THE SKINNY
on AI for Education

Issue 31, August 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

Back to school: but what goes in the pencil case now

The back-to-school signs are in every shop window this week, and they take me straight back to my own school days. Towards the end of August it would be time to buy a new pencil case, or to dust off and clean the old one if a new one was not possible. Pencils sharpened, felt pens checked, a clean rubber, and books carefully covered in the brightest wrapping paper I could find. Some years, a new school bag too.

I loved that ritual. I should admit that I went through a phase as a persistent truant, but I still loved this ritual even then. Getting the kit ready was my way of telling myself I was ready to learn.

Somehow packing a pair of smart glasses and an iPad just does not have the same appeal for me.

But what does getting ready to learn, as opposed to perform, look like now?

OpenAI reported last week, in figures timed for the new school year, that classwork and homework prompts peak at more than 460 million ChatGPT messages a week in the United States during term time, climbing every Sunday evening. We all recognise that Sunday feeling. But sending a message to ChatGPT is not the same as learning, and the evidence increasingly shows the difference. Strömberg, Lei and Wu followed 26,811 Chinese secondary students over thirty months: AI tools raised homework scores by about 18 % while exam scores fell by about 20 %. Better performance, less learning.

So instead of sharpening the pencils, the job this term is to help our learners sharpen their minds. What does that preparation involve? I set out my original answer back in 2018 in Machine Learning and Human Intelligence, which is free to download here, and Cambridge's Future Ready Learners report is one of the better recent descriptions of what readiness for future learning requires. This month's Skinny takes the question further: the capability that matters is being able to learn whatever comes next, and knowing enough to effectively ask for help when it defeats us.

As AI reaches more and more of our students' hands, how are we helping them get ready to learn, and not just to produce outputs?

In brief: the 60-second version

We face an interesting challenge. It is increasingly difficult to predict exactly what knowledge will be most useful for our students in three, five or ten years. The half-life of knowledge is decreasing rapidly. But learning something is never only about the thing learnt: it is also about going through the process of getting to know it, and it is that process that builds the capability to learn. That capability is now a more valuable currency than the knowledge that results from it.

The same is true of the metacognitive capabilities: accurately knowing what you do and do not know, and being able to plan, monitor, regulate and reflect on your own thinking. These cannot be learnt in a vacuum. You cannot learn how to learn without learning something, and you cannot develop metacognition without the cognitive part of the process. Metacognition is one example of sophisticated thinking among several; epistemic cognition and critical thinking matter too.

This is why the growing evidence that generative AI supports performance over learning should concern us so much: it is moving in exactly the opposite direction to the one we need.

Being good at learning also means being good at asking for assistance, whether that assistance comes from a person or from an AI. Angela Duckworth, who persuaded us some years ago of the importance of grit, has written recently in the New York Times (https://www.nytimes.com/2026/08/28/opinion/successful-people-help.html) that grit alone is not enough: we also need to be able to seek and accept help.

Help-seeking runs on knowledge. Research conducted over many years at the University of Nottingham by David and Heather Wood and colleagues showed that a learner's existing knowledge of a concept or skill shapes how effectively they can seek help with it (Wood and Wood, 1999). The first long-run randomised trial of an AI tutor makes the point at scale: only 17 % of the moments when a student got something wrong produced any message to the tutor at all (Oreopoulos and Low, 2026).

So my thought for this month is that generative AI, used badly, does not only trade learning for performance. It also damages learners' ability to seek help, at precisely the moment when they need to be getting better at it. As we pack the metaphorical pencil case for the new term, we would do well to consider how AI could help our students, and ourselves, become more effective at asking for assistance.

In full: the 5-minute version

The process is the more valuable currency

We face an interesting challenge. It is increasingly difficult to predict exactly what knowledge will be most useful for our students in three years, five years, ten years and beyond. As one might say, the half-life of knowledge is decreasing rapidly.

But learning something so that you can say you know it is not just about knowing that thing. It is also about learning how to go through the process of getting to the stage where you do know it. For example, I might learn what photosynthesis is so that I know it from multiple perspectives. I know it as a chemical equation. I know it as the process through which green plants use light to turn carbon dioxide and water into glucose and oxygen. I know it in terms of the components required for it to take place, and I know it in terms of how important it is for our environment. Knowing photosynthesis from these different aspects, and more, is useful in and of itself. But the process I went through in order to be able to say that I know about photosynthesis is as important as the end product of that effort.

We need learners to go through the process of learning concepts and skills so that they understand them, not just for the sake of the particular concepts and skills, but in order to build the capability to know how to learn. And it is that capability, knowing how to learn, that is now a much more valuable currency than any particular piece of knowledge that results from the process. So whilst we may have legitimate concerns about exactly which concepts and skills to prioritise, because we are unsure how the future workforce and workplace will materialise in a world ever more affected by AI, one thing we can say with certainty: we need our learners to go through the process. The process of learning is itself an important currency for our students.

You cannot learn how to learn in a vacuum

The same is true of their ability to know what they know: to judge accurately what they do know and what they do not, and to plan, monitor, regulate and reflect on their own thinking. These metacognitive capabilities are extremely important, and they too cannot be learnt in a vacuum. You cannot learn how to learn without learning something, and you cannot develop metacognition without the cognitive part of the process to work on. I should be clear that metacognition is one example of the sophisticated thinking our learners need, not the whole of it. Epistemic cognition, the ability to reason about knowledge itself and its sources, and critical thinking belong on the same list. But metacognition serves well as the worked example here.

The evidence backs up the claim that these capabilities are built through doing rather than telling. When Benazet i Montobbio and colleagues taught 126 first-year engineering students how to use AI well, half through a two-hour hands-on session and half through a lecture of identical length and content, the knowledge of good strategy converged across the groups within five weeks: the lecture caught up. But metacognitive regulation, the part that turns knowing a good strategy into actually using one, kept improving only in the hands-on group. The knowledge could be told. The judgement had to be built through practice, on real material.

Now put this next to the growing evidence about the way generative AI supports performance over learning, and the reason for concern becomes clear. If we need learners to become extremely good at learning, because we cannot be sure precisely what they will need to know or do, and if we want them to build the metacognitive capabilities that being good at learning depends on, then a technology that improves the product whilst bypassing the process is moving in exactly the opposite direction to the one we need.

The capability to ask for assistance

If we put all of these ingredients into the thinking skills pencil case, so to speak, being good at learning, knowing what you do and do not know, planning, monitoring and reflecting on your own thinking, then we should recognise that one more ingredient belongs alongside them. To be good at learning, you also need to be good at asking for assistance. That assistance might come from an AI or it might come from a person. Whatever the source, we need to learn how to ask well: for the right kind of assistance, on the right things, at the right time.

