AI should be the Guide, not the Ghostwriter

Illusion of competence

In their 2006 research, Karpicke and Roediger used rereading as the control, as that is what most learners do, see Karpicke, Butler and Roediger (2009). In a survey of 117 students, they asked them to list their study strategies, then also choose from a list of set strategies. The majority chose ‘rereading’ as a strategy with relatively few using self-testing or free recall. They christened this the ‘illusion of competence’. 

Bjork, Karpicke and Roediger believed that both teaching and learning can be improved and optimised by introducing techniques that force cognitive effort. Interestingly, they believe that teachers and learners are often delusional about learning, assuming it happens with little effort other than attending lectures, rereading and highlighting text. This is why they, and others recommend, among other techniques, generative learning. Bjorg builds his recommendations on the idea of desirable difficulty.

Desirable difficulty

Generating words, knowledge and solutions is better than simply reading, highlighting text or getting AI to do it for you. Acts of personal generation provide the context for greater understanding and subsequent recall. This is why we set essays and assignment for learners. They need to learn by genuine effort and this is achieved using ‘desirable difficulty’. 

This is a short-term pain, long-term gain idea, where desirable difficulties are learning challenges that make the learner study harder in the short term to improve long-term retention and understanding. That doesn’t in itself tell us how this should be achieved, so let’s bring in the fundamental reason for making learners create essays and assignments; generative learning.

Generative activities

Wittrock developed a generative theory of learning, as well as researching its effectiveness and applying it in practice. Learners, for Wittrock, are not passive receivers of knowledge, they are active reorganisers of knowledge creating meaning from their own generative activities. His generative learning theory was built on the idea of learners integrating new knowledge and skills into what they already know through generative activities, where effective teaching facilitates leavers to construct meaning from various generative experiences.

His model encourages learners to generate meaning and understanding from instruction through effortful, generative activities. The model has four major processes; 1) Attention, 2) Motivation, 3) Knowledge and preconceptions, 4) Generation.

Attention is directing generative processes on relevant incoming material and stored knowledge. This is what most learners use AI for, to get that initial attention on a position or starting argument. AI also helps with motivation, getting started and a willingness to really invest the time and effort to make sense of material. In particular, it helps with build that initial platform of knowledge and preconceptions. The problems come with generation, the sense making. Delegate this to AI and learning may suffer.

Scaffolding

Bruner’s four principles addressed the issue of assisting learners as they move forward in their learning process, with some concrete recommendations. 1) Readiness, 2) Structure, 3) Sequence and 4) Generation. His point is that just throwing an essay title at them is far from adequate. The learner must have a readiness in terms of a predisposition to learn and so their experiences and context must be considered. If you set them up without adequate support they will treat it as a transactional demand and respond by taking short cuts with AI. It is surely worth providing or pointing towards some sort of structure and sequencing, so that it can be grasped by the learner. But it is in the generation brings in extrapolation, manipulation, a filling in the gaps and expansion beyond the learners existing knowledge.

Bruner saw the solution these problems as scaffolding. He gave us the word ‘scaffolding’ in educational theory, and the recognition that learners need to be either self-aware or helped to build on existing knowledge, is certainly a useful device, albeit a little hazy. The problem with this constructivist generalisation is that it immediately begs more detailed questions about what we mean by ‘structure’, ‘sequence’ and ‘scaffolding’. Here, I think, the use of AI can be used to good effect.

Students should be encouraged to use AI to find, support, sequence and critique their work, not generate it from scratch. This is a vital distinction – the generation by AI of support, as opposed to solutions.

AI should be used as a dialogue between the learner and an imagined mentor. We need to accept that it will be used, because it is a useful mentor, not a generative tool in itself. It is a mistake to see contemporary AI as simply generating text. It has reasoning and can support, critique, identify gaps in reasoning and generally act as an expert mentor providing an external perspective.

What can this well researched work tell us about the use of AI by learners in learning? I’ll describe this as instructions to the learner.

Using AI as a mentor, not a ghostwriter

Many learners are simply overwhelmed by the blank piece of paper when set an essay or assignment. They can’t find a way in, a solid place to start. It seems like a mountain to climb. AI can help alleviate that fear, by helping them get started, providing the right equipment, mentoring them forward, to reach the top.

There is a big difference between AI writing an assignment for you, where you are outsourcing the effort, to using AI to help you move forward, keep up momentum and critique your own effort as you go. Do not get discouraged because you feel you don’t know what to do next. Use AI as your geode on that route, pointing out which way you should go. 

The trick is to preserve the desirable difficulty, the real learning gain in writing or doing an essay or assignment, rather than you floundering or taking shortcuts. Rather than using AI to write an essay or assignment, use it as a mentor to help you climb, one stage at a time. Imagine it is your teacher, tutor or lecturer, sitting next to you, gently guiding you forward. Let AI hold the ladder, not climb for you.

1. Just get started

Do NOT see AI as writing on your behalf, but as a tool to make your thinking more rigorous. Rather than asking AI to produce your essay or assignment submission from scratch, write a rough argument, an outline, set of notes, statement of the topic, suggested reading list or sources, anything, just to get started. Do not worry if this first version is messy, see it as a very rough starting point. Generate this on your own to start with or ask AI to generate several one paragraph starting points, then use your judgement to think about what you want as a starting position.

