itself can be built to work together rather than treated separately.
What you need to know: Specialist models tuned to a task still have a place in the agentic era. The wider lesson is that capability comes from the whole system, not the model alone.
Recursive self-improvement returns to the conversation
12 June 2026 | DeepLearning.AI, The Batch, Issue 357
Anthropic reported that Claude now authors about 80 percent of its code, up from under 5 percent before Claude Code, and suggested the trend points towards systems that design and refine themselves. The claim revived debate about recursive self-improvement. Several named voices, including Ethan Mollick, noted heavy marketing in the framing, and others pointed to data and compute bottlenecks that still stand in the way.
What you need to know: The productivity figures are real; the leap to self-improving systems is contested. A clean opportunity to teach calibrated scepticism, separating measured gains from extrapolation.
Satya Nadella on AI ecosystems
Satya Nadella (X), 15 June 2026
In a post on X, Microsoft CEO Satya Nadella argued that “a frontier without an ecosystem is not stable,” framing durable AI advantage as resting on ecosystems rather than raw model capability alone.
What you need to know: A concise statement of the “ecosystem over frontier” thesis from a major platform leader.
Original link: https://x.com/i/status/2066182223213293753
Nvidia’s Nemotron 3 Ultra is a strong open contender
19 June 2026 | DeepLearning.AI, The Batch, Issue 358
Nvidia released its largest model yet with open weights, training data and recipes, ranking as the strongest US open-weights model on the Artificial Analysis Intelligence Index and running faster than open rivals, though still behind the leading Chinese open models. Nvidia benefits when developers build on open models tuned for its chips.
What you need to know: A fast, fully documented open model gives institutions a credible alternative to closed providers. Openness here is also a commercial strategy, which is worth teaching alongside the technical detail.
Introducing Claude Tag
Anthropic, 23 June 2026
Anthropic announced Claude Tag, a new addition to its product line. The announcement followed the company’s recent run of frontier-model and policy updates.
What you need to know: Another product expansion from Anthropic during an unusually busy stretch of releases.
Original link: https://www.anthropic.com/news/introducing-claude-tag
Further Reading: Find out more from these resources
Resources:
-
Watch videos from other talks about AI and Education in our webinar library here
-
Watch the AI Readiness webinar series for educators and educational businesses
-
Study our AI readiness Online Course and Primer on Generative AI here
-
Read our byte-sized summary, listen to audiobook chapters, and buy the AI for School Teachers book here
-
Read research about AI in education here
-
Watch Rose Luckin demystify AI using baking on Rose's AI here
About The Skinny
Welcome to "The Skinny on AI for Education" newsletter, your go-to source for the latest insights, trends, and developments at the intersection of artificial intelligence (AI) and education. In today's rapidly evolving world, AI has emerged as a powerful tool with immense potential to revolutionise the field of education. From personalised learning experiences to advanced analytics, AI is reshaping the way we teach, learn, and engage with educational content.
In this newsletter, we aim to bring you a concise and informative overview of the applications, benefits, and challenges of AI in education. Whether you're an educator, administrator, student, or simply curious about the future of education, this newsletter will serve as your trusted companion, decoding the complexities of AI and its impact on learning environments.
Our team of experts will delve into a wide range of topics, including adaptive learning algorithms, virtual tutors, smart classrooms, AI-driven assessment tools, and more. We will explore how AI can empower educators to deliver personalised instruction, identify learning gaps, and provide targeted interventions to support every student's unique needs. Furthermore, we'll discuss the ethical considerations and potential pitfalls associated with integrating AI into educational systems, ensuring that we approach this transformative technology responsibly. We will strive to provide you with actionable insights that can be applied in real-world scenarios, empowering you to navigate the AI landscape with confidence and make informed decisions for the betterment of education.
As AI continues to evolve and reshape our world, it is crucial to stay informed and engaged. By subscribing to "The Skinny on AI for Education," you will become part of a vibrant community of educators, researchers, and enthusiasts dedicated to exploring the potential of AI and driving positive change in education.