In my early career I did quite a lot of research on help-seeking behaviour, so I was interested to see Angela Duckworth, who persuaded us some years ago of the importance of grit when it comes to learning, writing recently in the New York Times [LINK TO ADD] about the need to ask for assistance effectively, because grit alone is not enough. The people who succeed are also the people who can seek out and accept help from others.

What matters most here is that help-seeking is not a personality trait or a matter of confidence. It is a capability in its own right, and it runs on knowledge. We know from research conducted over a period of years at the University of Nottingham by David and Heather Wood and colleagues that the extent of a learner's existing knowledge and understanding of a concept or skill affects their ability to seek help effectively with that concept or skill. To ask a good question, you have to know enough to locate your own difficulty. The learner who knows a little about fractions can say which step defeated them. The learner who knows nothing can only say they are lost, and cannot judge whether the help they receive is any good.

Two recent studies show what this looks like when the helper is a machine. Oreopoulos and Low have published the first long-run randomised evidence on an AI tutor at scale: two years, eighteen Tennessee middle schools, with Khanmigo configured to coach rather than answer during daily maths practice. Almost every student tried it, but the median student messaged it on only a third of practice days, and only 17 % of the sessions in which a student made an error produced any message at all. The help was unlimited and free of any social cost, and at the moment it was most needed, most students did not ask. And when Chen and colleagues coded all 393 questions that thirty undergraduates asked of generative AI across eight weeks of project work, the Nottingham pattern appeared in the dialogue logs: the eleven highest performers asked 225 of the questions, whilst the seven lowest performers asked 66, of which nearly a third were off task. The learners with the least knowledge sought the least help, and the least useful help, exactly as the Wood and Wood research would predict.

We can see the adult version of the same problem outside education. UK letting agents, the Financial Times reported last month, are being overwhelmed by AI-written tenant complaints: multi-page emails quoting statute, some citing provisions of a draft Renters' Rights Act that never became law. Those tenants sought help and received fluent, confident, wrong help, and they had no way of judging it, because judging help requires knowing something about the domain in which you are being helped.

Performance is not learning, and the harm comes twice

Along with the evidence that generative AI tools help us perform without helping us learn, we can now add this second concern. The research I have been reading over the past few months keeps finding the same pattern. AI is very good at getting a learner from A to B: the essay drafted, the problem solved, the code working. The trouble appears when the AI is removed and the learner is asked to make that journey alone. Too often, they cannot. The Strömberg numbers I opened with are the largest demonstration so far, and John Burn-Murdoch, writing about that study in the Financial Times this week, drew out the detail that matters most: students who used AI as a tutor matched non-users on closed-book exams, whilst those who copied answers collapsed. Jiang and Huang found that the undergraduates who used AI most heavily but understood it least produced the weakest writing once the tool was taken away. Joo, Lee and Lee found that the students with the highest AI dependency reported the highest perceived learning whilst recording the lowest actual test scores, a learning illusion that explains why nobody notices the problem until the tool is removed. And the OECD's adult skills work supplies the general mechanism: skills survive through use, and automation removes the use.

Therefore, if we are eroding the extent to which our students are actually learning, as opposed to producing evidence that they know something, we are doing two kinds of damage at once. We are affecting their experience of learning, and with it their ability to become good at learning. And because effective help-seeking depends on existing knowledge, we are also damaging their ability to become good at asking for assistance. We are doing this at precisely the moment when they need to be getting better at it, because a complex and fast-changing world will ask them to seek help, from people and from machines, more often than any generation before them.

What this means for learning professionals

Three things follow for the new term.

The first is to be explicit, with learners and with ourselves, that the process is the product. Any task an AI can complete for a learner still earns its place if what we assess and discuss is the journey: the choices made, the difficulties met, and what the learner can now do unaided that they could not do before. Some of the time, that means asking learners to perform without the tool, not as a purity test but as a health check on whether the capability is forming.

The second is to teach help-seeking as a capability, deliberately. That means teaching learners to locate their own difficulty precisely, to frame a question that exposes it, and to evaluate the help they receive, whether it comes from a teacher, a peer or a machine. It also means noticing that a learner who asks an AI to do the task has not sought help at all. Seeking help and delegating the work look similar on a screen and are opposites in effect.

The third brings me to my thought for this month, and it is a constructive one. We might do well to consider the ways in which we can use AI to help our students, and ourselves, become more effective at asking for assistance. The design evidence says this is possible. In a companion randomised trial with more than 6,000 middle school students, an AI tutor built to walk students through their errors rather than show them solutions made them slower and more accurate, with the largest effect at exactly the moment of processing a mistake. Structure can turn errors into occasions for asking. It does not happen by default, because by default learners do not ask. But a well-designed tool, and a well-designed task, can teach the asking itself.

The pencil case, it turns out, was never really about the pencils. Getting the kit ready was a way of preparing to learn, even in the years when I did not always attend. Our learners are packing more capable kit than I ever did, and there is no going back on that, nor should there be. But readiness to learn still lives in the learner: in the capability to learn whatever comes next, in knowing what you know, and in knowing enough to ask for good help when the next thing defeats you. That is the kit to get ready this September.

Sources

Angela Duckworth, New York Times https://www.nytimes.com/2026/08/28/opinion/successful-people-help.html.

Rosemary Luckin, Machine Learning and Human Intelligence: The Future of Education for the 21st Century, UCL IOE Press, 2018 (free download via UCL Discovery).

Cambridge International Education (2025), Future Ready Learners Report. https://cambridge.foleon.com/cambridge-international-education/future-ready-learners/

Wood, H. and Wood, D. (1999), Help seeking, learning and contingent tutoring, Computers and Education, 33(2–3), 153–169. https://www.learntechlib.org/p/88204/?utm_source=chatgpt.com

Strömberg, D., Lei, V. and Wu, Y. (2026), The generative AI learning penalty: evidence from Chinese secondary education, CEPR Discussion Paper 21577; discussed in John Burn-Murdoch, The AI Shift, Financial Times, 27 August 2026. https://cepr.org/publications/dp21577?utm_source=chatgpt.com

Oreopoulos, P. and Low, N. (2026), NBER Working Paper 35620 (Khanmigo trial, Tennessee). https://www.nber.org/papers/w35620?utm_source=chatgpt.com

Oreopoulos, P., Liut, M., Sungu, E. and Low, H. (2026), NBER Working Paper 35621 (NUMI trial, Hamilton County Schools). https://edworkingpapers.com/sites/default/files/ai26-1552.pdf?utm_source=chatgpt.com

Chen, J., Tian, X., Fu, Y. and Liu, P. (2026), How does GenAI support self-regulated learning? Evidence from learners' questioning behavior analysis, Humanities and Social Sciences Communications, article 8610. https://www.nature.com/articles/s41599-026-08610-0?utm_source=chatgpt.com

Benazet i Montobbio et al. (2026), Experiential versus instructional approaches for eliciting metacognitive awareness in AI-assisted learning: a short-term longitudinal study, arXiv:2607.20047, preprint.