2. Build your first draft

Now use AI diagnostically. Instead of asking AI to produce a full essay, ask the system to suggest the best sources for researching or building a case, solution or argument. It is here you can get AI to give you pieces of the jigsaw that allow you to create a first draft. Write a first draft, in your own words. Keep it structured and simple. It is here that you will get some idea of structure and sequence. A page will suffice.

3. Diagnose weaknesses

Now use AI to identify weaknesses in your reasoning; weak logical jumps, unsupported claims, vague terms, hidden assumptions and places where an expert would say you can’t make that jump or take that direction up the rock face. AI is not helping you avoid the effort, it is helping you expose where the next stage of the work still needs to be done, a direction of travel. Use AI to critique, not compose

4. Rebuild your argument

Once those weaknesses have been identified, start climbing again, rebuild your argument. This is a crucial stage, because the improvement has to come from your own judgement. Tighten your logic, clarify any vague concepts or ideas, remove exaggeration and add supporting evidence or explanations for what was identified as missing. The value of this process lies in the climbing, looking for secure hand and footholds.

5. Check against sources

Now check your argument against the actual readings, papers or evidence you are using. AI can be helpful here by highlighting where an argument oversimplifies an author, contradicts the source material or misses an important distinction. Used properly, you push yourself up on a solid route. Instead of just citing sources you have skimmed, you are forced into a more genuine understanding of what those sources actually say and how far their own claims can be supported.

6. Identify what’s missing

Now that you have established your route upwards, now you can move into a deeper level of reflection and quality of climb. Again, ask what is still missing from your arguments, what counterarguments have not been addressed, what assumptions remain undefended and what a critic from another perspective might say. This is often the stage that turns a reasonable piece into a genuinely thoughtful one, because it forces you to see your own arguments from other, outside perspectives and recognise their limits. This allows you to push on to towards the summit.

7. Test conclusion

Before conquering the peak, test your conclusion. It may be a false peak, the real peak being just beyond what what you see, A conclusion often falls short claiming less or more than your work really justifies. Ask whether the conclusion really does follow on from the evidence, whether it overstates the case or whether there is another bolder conclusion that can be justified. You can then make it more precise and defendable.

8. Write final version yourself

Only then should you write the final version. At every point on the climb, you will have understood what you were reading and writing. At this point, you produce the finished, polished work in your own words, built on strong foundations than you have managed alone. AI has not written the essay, it has served as a critic, a preserver of method and difficulty a guide pushing you forward to your peak.

Conclusion

The purpose of setting an essay or assignment is to make you think and learn, so use the tools that allow you, not to flounder and struggle, but move forward with confidence, overcoming difficulties as you proceed. Keep the useful difficulty and preserving the learning. This gives you work that is your own, defensible and constructive.

AI assessment apocalypse – a 5-step solution

Some describe the assessment issue in higher education as apocalyptic, destroying the very fabric of higher education. This is exaggerated. This piece offers an alternative,  to turning learning and assessment into a toxic cat-and-mouse game, where there are many more mice, and the mice are winning, It is about better and more authentic assessment.

There is no silver bullet, as this is a multivariant problem involving student motivations, teaching, institutional practices and technology. Those who simply shout ‘bring back in-person exams’ are ignoring the causes and not offering adequate solutions. The solution involves several steps.

Step 1: Stop the blame

The first step is to admit the serious nature and scale of the problem but also accept it is not the students’ fault. There is a bias in teaching, research and assessment, towards the teacher and researcher. They are the means to an end, not the end in itself. The profession all too ready to blame students and dismiss AI, when these problems were explicit before AI hit the scene. It is hopelessly utopian to expect learners not to use AI. 

Rather than accuse, the solution has to be involve not tempting them with shortcuts that resut from an accusatory environment and poor assessment. Redesign the system to allow more time for teaching and eliminate the temptation that may lead to a toxic environment of accusations, false positives and expulsions; a life-changing disaster for any young person. 

It may also be a life-changing disaster for the faculty member or administrator who ends up making a false accusation. This has already happened with a Minister for Higher Education and leaders of major educational institutions, accused of plagiarism themsleves, and removed.

Most students do not want to cheat. However, when the pressure is overwhelming, from the perception of peers (everyone is doing it, I’d be a fool not to) and parents, tteaching not as good as it could be, assessments poorly designed; students will take available shortcuts. Step 1 is to recognise that cheating is normal and that in high stakes exams, people will take high stakes risks, so don’t blame the students.

We also need to cool down on ths idea that using idea is simply congitive surrender, destroying the leaner's ability to learn. There is a fundamental flaw in most debate about cognitive surrender to AI. The argument that we should be keeping learning difficult, is very different from the idea of useful, deliberate difficulty. As I said earlier, dull lectures, poor teaching, obscure content and accessibility is a big problem that requires more focus on teaching.

Students use AI because they find it useful in learning, to find things out, expand on concepts, unravel things they find difficult, test themselves, produce flashcards for revision and practice. To ban AI would be to throw the baby out with the bathwater, and the bath.

Step 2: No silver bullet

Cheating has always been rampant in education. It was there before AI, with a range of techniques and technologies. I wrote about this in my book ‘Learning Technologies’. There was a whole section on cheating technology from Confucian silk cheat sheets to repurposed calculators, false arms and even surgical implants. Cheating has been an intrinsic feature of educational assessment. As long as there are exams, people will try to take shortcuts.

Let’s take a cold, hard look at in-person essay-based exams. For generations, smart students have looked at past papers, worked out the probability of topics appearing in their exam, pre-writing essays, then memorising them for regurgitation in the exam. The assumption was that we were testing critical thinking. Yet no one who has ever written anything using critical thinking would claim that a piece of writing, written in pen from top left to bottom right of page, without redrafting, reordering and rewriting, even approximated critical thinking. Critical thinking is an internal dialogue where you think, reconsider, revise, seriously reorder and rewrite, as you proceed. This is as far away from regurgitating essays in exams as you can get.

Even in formative essay assessments, students would readily beg, steal and borrow essays from each other, get help from their graduate parents or pay essay mills. These mills were huge enterprises with tens of thousands employed in Nairobi, China and elsewhere, where the well-educated poor provided essays and dissertations for the rich. It was generally ignored by the system (no real moral outrage, as with AI) even though everyone knew it was endemic, especially among students who were studying in their second language. Why? This became a lucrative source of income, the real reason for sliding the problem under the carpet. AI suddenly became one big essay mill, free or cheap, and everyone had access to its services. The revenues of the known cheat companies plummeted.

This is why the current emergency over assessment is really the surfacing of an old and existing problem. We can pretend it is all about AI but AI has merely surfaced a deep, existing problem. It is not fundamentally an AI problem, it is a system and human nehavu=ioural problem. A lot of cheating is the artefact of of the existing teaching and assessment processes and design. It is the same problem that pushes parents to help students with their assignments, hire tutors and pay for exam prep.

Step 2 is to recognise that in-person exams may help, and are not to be scoffed at, but they are not the whole solution, not the silver bullet.

Step 3: Create a Hub

Policies are, at best sticking plasters, at worst they exacerbate the problem. They are certainly not a solution to a large and evolving problem, as they suck up collective effort, are often ignored, then just sit there unrevised and unloved until out-of-date in relation to the advances and uses of AI. 

A policy is one thing, strategy another. Rather than rant and rail around academic integrity, or blaming students, one must understand the problem, then come up withe ‘workable’ solutions. 

There is a lot of hand-wringing and ethical hubris centred around words like integrity, responsible, ethical, trust and so on. This form of abstraction comes easy to academics and administrators but it does not tackle the problems head-on. This is a practical problem that needs workable, pragmatic design solutions.

AI was the fastedst adopted technology in the history of our species and has continued to get better, as it learns. The solution, therefore, needs to involve a process, not a single policy or event. That process needs to be owned and maintained by the institution or cluster of institutions, even nationally. At first, this needs to be a one-stop-shop for advice, tools and services on assessment. This can be part of a wider hub for the use of AI in general, by all in the institution; administrators, researchers, teachers and students. A technology that is globally universal, used by almost all students to learn, warrants this level of attention. Create that hub and keep it up to date.

Step 4: Multimodal assessment

In many subjects, if you depend on just writing as proof of learning, you do not have an AI problem, you have a learning design and assessment problem. When writing is treated as the sole proof of learning for everything, AI exposure isn’t the flaw. the flaw is in the assessment design.

In the real world, all jobs involve the doing of things, dealing with people, using tools, practical tasks. 80% of jobs in the world are deskless, and those that are deskless are being automated by AI. If you do want to switch towards skills that just focus on expression through text, then other forms of teaching and assessment are necessary.

We are now on the other side of  the Gutenberg Parenthesis, where more is available in multiple media formats, from which one can teach and learn. teachers actually speak and listen, we now listen to audiobooks and podcasts, recorded lectures, videos and audio dialogue using AI.

Multimodal assessment is the optimisation of assessment by moving beyond text, now made possible, as AI has become truly multimodal. Models have integrated all media types and can ingest text, audio, images and video, as well as output these media from your text. This offers you the opportunity to free assessment from the tyranny of pure text.

One useful shift, in a world where listening and speaking to others is likely to be more useful than simply writing, is to record oral assessments and use AI to grade them. The argument against oral exams is that they take too much time, but transcription and automatic grading can speed up the process. You also eliminate the stress and problems of worrying about who used AI. Hartmann's (2025) reoriented an upper-level humanities course around oral exams and tracked the time to show that oral exams can verify student understanding directly and, importantly, that they may not take more instructor time than essay.“ Instructor time investment proved comparable to traditional paper grading” with oral exams taking 13 hours , compared to 15 for grading papers. The main point made by the paper was that oral assessment can be integrated back into your existing courses. They may even force your students to do the work, through the idea that they will be properly assessed. 