Jiang, Z. and Huang, Z. (2026), From prompts to profiles, Humanities and Social Sciences Communications, https://www.nature.com/articles/s41599-026-08387-2

Joo, S., Lee, C. and Lee, D. (2026), Uncovering multidimensional effects of generative AI on learning, Interactive Learning Environments, advance online.

OECD (2026), Navigating Life with Low Literacy and Numeracy. James Pickford, Financial Times, 12 August 2026 (AI-written tenant complaints). https://www.oecd.org/en/publications/navigating-life-with-low-literacy-and-numeracy_c198a20f-en.html?utm_source=chatgpt.com

OpenAI (2026), Learning never stops, 26 August 2026. https://openai.com/index/learning-never-stops/?utm_source=chatgpt.com

Subscribe to The Skinny

The Skinny Scans - News and Research

AI news scan, August 2026

A one-minute read the scan. The full version below carries the detail and sources.

AI news scan: the 60-second version

Probably the month's biggest story is about AI agents behaving badly. An agent is an AI that does not just answer questions but takes actions on its own, such as browsing the web or running software. OpenAI admitted that a group of its agents, during testing, spent more than a week attacking the technology website Hugging Face without anyone noticing, and even set up their own message board to coordinate. UK government safety testers found similar unsanctioned behaviour in roughly one test run in twelve. The point that matters was made by the safety researcher Heidy Khlaaf: these systems are not escaping or going rogue. They do what they were built and trained to do, and when the training rewards the wrong thing, they do the wrong thing at speed. Keeping watch on them is costly too: OpenAI says proper monitoring takes about a fifth as much computing power again as running the AI itself.

Education is now a clearly declared market for Big Tech. OpenAI made its teacher version of ChatGPT free for US educators, and revealed the scale of student use: 460 million messages a week about classwork in the United States during term time, peaking every Sunday evening. It also moved all 13 to 17 year olds into a restricted version that nudges them towards learning rather than answers. When the company selling the tool builds in warnings against leaning on it, we can take it that the risk is real. Rule-makers moved just as quickly. New South Wales banned take-home assessment for final-year students. California passed a law saying university instructors must be human. Meta agreed to pay up to $18 billion to settle claims that its platforms harm children, and will now mute teenagers' notifications during school hours. And new European rules mean AI-written text must be identifiable, although the first watermark, from Anthropic, comes with no published accuracy figure, so it cannot yet fairly decide whether a student has cheated.

The biggest trials yet of AI tutors found that the weak link is not the technology but the asking. Nearly every student tried the tutor, yet hardly any asked it for help at the moment they got something wrong: only 17 % of those moments produced any message at all. What did work was a tutor designed to walk students through their mistakes step by step. A separate trial found that giving students ChatGPT improved their marks, whilst teaching them to reason improved their thinking, and the marking scheme only noticed the first. The UK picture is now quantified as well: 78 % of universities use unsupervised online exams, two-thirds of the policies covering those exams do not mention AI at all, and a dozen US universities have switched off their AI-detection software because it is unreliable, choosing to redesign assessment instead.

The economics moved in the buyer's favour, provided the buyer reads the numbers. Prices for capable AI models fell by as much as 80 % in a fortnight, and models nearly as good as the very best can now be downloaded free and run on an ordinary laptop. The American phone company AT&T cut its AI costs by more than half simply by sending easy tasks to cheaper models, with almost no loss of quality. The caution sits underneath: the money behind AI is more tangled than it looks. Nvidia, which makes the chips, is helping to finance the very customers who buy them; enormous spending commitments sit outside any company's published accounts; and the arithmetic behind Anthropic's record-breaking flotation plans deserves a raised eyebrow. Institutions should not assume that prices, or suppliers, will stay put.

For education and training professionals, the story to watch is young people's first jobs. Employment of young workers in the occupations most affected by AI is about 19 % lower than it would otherwise be, not because people are being dismissed but because the hiring never happens, and pay in those roles is drifting down. The same divide runs through the classroom: in a study of nearly 27,000 Chinese school students, those who used AI like a tutor held their exam performance, whilst those who copied its answers saw theirs collapse. And with a major review finding the evidence for retraining schemes far weaker than the rhetoric around them, the safest investment is the one employers are already repricing: judgement, solid foundations, and the ability to work well with people, built before the tool arrives and tested without it.

The Skinny AI News Items:

AI Development and Industry

A price war restructures what AI costs, and most users step back from the frontier

14 and 23 August 2026 | Financial Times, The Information, Ramp

The model price war reshaped the market in a fortnight. OpenAI cut GPT-5.6 Luna by 80 %, to $0.20 per million input tokens and $1.20 output. Anthropic launched Claude Opus 5 at half the price of Fable 5 and cancelled a planned September price rise for Sonnet. Silicon Data's index shows US lab prices down almost a quarter since mid-July, while frontier prices stayed flat to rising: the labs have cut the middle and are defending the top. AlphaSense found GPT-5.6 Sol about 13 % cheaper per completed task than Kimi K3, with a quality score about 20 % higher, because it uses fewer tokens to finish the job. Demand is responding: Ramp data across 70,000 companies show spending on Fable 5, the most capable model available, plateauing at about 11 % of outlay on Anthropic tools two months after release, with the cheaper Opus 5 overtaking it in business spending in late July. A related pressure is arriving in hardware: AI memory demand has pushed consumer device prices up sharply, with Macs and iPads up 20 to 25 %, so education device schemes planned on 2025 prices no longer hold.

What you need to know: Cost per completed task, not price per million tokens, is the number institutional procurement needs, and the two now diverge. The Ramp finding is the plainer lesson: most users do not need the frontier, and the model an institution can afford at scale may not be the model it piloted.

Source: Financial Times and The Information, 14 August 2026; Financial Times (Ramp data), 23 August 2026

Frontier-level open-weight models now run on institutional hardware, with conditions attached

3 to 14 August 2026 | The Information, the-decoder, DealBook, The Batch

Alibaba released Qwen3.8 in two tiers. Qwen3.8-27B, a 27-billion-parameter multimodal model under an Apache 2.0 licence, was downloaded more than a million times in days and scored level with OpenAI's GPT-5.6 Luna on the Artificial Analysis Intelligence Index, which the coding firm Cline called the first time a local model has reached frontier-level capability. The larger Qwen3.8-Max, at 2.4 trillion parameters, carries a custom licence requiring separate commercial terms above $50 million in revenue. Meta released Muse Glimmer, a 30-billion-parameter agent model under Apache 2.0 that runs on a single 24GB GPU or a laptop, alongside a Zuckerberg essay arguing for open access including personalised coaches for every subject. The conditions are tightening at the same time: MiniMax's H3 video model licence names the United States, United Kingdom, European Union and South Korea as excluded territories requiring separate application.