Video based assessment of performance is also becoming possible as AI can recognise what the person is doing in an uploaded video. This swings effort away from detecting AI, towards designing assessments that assess student performance. All outputs from learners can be ingested and interpreted by AI; text, audio, images and video. 

Digial portfolios are also an option for gathering evidence across the course. What is often missed is that is shows the student thinking, the ability to recall the foundational knowledge that allows thought to flourish and build a case. It shows skills, not in just writing, but a fuller form of expression and doing. This is not to say it is right for all subjects and skills but portfolios and oral exams should be part of the teaching and assessment toolkit.

A multimodal approach also offers a solution to other problems: accessibility and dyslexia. Dyslexics love AI, as it has for years offered a text to speech option. Others find that rewriting and summarising through AI translates the content into something they can understand more readily than often abstruse academic language. In other words, AI is often used by learners to simply ‘access’ and understand content. So widen out your assessment options and think beyond just text.

Step 5: Automate assessment

Automating assessment has become possible in many cases. This is not to say that all assessment should be automated, only ‘optimally’ automated. There can still be expert validation and quality checking. This has the additional advantage of freeing up busy teaching time for actual teaching.

The simple generation and marking of quizzes in formative assessment can clearly be automated. Students do this routinely. They instinctively know the ‘test effect’ works, so build their own quizzes and flashcards with spaced-practice, using AI. Many now use specialist tools like NotebookLM and ChatGPT’s education features, to help them improve the productivity of their own learning. Learning is also being integrated into tools like Google Translate, so that you can practice learning a language through role play and immersion. There is a strong argument for automating much formative assessment via platforms which gives data back to teachers about individual student performance.

Summative assessments have to have strong input by faculty but the questions, rubrics and marking can often be automated by AI. Automating marking is the single most effective way to free up time for teaching, research and other activities. This also includes the marking and feedback of essays.

We can also automate much more detailed feedback. As Dylan Wiliam has been saying for years, far too much assessment has no forward-looking pedagogy. It is seen by students as an end-point, when it should be feeding forward. AI can do this. Have experts in the loop by all means but look at ways to automate the bulk of the work. This as a field that is advancing rapidly, as AI capability progresses.

Step 6: Test Centres

The Opposite of Cheating by Tricia Bertrand Gallant presents a different set of perspectives. She and fellow author David Rettinger, flip the argument and start, not from the institutional, but student perspective.

She argues, based on flipping the debate towards students’ needs, that testing should be the responsibility of separate and shared ‘Test Centres’. These would provide assessment expertise and the ability to design and manage the delivery of assessments. This removes the pressure on faculty, who, on the whole, do not have the necessary expertise on assessment or its delivery. These centres would look at automating as much as possible, while being careful about verification and standards. It is clear that as AI improves, and it is at a blistering pace, so the automation of assessment and marking will become easier, better and cheaper.

Assessment is a rapidly evolving problem that needs this rapid and adaptive response. This is not a final solution as such, but a new approach to a growing problem that focuses expertise, while relieving the system of the unbearable pressure of producing and policing assessment. Teachers are not cops.

The question is whether this should be a single, clustered or national initiative. Huge savings would be possible if it were organised by the sector nationally. This is unlikely, as there is no real legal or political mechanisms for such a strategic approach. Tertiary education institutions are not known for their sharing, so even clusters are unlikely. That does not invalidate the strengt hof the idea.

These test centres could be physical but more sensibly virtual. This would allow testing at any time, on any subject. It is bizarre that one can only get tested on one day of the year, resits often not available for months, even a year later. Imagine of this were true of driving? A Test Centre should allow testing on demand. 

Conclusion

Oddly, the AI prwessure is forcing tertiary education to rethink, and reassess its own assessment. This is long overdue. It is acting as a catalyst for reshaping the role of the teacher and learner in relation to technology, accepting that AI is here to stay. The alternative is to boil like the proverbial frog, fail to respond with anything other than a policy document, constantly accusing students and/or institution for failing to properly assess learners. This is the road to ruin and regret, not the road to success.

Bibliography

Hartmann, C. (2025). Oral exams for a generative AI world: Managing concerns and logistics for undergraduate humanities instruction. College Teaching.

AI Confidential with Hannah Fry - hatchet job that lacks balance

AI Confidential with Hannah Fry is suspect. None of the three episodes makes any attempt at comparing upsides and downsides. Almost all of the time is spent, often laboriously, on the downsides, so they come across as hatchet jobs. For a documentary series on AI, there is a serious lack of data, studies and balance. That’s a shame.

Episode 1  AI girlfriends

First episode, on Jaswant, who broke into Windsor Castle (and was caught) was just weird. He had visited Amritsar, site of a massacre of Indians in 2018, became obsessed about colonial injustice, failed his exams, lonely during COVID, depressed, thought he was a Sith Lord and was seriously psychotic, so mentally ill that he was diagnosed as being in full blown psychosis when he climbed the wall and was eventually not charged with the crime but locked up in Broadmoor. 