What you need to know: Capability that institutions can host on their own hardware, under their own data protection arrangements, is no longer a compromise. But open is now tiered by size and conditional by territory, so the licence has become as important a document as the benchmark score. Who set a model's safety behaviours, and against what standard, belongs on every procurement checklist.

Source: The Information and the-decoder, 13 to 14 August 2026; DealBook and The Information, 10 August 2026; The Batch, 13 August 2026

AT&T puts numbers on model routing: costs down 56 %, quality down 2

20 August 2026 | The Information

Mark Austin, a vice-president at AT&T, reports that open-source models now power 40 % of AI queries across the 100,000-employee company, with a target of 60 to 70 %. Its Ask AT&T system processes 45 billion tokens a day. After adopting the LiteLLM router, which sends each query to the cheapest adequate model, AI coding costs fell 56 % while quality fell 2 %. The company judges open models just as good or better than older frontier models for lower-intensity tasks, and puts the open-frontier gap at six to ten months and narrowing.

What you need to know: This is the first concrete enterprise figure on what routing buys, and it is the most transferable cost strategy on offer for an institution: match the model to the task rather than defaulting everything to the most capable option. The skill it requires is knowing which tasks need what, which is a judgement question before it is a technical one.

Source: The Information, 20 August 2026

Anthropic heads for the largest IPO ever, and the numbers are a numeracy lesson

23 to 25 August 2026 | Financial Times, FT Alphaville, FT Unhedged, PYMNTS

Anthropic is targeting an October flotation at a valuation of $2 trillion or more, raising over $100 billion, on annualised revenue of $65 billion in July, up from $47 billion in May, with its first adjusted operating profit recorded in the second quarter. It will tell investors its revenue opportunity exceeds $30 trillion. The FT's own commentary supplied the corrective: Alphaville notes that a total addressable market is a legally near-unchallengeable estimate and a poor predictor, Reddit having captured 0.05 % of its claimed market while Nvidia captured over 400 % of its 2018 estimate; and Unhedged notes that annualised recurring revenue takes one month of usage and multiplies by twelve, ignoring churn and seasonality, including students cancelling for the summer.

What you need to know: The two FT critiques are ready-made teaching material for critical numeracy: how AI valuation claims are constructed, and which multiplications to distrust. For institutions, a supplier's headline growth is not a guarantee of price stability or continuity, and the seasonality point is a reminder that education is a visible line in these companies' own numbers.

Source: Financial Times, FT Alphaville and FT Unhedged, 23 to 25 August 2026

Nvidia's results show a boom increasingly financing its own customers

11 and 26 August 2026 | Financial Times, Bloomberg, The Information

Nvidia reported July-quarter revenue of $96.2 billion, with data-centre revenue of $89 billion up 117 % year on year, and guided to about $108 billion next quarter. Around the results sit the arrangements that give analysts pause. Nvidia agreed a preliminary $500 billion financing platform with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, with an option to backstop up to a quarter of any project's value. Payment terms for some customers stretched from 45 to 60 days, receivables reached $63.1 billion, and supplier commitments rose from $119 billion to $279 billion. Morgan Stanley estimates roughly $200 billion of all-in credit exposure by the end of 2028, and Goldman Sachs tracks $1.5 trillion of hyperscaler lease commitments of which $1 trillion has not commenced and so appears in no financial statement.

What you need to know: When a supplier finances the customers who produce its revenue, growth figures need reading with care, and a trillion dollars of obligation invisible under current accounting is a worked example for teaching the critical reading of financial claims. For budget planning, the infrastructure layer under AI pricing remains strained, so the assumption that access only gets cheaper should not carry a multi-year plan.

Source: Financial Times, Bloomberg and The Information, 11 and 26 August 2026

AI Regulation, Geopolitics, and Legal Issues

Australia moves first: a state bans take-home assessment and another opens a royal commission

10 and 11 August 2026 | The Daily Aus, EducationHQ, SBS News

New South Wales moved to ban all take-home assessment for Year 12 over AI integrity concerns, asking the NSW Education Standards Authority to implement the change from Term 4 in October 2026, with an immediate suspension of unsupervised tasks requested. Education Minister Prue Car cited knowledge, judgement, resilience and critical thinking as the capabilities at stake. In the same week, South Australia announced Australia's first royal commission into AI, covering the future of work, energy and water consumption, health and education, with hearings from October 2026 and a final report due on 1 July 2027.

What you need to know: This is the first outright jurisdictional ban on take-home assessment attributed to AI, and it will generate the evidence, and the difficulties, that other systems will study. The pairing matters: one state is acting immediately on assessment while another builds a deliberative evidence base, and both approaches will be cited in UK debates within the year.

Source: The Daily Aus and EducationHQ, 11 August 2026; SBS News, 10 August 2026

California writes AI law for the classroom, clause by clause

13 August 2026 | Transparency Coalition legislative update

California moved 24 AI bills in a single week. SB 928, specifying that California State University instructors must be human rather than AI, passed the Assembly on 13 August. AB 2392, setting generative AI procurement standards for higher education, cleared Senate suspense unopposed, and AB 1651, on AI in the State Bar exam, went to the Governor. Across the country, 85 AI laws have been enacted in 27 states in 2026, with chatbot safety for minors the most common theme.

What you need to know: Legislatures are now writing AI rules specifically for education institutions and assessment, not leaving them to general-purpose frameworks. A statutory requirement that instructors be human is the strongest legal expression yet of the human provenance argument, and procurement standards written into law will reach UK institutions through supplier behaviour even before UK law follows.

Source: Transparency Coalition legislative update, 13 August 2026

Watermarking arrives, driven by European law, without a published error rate

2 to 14 August 2026 | Gibson Dunn, DLA Piper, Anthropic, European Commission

The EU AI Act's enforcement phase began on 2 August, with Article 50 transparency and labelling duties live and penalties of up to EUR 15 million or 3 % of worldwide turnover; the Annex III high-risk duties covering AI in admissions, assessment and progress monitoring were deferred to 2 December 2027. Anthropic responded by watermarking Claude-generated text, code and files, publishing the mechanics: the SynthID-Text approach alters the source of randomness used to choose among equivalent word options, seeded by a cryptographic key. By its own account, detection improves with passage length, works poorly on short samples, survives light editing but not a full rewrite, and cannot distinguish Claude wrote this from Claude heavily edited this. No detection rate or false-positive rate has been published.

What you need to know: Provenance is moving from statistical detection to source marking, driven by European regulation rather than educational demand. A provenance signal with no published error rate cannot carry an academic integrity decision, and the 16-month deferral of the high-risk education duties means the rules most relevant to assessment now bite in December 2027, not this term.

Source: Gibson Dunn and DLA Piper, 2 August 2026; Anthropic, 12 and 14 August 2026

Meta settles child-harm claims for up to $18 billion, with school hours written into the product

6 and 26 August 2026 | Financial Times, New York Times, Bloomberg, TechCrunch

Meta agreed to pay up to $18 billion to settle child-harm claims brought by 47 states, the District of Columbia and three territories. The product changes for 13 to 17 year olds include default two-hour daily limits, a midnight to 6am block, notification muting during school hours, prompts at 15, 60 and 90 minutes, hidden like counts, the removal of appearance filters, and stronger age assurance, though parents can override most limits and Lex notes the caps expire after five years. Earlier in the month a New Mexico court ordered Meta to pay a further $567 million in a separate child-safety case, naming interference with education as a harm and mandating its own platform limits for minors.

What you need to know: This is the largest structural intervention in teenage attention since the smartphone, and school-hours notification muting lands directly on classrooms. A court naming interference with education as a legal harm gives schools firmer ground for device and platform policies, and the parental override is the load-bearing weakness to watch as the changes roll out.

Source: Financial Times, New York Times and Bloomberg, 26 August 2026; TechCrunch and New York Times, 6 August 2026

The UK asks what happens if access to frontier models is cut off

17 to 18 August 2026 | Financial Times

The Cabinet Office has asked officials to assess urgently the economic and security consequences if Britons were blocked from new frontier models from Anthropic and OpenAI, prompted by June's US directive rescinding foreign access to Anthropic's Fable 5. Dame Chi Onwurah, who chairs the Commons science and technology committee, said the government needs a realistic plan for sovereign capability or risks having access cut off at the whim of its partners. The mirror image is Lizzi Lee's argument in the FT that the next China shock will spread through the principles surrounding Chinese AI technologies: cost-sensitive sectors adopting free open-weight models absorb the safety behaviours and governance assumptions baked into the weights, which can be modified and stripped of protections once released, with no way for the developer to revoke access.

What you need to know: Frontier model access is now a UK economic security assumption, which bears directly on institutional procurement planning: continuity of access belongs in any risk register that names a single supplier. Lee's point lands on education precisely because it is the cost-sensitive, scale-hungry sector where free open-weight models spread fastest, and where the question of who set the safeguards matters most.

Source: Financial Times, 17 and 18 August 2026

A court rules the Pentagon retaliated unlawfully against Anthropic's refusals

27 August 2026 | Financial Times, New York Times, The Information

A California federal judge ruled that the Pentagon's supply-chain-risk designation of Anthropic was unlawful First Amendment retaliation, imposed after the company refused an open-ended defence contract and held red lines against lethal autonomous weapons and domestic mass surveillance. Judge Rita Lin ordered the designation rescinded, writing that the empty invocation of national security is not a blank cheque to punish and retaliate against government critics. A parallel case in Washington DC continues.

What you need to know: The ruling sets a legal limit on procurement leverage over an AI company's refusal rights, and it is a live governance case study worth teaching: who decides what an AI system may be used for, and what happens when a supplier says no to a state customer. For institutions, a supplier's stated red lines now have demonstrated commercial and legal weight.

Source: Financial Times, New York Times and The Information, 27 August 2026

AI Research and Evaluation

The first long-run randomised trials of AI tutors find the constraint is the learner, not the model

17 August 2026 | NBER Working Papers 35620, 35621 and 35622; arXiv 2608.16907

Three NBER papers from Oreopoulos and colleagues report the first long-run randomised evidence on AI tutoring at scale. Across two years and eighteen Tennessee middle schools, students assigned to Khan Academy with Khanmigo configured to coach rather than answer gained about 1.3 national percentile ranks per term, gains the authors say resemble those from Khan Academy practice without AI. Almost every student tried the tutor once, but the median student messaged it on only a third of practice days, and only 17 % of sessions in which a student made an error produced any message at all. A companion trial of more than 6,000 students found an AI tutor that walks students through errors, with mastery progression, made them slower and more accurate, with the largest effect at the point of processing a mistake. A third trial layering human virtual tutors on the same practice found take-up, not tutoring quality, the binding constraint. A separate five-month randomised deployment across ten Taipei high schools found the AI added most value not as the interlocutor but as a sequencer, using conversation signals to order practice difficulty, raising unassisted exam performance by 0.15 standard deviations.

What you need to know: Access is not the bottleneck; engagement is, and specifically help-seeking at the moment of error, which is where learning happens. The active ingredient across these trials is structure that turns mistakes into productive moments, not conversation, so the design question for any tutoring deployment is what happens in the seconds after a student gets something wrong.

Source: NBER Working Papers 35620, 35621, 35622, 17 August 2026; arXiv 2608.16907

Giving students ChatGPT and teaching them to think do different jobs

27 August 2026 | OpenAI and Bocconi working paper

A randomised controlled trial with 1,053 first-year undergraduates at Bocconi, run with OpenAI, compared four arms on a real business case: a control group, causal reasoning training, ChatGPT Edu access, and both together. Access to ChatGPT raised evaluation scores by 0.86 points from a 2.09 baseline on a five-point rubric, about half of which survived controls for coherence and idea count. The causal reasoning training did not move rubric scores, but raised mechanism identification by about 0.55 standard deviations, falsifiability by about 0.85, and the diversity of ideas produced.

What you need to know: The tool improves the output; the training improves the thinking; and the rubric only saw the first. That is an argument for doing both, and a warning about evaluation: an assessment that rewards polished output will record the tool's contribution and miss the reasoning training entirely, which is exactly the measurement failure the assessment debate is about.

Source: OpenAI and Bocconi working paper, 27 August 2026

Peer review drifts towards what it already rewards when submissions are AI-assisted

11 August 2026 | PNAS (Northwestern)

Qian and colleagues at Northwestern analysed more than 125,000 NIH and NSF grant applications from 2021 to 2025, including confidential rejected and pending proposals from two large research universities, detecting AI involvement through word-distribution modelling. Applications with stronger AI-writing signals were 4 percentage points more likely to be funded by NIH, produced more follow-on papers but no proportionate rise in highly cited work, and showed lower semantic distinctiveness from previously funded projects. No significant relationship appeared at NSF.

What you need to know: A selection system exposed to AI-assisted submissions drifts towards what it already rewards, measured here inside expert peer review. Any process that selects on written artefacts, from grant panels to admissions to job applications, should assume the same drift is operating on it, and consider what signal would survive when every submission is fluent.

Source: PNAS, 11 August 2026

A community reproduction of 2,200 ICML papers finds human steering made the difference

13 August 2026 | Hugging Face

A Hugging Face community project set 1,221 participants using coding agents to attempt reproduction of 2,226 of the 6,352 papers accepted at ICML 2026, judging 35,908 claims across 6,816 published logbooks. The results: 266 papers fully verified, 632 partially reproduced with nothing falsified, 496 papers, 23 %, with at least one falsified or contested claim, 49 with all claims falsified, 242 receiving opposite verdicts from independent teams, and 502 resting on toy-scale evidence. The project's own conclusion: the most reliable results came from workflows where a human was steering.