Yet his Replica girlfriend bot is presented as the cause, a primitive pre-LLM ELIZA. The problem is a complete lack of balance and consideration of multiple causes. The police thought he would have done this anyway without the bot. They find this obsessive behaviour in many young terrorists. The mistake, common when fingering AI, is to see such things in terms of a single cause, when it is clearly muti-variant. And when things are multivariant, the new kid on the block, in this case AI, gets the blame.

I’d recommend you read ‘Love Machines’ by researcher James Muldoon, who takes a wider look at how bots are used by millions as companions, friends, girlfriends, lovers, mentors, therapists, advisors, coaches, and deathbots. There’s good evidence to show that far from manufacturing assassins, the technology helps with loneliness, avoiding embarrassment, as it is non-judgemental and affirmative. People feel heard, understood and supported, as the bots are calm and anonymous. They can break the silence and stave off the isolation. It is balanced in that it compares the upsides and downsides. Fry’s three programmes are all lopsided.

Episode 2 Self-driving cars

Rafaela Vasquez, the safety driver in Uber's self-driving Volvo, was distracted and looking down at her personal cell phone, even before she was on the road, streaming an episode of ‘The Voice’ on Hulu. In the moments leading up to the crash, that killed pedestrian Elaine Herzberg,  Dashcam footage showed her looking down for about 5.3 seconds immediately before the impact. and the National Transportation Safety Board determined she spent roughly 34% of the trip looking at her phone rather than monitoring the road. The police deemed the crash ‘entirely avoidable’ if she had been attentive. Vasquez was charged with negligent homicide in 2020 and pleaded guilty in 2023. She pleaded guilty and was sentenced to three years of supervised probation. Fry at this point blamed the car, when this was clearly a human error. You’d think she was innocent when listening to Fry. She may be a mathematician but she’s no journalist. 

Enter George McGee was found culpable in the civil lawsuit stemming from the April 2019 fatal crash in Key Largo, Florida. He admitted to police that he dropped his phone, while driving on ‘cruise’, and looked down to retrieve it while driving his Tesla on Autopilot. Note he was using Tesla's Autopilot at the time of the 2019 crash, not Full Self-Driving, which was not available until 2020.

This caused the vehicle to run a stop sign at around 62 mph and strike a parked SUV. This killed Naibel Leon and injured her boyfriend, Dillon Angulo. A jury in 2025 assigned him the majority of the blame 67%, with Tesla at 33%, and he settled privately with the plaintiffs prior to the trial against Tesla. It is true that Tesla paid the dead woman’s family a massive amount of money but this is a complicated case of a largely culpable driver and some AI.

In a California incident, they did not mention that the NTSB investigation revealed that Autopilot was engaged for nearly 19 minutes prior to the crash, and Huang's hands were not detected on the steering wheel for the six seconds before impact; data also showed he was playing a video game on his phone during the drive. 

In the Utah incident, Heather Lommatzsch's Tesla Model S, with Autopilot engaged, rear-ended a stationary fire truck while she was looking at her phone. She sustained a broken foot and the fire truck driver reported minor whiplash; no fatalities. Police determined she was culpable for distraction and over-reliance on Autopilot, issuing her a misdemeanour traffic citation for failure to keep a proper lookout. 

Overall, this was a hatchet job. No overall safety stats, which are positive but happy to feature footage of nutjob luddites in balaclavas. 

Episode 3 Healthcare

This was the strangest of all. As it was not AI in the dock but a US medical insurance company who were gouging customers. It’s a business model, not AI to blame. Sure they were denying people proper care but this is a feature of the American system, that maximises profit not care.

It was a bait and switch story, where the murder of the CEO by Luigi Amlioni, was warped into AI being the villain. Al was certainly being used but if this were spreadsheet with a formula, even humans making these decisons, the bottom line is greed not maths. Again, no balance.

I did a lot of research on AI in healthcare for my book AI and Productivity and have an entore chapter on the subject. I feel that Fry has focused too much on one case, in one business, in one country, doing one task, when there is evidence that AI is being beneficial across many areas in healthcare.

Conclusion

I’ve given up on the BBC Radio 4 stuff on AI, as it is truly awful, but expected more from Fry, as she’s a good presenter, smart and has the background to understand the technology. What she doesn’t have is the journalistic training and experience to see the big picture, so this series descends into rather long-winded hatchet jobs.

AI fits like a glove in EFL and second language teaching

“AI is your friend, not your enemy” was the title to my Keynote. That statement can sound provocative in a hall full of language teachers. We have all seen the headlines warning that translation tools are destroying motivation, that chatbots are replacing writing, that students are outsourcing their thinking. It is understandable to feel cautious. Yet when we look at the evidence and, more importantly, at what is actually happening in classrooms and with actual teachers and language learners, a more balanced and optimistic picture emerges.

There is a persistent fear that instant translation will remove the need to learn another language. If a phone can translate, why struggle with grammar, vocabulary and pronunciation? And yes, some learners report that translation tools reduce their motivation, but most do not. The majority remain motivated to learn languages even while using AI tools. Large international surveys of teachers show that human-led language learning remains irreplaceable, and scepticism that AI will replace teachers any time soon. Motivation has not collapsed. It has shifted. The environment has changed, and our role as teachers is not to resist that environment but to shape how students learn within it.

The reality is that students are already using AI, and many are using it daily. They practise speaking and receive feedback, summarise texts, brainstorm ideas, generate content, build flashcards, simulate exams and converse with chatbots. For many learners, AI has quietly become a personal assistant. Ignoring this will not make it disappear. Harnessing it, however, can transform our teaching.

One of AI’s most powerful contributions is psychological. Language learning is emotionally demanding. Students fear embarrassment. They worry about making mistakes. They hesitate to speak. AI provides a non-judgemental partner that is infinitely patient. It offers immediate help, avoids public correction, reduces anxiety and gives affirmative feedback. Learners can rehearse privately before speaking in class. For shy or anxious students, this is not a minor advantage; it can be transformative. Increased confidence often leads to increased engagement and engagement is the engine of progress.

The research evidence supports this. I showed the two major teta-studies examining AI tools in English language learning that report significant improvements in achievement. Reviews of AI-powered chatbots for speaking practice describe strong effects on oral proficiency, interaction and motivation. We are not discussing novelty tools; we are seeing measurable impact. When AI enhances interaction, it enhances language learning, because language learning is interaction.

Voice-based AI is particularly powerful for EFL. A learner can say, “Be my Spanish tutor at B1. Speak slowly, use everyday topics and stop often with questions,” or “I’m A2 French. Keep sentences short and correct me after I finish.” They can switch between languages, practise hotel check-in roleplays, request dictation, drill minimal pairs such as “ship” and “sheep,” practise connected speech like “Whaddaya wanna do?” or complete article drills choosing between “a,” “an,” “the” and zero article. The feedback is immediate and personalised. Each learner can operate at their own level, repeat as often as needed and progress at their own pace. This is differentiation without leaving anyone behind.

AI fits beautifully with Papert’s principle of high ceiling, low floor and wide walls interface. A beginner can have simple, structured conversations – low floor An advanced learner can debate ideas, analyse arguments or rehearse professional interviews – high ceiling The same tool accommodates a wide range of ability levels and creative directions – wide walls.

There is also a deeper pedagogical reason why AI works so naturally in language teaching. From Socrates onward, learning has been understood as dialogue. Socratic questioning draws out thinking. Bakhtin emphasised that meaning emerges through multiple voices in interaction. Vygotsky described the “knowledgeable other” who mediates learning within the learner’s Zone of Proximal Development. AI does not replace the teacher, but it can function as an additional knowledgeable other, available whenever the learner needs it. The teacher remains the orchestrator of learning.

Many teachers are already experimenting. They use AI to create lesson plans, generate materials, personalise exercises, design assessments and increase student engagement. At the same time, many feel underprepared. This is not a reason to retreat; it is a reason to invest in professional development. Teachers need permission to prompt, and students need permission to prompt wisely.

Assessment is another area where AI fits naturally. It can analyse spoken presentations, generate targeted feedback, turn transcripts into personalised error-based flashcards, create grammar drills based on actual student mistakes and help design video-based assessments. What once required hours of marking can now become rapid, formative feedback. Used thoughtfully, AI strengthens rather than weakens assessment for learning.

Perhaps most inspiring is the global dimension. AI systems now support hundreds of languages, including many that have historically lacked digital presence. Initiatives focused on low-resource languages aim to reduce language inequality in technology. Oral traditions can be digitised through mobile devices. Minority languages can gain visibility and vitality online. AI is not simply serving English; it is expanding linguistic possibility.

So where does that leave us as teachers? Precisely where we have always been: at the centre. AI cannot build human relationships. It cannot replace cultural nuance, empathy, humour or inspiration. It can amplify what we do. It gives learners rehearsal space, lowers anxiety, personalises practice, provides role-plays,  and multiplies exposure to language. It gives teachers creative leverage and new forms of feedback. It extends dialogue beyond classroom walls.

AI does not diminish language teaching. Used well, it strengthens it. It fits like a glove. The glove does not replace the hand. It enhances what the hand can do.

US to EU software: you can’t regulate people into better tech

The most obvious problem with a European social media platform and software migration, touted by the EU, is that a political stance is not enough to shift the dial. I’d also like to explore the other reasons for expressing doubt. Network effects don’t care about your values, which is why a ‘European alternative’ rarely become mainstream.

A political push will attract the Europhile fanboys and girls, so become a European BlueSky, a blue bubble of insufferable conformity to the view of almost everyone, part from those who are more libertarian, which is a vast number of even European citizens. Activism and a skewed political push are never wise when it comes to neutral platforms.

The EU’s tech problem isn’t politics,  it is product. We saw this with the European Mastodon and the Fediverse, both disastrous in terms of user experience and never mainstream, despite the efforts to push Mastadon as an alternative to X. BlueSky, is US owned, and has turned into an echo-chamber for those who don’t like X and Musk. These sprove how difficult it is to shift users from the incumbent platforms.

Quality matters

Quality has to be superior to effect a large migration, yet most of the options on the table are significantly inferior to their non-European comparators. There’s a good reason why many have a tiny fraction of users compared to US tech; they are more expensive, unreliable, have poor security and above all have much poorer functionality. 