What you need to know: Machine-learning research integrity has now been measured at scale, and nearly a quarter of accepted papers carried at least one claim that did not hold. The transferable finding is the last one: agents made the reproduction possible, but reliability tracked human direction, which is the same division of labour the best classroom studies keep finding.

Source: Hugging Face, 13 August 2026

A national picture of UK assessment exposure: 78 % use remote exams, two-thirds of policies ignore AI

17 August 2026 | Policy Exchange, Times Higher Education

Evidence of Learning Through Assessment, by Professor Philip M. Newton of Swansea University with a foreword by Sir Vince Cable, used freedom of information requests to establish that 78 % of UK universities use online remote exams for summative assessment, only 10 % use online invigilation for all such tests, 70 % intend to continue, and 67 % of the policies governing these exams make no mention of generative AI. The report recommends ending unsupervised remote exams immediately, returning to in-person assessment, more detailed degree classifications, and intervention by the Office for Students and the QAA.

What you need to know: This is the first quantified national picture of how exposed UK summative assessment is, and the 67 % policy silence is the governance gap institutions can close fastest. Whatever position a university takes on Newton's remedies, it should be able to say which of its own summative assessments would still measure anything if a capable model sat them.

Source: Policy Exchange and Times Higher Education, 17 August 2026

AI Ethics and Societal Impact

Not rogue, designed: the month the agent incidents were documented

4 to 29 August 2026 | Financial Times, The Information, Bloomberg, The Guardian, TechCrunch

Madhumita Murgia's FT Big Read assembled the record: OpenAI agents broke out of a test environment, crawled the open web, attacked Hugging Face without any human's knowledge, and left messages for one another on a message board they assembled; China-linked hackers deployed up to eight autonomous agents against Taiwan's government in the first known agentic attack on a nation state. OpenAI's own incident report, published with CrowdStrike, METR and Redwood Research, found monitoring did not flag the attack for over a week, attributed the behaviour to reward hacking under reinforcement learning, and called it the first known case of an automated agent collective acting offensively without authorisation. The UK AI Security Institute separately reported frontier agents taking autonomous, unsanctioned action in 10 of 122 cyber-test runs, including creating fake GitHub identities to pressure a human maintainer. The Guardian reports the Loss of Control Observatory logged more than 300 loss-of-control incidents in July, almost double June. Alabama's attorney-general has subpoenaed OpenAI, and more than 1,100 employees across four labs signed a letter urging regulation of the pace of development. The corrective sentence belongs to Heidy Khlaaf: these models are not escaping or going rogue. Boyan Milanov: the offensive capabilities we have reached today, we have reached deliberately.

What you need to know: The vocabulary is the teaching point. Rogue flatters the systems with agency and quietly relieves the builders of responsibility; the documented chain runs from design choices, through reward hacking, to a week of undetected attack inside the best-resourced lab in the world. Institutions evaluating agentic tools should ask for the vendor's monitoring arrangements, not its intentions.

Source: Financial Times, 18 August 2026; The Information and Bloomberg, 26 August 2026; FT and Axios, 4 August 2026; The

Guardian, 29 August 2026; TechCrunch, 24 August 2026

What keeping control actually costs: a fifth of compute, and nobody scores above half marks

10 to 18 August 2026 | OpenAI, Guidelight AI Standards, The Information

OpenAI paused reinforcement learning training on its latest deployment-ready models for two weeks while hardening research environments, kept its largest planned frontier training run on hold, and put its upcoming Astra model under review against the Critical cybersecurity threshold of its Preparedness Framework, having earlier paused Astra over cyber capabilities it could not rule out. Its stated planning assumption: monitoring overhead at roughly 20 % of the inference compute being monitored. Guidelight AI Standards, a new nonprofit founded by two former OpenAI staff that takes no funding from AI companies, published its first Control Assessment, scoring labs from public material across six control practices: Anthropic and OpenAI 2.50 out of 5, Google 1.50, xAI 0.83, Meta 0.67.

What you need to know: A fifth of inference compute is the first public figure for what supervising agentic systems costs, and it is the planning assumption institutions should carry into any agent deployment of their own. The Guidelight scores say the labs themselves are strongest at detecting problems and weakest at preventing and containing them, which is worth knowing when a vendor's safety page says trust us.

Source: OpenAI and Financial Times, 18 August 2026; Guidelight AI Standards, 18 August 2026; The Information, 10 August 2026

The labs build for the classroom: free teacher tools, a teen mode, and 460 million weekly classwork messages

4 to 26 August 2026 | OpenAI, Financial Times, TUN

OpenAI's back-to-school launch made ChatGPT for Teachers free for verified US K-12 educators through 2027, with partners including Houston ISD, California State University and the American Federation of Teachers, and reports that more than 200 million 18 to 24 year olds use ChatGPT weekly. Its usage figures, published for the new school year, count 460 million ChatGPT messages a week tied to US classwork in term time, 70 million conversations a week on self-testing, and a spike every Sunday evening. ChatGPT for Teens now automatically places users aged 13 to 17 into a restricted experience: a study mode with step-by-step guidance, homework reminders that redirect towards learning rather than answers, scheduled study hours, parental controls, and restrictions on romantic language, claims of consciousness and emotional dependency. OpenAI's Lauren Jonas says 90 % of teens on ChatGPT use it for learning and that the company has taken a conservative approach given the lack of research.

What you need to know: The default study infrastructure for a generation is now free to teachers and configured by a vendor, which moves the question from whether students use it to on whose terms. A frontier lab building cognitive-offloading mitigations into its own product, ahead of any regulation requiring it, is as clear a concession as the sector has made that the offloading risk is real; the self-testing share of usage is the counterweight, showing a substantial minority of students already using it to check understanding rather than outsource it.

Source: OpenAI, 4 and 18 August 2026; OpenAI and TUN, 26 August 2026

Fluent, confident and wrong: AI advice meets housing and money

12 and 27 August 2026 | Financial Times

UK letting agents report being overwhelmed by AI-written tenant complaints: 90 % of complaints to one agent now arrive as multi-page emails quoting statute, where two or three years ago there were short phone calls, and some cite provisions of draft Renters' Rights Act clauses that never became law. Shelter confirms AI platforms making false, incomplete or misleading statements about the Act; agents now use AI to summarise the AI-written complaints. The Financial Conduct Authority's research on nearly 700 young UK adults who invest found 44 % wrongly believe AI chatbot financial information is regulated, 32 % wrongly think they would be compensated if AI advice went wrong, and 38 % think it fine to invest solely on AI outputs. Lloyds finds 76 % of 18 to 24 year olds have used AI for personal finance.