Europe can’t migrate its way out of big tech

Consumer inertia also means that people stay with what they’ve got, as they are familiar with the platform. You can leave big tech, but your friends won’t. Then there’s the cost of change, losing your past data, friends and have to relearn a new platform and build again from scratch. It’s not often a monetary but time and social cost. But the biggest barrier is the network effect, where you are literally enmeshed in a vast network of friends , others and potential viewers. That’s hard to leave, like leaving home.

Facebook has faced endless claims of imminent death and decline. It has not happened. Network effects and the fact that we are creatures of habit, with a huge backlog of posts and memories on the platform, means we stay. We are not only creatures of habit, we are creatures of memory and identity. 

There’s a reason WhatsApp has not been replaces by Signal and Telegram. Millions said they’d leave but they didn’t. Platforms are sticky and when everyone else is on WhatsApp, you have to disentangle all of your existing social groups, and start again.

Organisational software

With organisational software, most European companies of any size use Teams. This will be fiendishly difficult to shift as they need to reduce risks on cybersecurity, outage and reliability. It is often a procurement as part of a wider Microsoft ‘house’ project. People complain about Outlook, Teams, Office, but the alternative, LibreOffice is worse. There’s also Google Workplace, which has oodles of great functionality. Google has a global revenue model, a massively integrated ecosystem, including its own operating system. It dominates search, video through YouTube and global maps with street view. There is no European entity that comes close to this. Both are solid as they have well-used file formats, are embedded into existing companies and their workflows, and people are familiar with their user interfaces and expected functionality. This inertia is hugely expensive to migrate to other platforms.

One way to push this European initiative is to mandate change in public institutions. But this may only result in even lower productivity in the public sector, compared to the private. That’s the last thing a stagnant economy needs.

Killer objection - AI

But the killer objection, for me, is the failure of Europe to embrace and utilise AI. This could have been a chance to break the US dominance but it is too late. Mistral is good but open source and now a bit player. The US and China are now unstoppable. More worrying is the failure of European companies to embrace AI within their own tech-stacks.

One exception

There is one exception to the rule, that’s Sweden’s Spotify. I’d have said SAP but their shareprice has dropped by a third in six months and they now look like a wounded, if not dying entity, under attack from those who can quickly code alternatives.

Good intentions, bad software

You can’t just legislate them out, there has to be choice, and as long as there is choice European tech companies are largely, but not entirely, screwed. Google, Meta, Microsoft, Amazon, Apple, OpenAI and Anthropic dominate in scale and global reach. Most of those major platform players are US firms but operate extensively in the EU and are regulated locally under laws like the Digital Markets Act, and include WhatsApp, Facebook, Instagram, LinkedIn, TikTok and YouTube.

Europe’s tech strategy is full of good intentions, but bad software. Most of the options are significantly inferior to what is available today. Using outdated technology will only set Europe even further back, through a clear drop in productivity, first in migration, then in poorer functionality and support.

AI Hierarchy of Expertise

There is a widening hierarchy of participants and drivers in AI. This is not unusual, as almost all technological revolutions show this type of fractional distillation.

1) AI creators (working on models & other source techniques)

2) AI experts (implementing structured projects and products)

3) Experts-in-the -loop (using tools like Claude code etc)

4) Humans-in-the loop (using chatbots)

5) Humans who don’t use AI

It is particularly acute in AI, as there are a huge conceptual and competence gaps between these different types of participants.

AI creators

Those who work in the large foundational model companies, such as OpenAI, Google, Anthropic, Meta and others are experts both in the fundamental mathematics and techniques. It is a mistake to focus entirely in the training of models here as the field has widened out into advanced techniques before and after training that are proving effective.

This is the most opaque group, as the surrounding system that makes those models useful, reliable and economical  viable in the real world are highly technical. Companies now rely on techniques to ground responses in external knowledge, tool use and function calling to let models act rather than just speak. They use agentic architectures to plan and execute multi-step tasks and long-term memory to maintain context and personalisation across sessions. There’s also serious evaluation and guardrails, synthetic and curated data pipelines, inference optimisation to cut costs and latency, even human-in-the-loop training. The centre of gravity has shifted to these more mature enhancing techniques because real value now depends on accuracy, trust, controllability, cost and integration into workflows. These problems that are solved at the system and product level rather than in the training run.

AI experts 

These are the companies and researchers who are building the next layer, on the basis of the fundamental technology. They may enable model orchestration, routing tasks across multiple specialised models as a service to customers. They are also creating vertical products in law, healthcare, training, education etc.

Many are involved in RAG and similar techniques to create useful projects and products across and within organisations. This is similar to the growth of internet companies in the early 2000s on top of the internet, now a massive portion of the world economy. I’ve been involved in this layer for over a decade. It’s pioneering, exciting and difficult but without this layer of experimentation the core technology would remain aloof and unusable. It takes commitment by CEOs and a real push to get through prototypes to product.

Experts-in-the-loop

As the technology matures a new group emerged which used more advanced techniques and tools, often the ‘pro’ models. This can be the sophisticated use of AI to handle and use data. But the dominant use has turned out to be coding and on a wider scale media production (text, graphics, images and video). My book on AI and Productivity’ explains why this emergent group is the key to solving the productivity problem and overcoming productivity paradoxes using AI.