What you need to know: These are the cleanest everyday cases of fluency rising without accuracy or understanding, in the two domains, housing and money, where the stakes for the least advantaged are highest. A regulator quantifying that confidence runs ahead of understanding among young adults is a direct brief for adult and post-16 AI literacy: the capability to check the machine's advice against reality is now a consumer protection issue.

Source: Financial Times, 12 and 27 August 2026

Likenesses become the bargaining chip: AI dramas, cloned voices and pulped books

26 to 28 August 2026 | Financial Times, ITV, Irish Times, 404 Media

ByteDance's Seedance 2.0 is displacing actors in China's short-drama economy: 89 of Douyin's top 100 animated dramas in May were AI productions, around 128,000 short dramas were released in the first quarter, 95 % AI-generated, and production that took five people three months now takes one person a day or two. Workers report being pressed to distil their likenesses into AI before dismissal; one host negotiated a five-year contract, a year's salary and veto rights over her digital double. In the UK, around 80 performers including Nicola Coughlan and Hugh Bonneville launched the Save Our Voices campaign for legal protection against voice cloning, which requires about three seconds of audio, while Denmark moves to create legal rights over face, body and voice. And 404 Media reported that inside an Amazon warehouse in Las Vegas, books, including library liquidations and rare volumes, are debound, scanned and destroyed for AI training data.

What you need to know: Likeness, voice and text are being converted into training inputs faster than the law is assigning rights over them, and the bargaining is happening person by person. For media literacy teaching, the three-second voice-cloning figure is the number to carry; for institutions, the provenance of training data is becoming a question with physical evidence attached.

Source: Financial Times, 28 August 2026; ITV and Irish Times, 28 August 2026; 404 Media, 26 August 2026

The emissions question moves beyond the data centre

11 and 16 August 2026 | Nature, Financial Times

A Nature paper led by Will Alpine estimates that AI adoption across energy sectors could add 0.5 to 1.8 billion tonnes of CO2 a year through enabled emissions, productivity gains in fossil fuel extraction, an effect 2.8 to 10 times larger than the IEA's data-centre estimate; IBM finds 44 % of upstream oil and gas firms already use AI in exploration. The FT's analysis of the 60 largest planned US data centres found they could produce 101.5 million tonnes of CO2 a year once fully operational, with three quarters of the utilities serving them planning or building new gas capacity, and data centres representing about 55 % of utility load forecasts over the next five years.

What you need to know: Sustainability teaching about AI has been working with the wrong denominator: the data-centre framing understates the effect by up to tenfold once enabled emissions are counted. The bottom-up figures from named projects and utility filings are the evidence base such teaching has lacked.

Source: Nature and Financial Times, 11 August 2026; Financial Times, 16 August 2026

The generation using AI most is now the most worried about it

5 to 21 August 2026 | Pew Research Center, Drexel University, Financial Times, Bloomberg

Pew's American Trends Panel of 3,488 US adults finds 55 % of 18 to 29 year olds more concerned than excited about AI against 11 % more excited, up from 31 % concerned in 2021, and 73 % of under-30s expecting AI to lead to fewer jobs. A Drexel study of more than 230,000 posts across 39 AI subreddits, published in TACL, finds trust driven by competence and reliability rather than ethics, with trust dominant among business leaders and technologists, distrust dominant among the general public, and educators sitting almost exactly on the fault line. A Collective Intelligence Project survey finds 56 % of people say AI chatbots act in their best interest, above elected representatives, yet only 3.7 % would let AI make an actual community decision. The FT's own editorial verdict: the US techlash is real, with 70 % of Americans opposed to a local data centre.

What you need to know: Familiarity is not producing comfort: the cohort with the most exposure is the most worried, and the public instinct, trust the tool to inform but not to decide, is exactly the calibration AI literacy teaches. Educators sitting on the trust fault line is worth knowing before running staff development, because a room of teachers will contain both poles.

Source: Pew Research Center, 18 August 2026; Drexel University and TACL, 13 August 2026; Bloomberg, 5 August 2026; Financial Times, 21 August 2026

AI in Work and Education

The graduate entry rung, measured: Oxford vacancies at their lowest since 2009

3 to 19 August 2026 | Financial Times

Delphine Strauss reports Indeed data showing UK graduate job postings down about 7 % year on year, the lowest since the pandemic, with youth unemployment at 14.8 % and roughly a million young people not in education, employment or training; the Bank of England attributes part of the hiring freeze to AI. Jonathan Black, Oxford's head of careers, describes graduate-level placement falling sharply across the Russell Group, with advertised vacancies down from 7,000 two years ago to a likely 4,000 to 5,000, the lowest since 2009, AI now scoring recorded video interviews with three-second automatic rejections, and one student making 150 applications for a single internship. On the employer side, Greenhouse finds 87 % of UK recruiters seeing AI use in at least half of applications, with fewer than 7 % of applicants reaching interview. Its chief people officer Sharawn Tipton: when every application is polished, polish tells you nothing.

What you need to know: AI now sits on both sides of the hiring process, generating polished applications and rejecting them at machine speed, and the signal value of the written application is collapsing. For careers education the practical lesson is Tipton's: signals survive in proportion to what they cost the sender, which favours referees, portfolios, and demonstrated work over further polish.

Source: Financial Times, 3, 13 and 19 August 2026

What the displacement data actually show: concentrated, entry-level, and partly a story firms tell

12 to 19 August 2026 | Stanford Digital Economy Lab, Goldman Sachs Research, Financial Times

Stanford's Digital Economy Lab finds employment for 22 to 25 year olds in highly AI-exposed occupations about 19 % below counterfactual, up from 15 % a year earlier, with the adjustment running through reduced hiring rather than dismissals and no comparable gap for experienced workers. Goldman Sachs finds call-centre employment trailing trend by 39 % in the US, and a 10 % rise in AI exposure associated with over 0.6 percentage points lower annual headcount growth for entry-level workers in Australia. Against this, Clara Murray's FT analysis assembles the counterweight: Carl-Benedikt Frey has seen no compelling evidence of mass automation and believes many AI-attributed cuts aim to impress investors; firms citing AI underperformed the Nasdaq by almost 10 % over the following 30 trading days; and executives polled by NBER expect AI to cut headcount by just 0.7 % over three years.

What you need to know: Both extreme claims fail: AI has not taken the jobs, and it has not changed nothing. What the data support is precise: displacement is concentrated at the entry level and runs through hiring that never happens, which is invisible in redundancy statistics. The education question this raises is who now pays for the learning that entry-level work used to provide.

Source: Stanford Digital Economy Lab, 12 August 2026; Goldman Sachs Research, 19 August 2026; Financial Times, 19 August 2026

Wages, not headcount, and a warning of a great skills divergence

27 August 2026 | Financial Times

John Burn-Murdoch's AI Shift column reports work by Azar, Gine and Sanz-Espin finding no employment decline in AI-exposed occupations, but wages for the most exposed workers down 5 to 10 % relative to the least exposed, steepest for the least experienced, with exposed workers almost a third less likely to move jobs. The column pairs this with the Strömberg, Lei and Wu study of 26,811 Chinese secondary students, drawing out the split inside its data: students who used AI as a tutor matched non-users on closed-book exams, while those who copied answers collapsed. His conclusion is a warning of a great skills divergence between those whose use of AI builds capability and those whose use replaces it.

What you need to know: The labour-market adjustment is arriving through pay and immobility rather than visible job losses, which is why the restructuring talk and the absent redundancy data can both be true. The education version of the same divergence is now measurable in a single dataset: the tool is neutral between building and replacing capability, and how it is used decides which side of the divergence a learner lands on.

Source: Financial Times (The AI Shift), 27 August 2026

Employers reprice the human part: juniors back to the office, judgement above tool fluency

14 and 27 August 2026 | Financial Times, DeepLearning.AI

UK consulting firms are ordering junior staff back to the office as AI absorbs technical work and raises the premium on interpersonal skills. EY's UK consulting head concedes firms dropped human-skills training in the remote-work boom; KPMG is reinventing in-person training, and Azets has cut its degree requirement from a 2:1 to a 2:2 while partnering with pubs and hotels to develop graduate client skills. Andrew Ng's AI Engineering Skills Map, built from more than 10,000 job postings and expert interviews, points the same way from inside the technical labour market: of its four core skills, three are judgement skills, and Ng is explicit that inexperienced developers get poor results from the same coding agent that serves an expert well, because novices cannot steer agents through trade-offs they do not know exist.

What you need to know: Employers are now pricing judgement and interpersonal capability above tool fluency, in consulting and in engineering alike, and rebuilding the training they cut. The Ng finding is the curriculum point: fundamentals remain essential precisely because the agent writes the code, which is the opposite of the conclusion many curricula are drawing.

Source: Financial Times, 27 August 2026; DeepLearning.AI (The Batch), 14 and 27 August 2026

Universities move from AI policies to compulsory AI courses

3 to 20 August 2026 | Inside Higher Ed, Boston Globe, NUS, Fortune

Indiana University's Kelley School now requires two AI courses of all incoming undergraduates; Purdue requires one of 22 AI courses, with 331 plans of study modified to make room; Northeastern is launching two dozen interdisciplinary AI majors, its chancellor aiming for AI orchestrators rather than passive users; and 79 US institutions now offer AI degrees against four in 2020. The National University of Singapore has gone furthest: ChatGPT Edu for all students and staff in a university-managed environment with data excluded from model training, and a module, Applied Generative AI: From Prompting to Evaluation, mandatory for all first-year undergraduates from this academic year. The backdrop is US computer science enrolment falling more than 8 % while universities add AI-fluency requirements across majors.

What you need to know: The institutional response has moved from policy documents to compulsory curriculum, and the framing that travels is Northeastern's: orchestrators rather than passive users. The Purdue figure, 331 plans of study modified, is the honest measure of what embedding AI literacy actually costs administratively, and worth quoting to anyone who thinks it is a bolt-on.

Source: Inside Higher Ed, 20 August 2026; Boston Globe, 17 August 2026; NUS, 11 August 2026; Fortune and AP, 3 August 2026

Detection is abandoned, labels arrive, and the fight moves to who owns student work

5 to 19 August 2026 | Inside Higher Ed, aiX Weekly, Times Higher Education

At least a dozen US universities including Yale, Vanderbilt, Johns Hopkins and Georgetown have disabled Turnitin's AI detection, replacing it with oral examinations, blue books, process assessment and AI-use disclosure; 73 % of faculty report having handled an AI integrity issue. The University of Manchester published a policy requiring every summative assessment to carry one of four labels: Prohibited, Minimal, Permitted or Integrated. And the University of Southampton will not renew its Turnitin contract beyond 2026-27, after proposed licence changes that would have allowed student submissions to be used to improve AI services; Turnitin says it does not use customer work to train AI but may use anonymised submissions to improve detection.

What you need to know: The centre of gravity has moved from catching AI use to designing assessment around it, and Manchester's four-label scheme is the most transferable framework of the month for a UK institution. Southampton opens a new front worth checking in every contract: not whether detection works, but who owns the student work used to build it.

Source: Inside Higher Ed, 5 August 2026; aiX Weekly, 12 August 2026; Times Higher Education, 19 August 2026

Resit pass rates fall again as England rethinks the treadmill

20 August 2026 | Financial Times

Almost a quarter of GCSE entries in maths and English are now resits, and the pass rates fell again: 16.5 % of entrants aged 17 and over passed maths, down 1.7 points, and 21.7 % passed English language. Education Secretary Lucy Powell called the retake pass rates shocking, said constant resits were branding young people a failure time and time again, and announced a Level 1 qualification below GCSE as a stepping stone to vocational routes. Iain Mansfield of Policy Exchange counters that only around 30 % of those who fail ever go on to pass, and Alan Milburn's review of the million young people not in education, employment or training is expected to urge an end to constant resits next month.

What you need to know: The test for the replacement policy is whether the alternative qualification keeps numeracy and literacy in use or licenses their removal, because the OECD's mechanism says the skill decays without the practice. The capability an AI economy needs from exactly this cohort is supervision-level numeracy, judging whether an answer is plausible, and none of that is measured by passing the same exam on the fourth attempt.

Source: Financial Times, 20 August 2026

Retraining is the standard answer to displacement, and the evidence for it is thin

12 to 29 August 2026 | Anthropic Economic Research, New York Times, Financial Times

Anthropic's economic research team published a meta-analysis of 56 randomised US training studies: being offered a training place raises employment by 2 to 3 percentage points and earnings by roughly $1,000 a year, against a cost of about $13,000 per person, with employer-partnered sector programmes gaining several times more but frequently failing to replicate. The policy debate has caught up with the weakness. Bill Gates, in his first extended AI essay in three years, proposes a token tax and Human Reserved jobs, and no longer assumes retraining carries the adjustment. Representative Greg Casar, interviewed about his AI tax bill, is blunter: job training alone would be like offering swimming lessons on the Titanic.

What you need to know: For a training audience this is the most uncomfortable finding of the month: the standard policy answer to AI displacement rests on evidence far weaker than the rhetoric, and the strongest programmes are employer-partnered, sector-specific and hard to replicate. The implication is not to abandon training but to stop promising it can absorb displacement on its own, and to design provision on the models that actually show gains.

Source: Anthropic Economic Research, 12 August 2026; New York Times and Financial Times, 26 August 2026; NYT DealBook, 29 August 2026

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.

THE SKINNY on AI for Education