Experts bring domain knowledge, this is usaally a mix of tacit and explicit knwoe3ldge and skiils that allow them to get the best out of the available tools. Coders make goof users fo Clade Code, graphic artists of image creation tools and will video creators in creating video. Experienced researchers will get the best out of the technology for research and so on.

Humans-in-the loop

By far the largest group are the ordinary users. One could split them into the two lower fremium groups of paid and free users but both sit on a rising tide of quality and functionality. They tend to single promtt and use the technology as a form of performance support, namely solving an immediate problem; finding something they don't knopw, communications issue, summarising, translating, rewriting and so on. They tend to prompt simply and as the technology improves get more and more bang for their buck.

Humans who don’t use AI

This last group are the disinterested or those who have no real need, which is fine. It also includes those who actively dislike the technology. They have all sorts of objections, some rational, others irrational and ill-informed. This is where the biggest, some say yawning, gap exists.

Least informed but loudest are this last group, at least on social media. They often invoke an eye-roll from the rest, not because they don’t have a point but because the claims are often unsubstantiated or just plain wrong. Typically they will refer to content from old models, rely on anecdote and make claims about the technology hitting brick walls.

Integration

There are people who find ways to integrate AI into organisational culture. They follow the hierarchy, as it depends on the degree of integration. Most organisational institutions focus on encouraging 'humans-in-the-loop' activity, with but this caps your productivity gains at mere text search. Capping effort at 'share prompts' and AI literacy courses, is to neither understand the capabilities of current AI, nor make serious organisational gains in productivity.

Conclusion

We should not be surprised by any of this. The fractional distillation into these sedimental layers of expertise is both a normal and necessary part of the evolution of a technology. In truth, as the technology matures and improves, this hierarchy tarts to massively expand in the middle layers, while the bottom layer shrinks. Think printing, cars, television internet. In this sense resistance is futile.

Ultimately the hierarchy turns in to fully realised rout and it becomes a normalised consumer product woven into the fabric of our lives. That will happen faster tyan most people think, within a few years. It is a hierarchy that automates, thereby blurring the lines, flattening it out and dropping the doubters.

At this stage of AIs development, only 3 years in, the last group is still the largest and therefore the loudest, that does not mean it is the most significant, merely the least informed and practiced.

Votes at 16 - no thanks

So Starmer, who is now working full time on digging his own grave, wants votes for 16-year-olds. This is a bad idea, as bad as teenage pregnancy. We also RAISED the age of marriage to 18, so why go backwards on this?

My first piece of evidence is myself. I was one. At 16 I has a Che Guevara poster on my wall, was a fake Marxist and had beliefs on politics, football teams even food that were absolutes... and absolutely ridiculous There were many hills I’d have gladly died on, metaphorically of course.

That’s the problem. We lack the tempering effect of experience, exposure to other groups of people, fiscal responsibilities, having kids. At 16 you’ve spent most of your time in one age group. Watching prank videos is not real life, neither are inordinate amounts of time devoted to influencers. Which is basically self-lobbying. Neither is taking pictures of yourself, adding filters or Tik Tok dances. Nor binging on Love Island or watching people eat enormous amounts of food. The teen brain is wired for novelty, peer validation and weird stuff online.

Teens are also curious, bored and boundary-pushing, so latch onto obscure causes and oddball stuff. That’s what being a teenager is all about. To be honest, that’s great, do it, enjoy it, just don’t vote for anyone on the basis of your fickle obsessions.

Also, given all we hear about mental health, do we really want teenagers to get heavily involved in what is essentially an endless tale of disasters and catastrophes? You have decades of this shit ahead of you, relax and enjoy your teenage years. Don’t destroy them with doom-laden tales of war, genocides and planetary disaster.

Let me tell you a secret, those kids that join a political party at 16, and you see giving those earnest speeches at political conferences. They are hated by both their peers and all sensible adults. There's nothing worse than a precocious teenager preaching to adults about how much they matter and if only we listened to them the world would be a utopian heaven. Is there anything more insufferable than those students who were active in the student Union? The tragedy of our age is that so many of these arseholes eventually get into the House of Commons.

Imagine thinking that politics is the solution to the rise in mental health and fall in wellbeing. Can you imagine a therapist saying “What you need is NOT more with your friends, having fun and feeling great, even ecstatic. What you need is to join the young Labour, Liberal, Conservative, Reform group and get out there arguing with people online.” There's a reason we don't allow 16 year olds to drink, they'd get wrecked too often. That's the problem, they're reckless.

Responsibilities and experience come later now, as paper rounds have disappeared and kids work far less than they did, as it is illegal. On top of that they all stay on at school until 18, many until their early 20s and, of course everyone needs a useless Masters degree to extend this period, and their debt, even further.

They will want to vote for open borders, free money, endless services, limitless holidays, everyone working from home, no cars and free travel. Until they have to pay for all of this as adults.

Also, across Europe, young people are now voting for the far right. If you think all of them are embryonic liberals in the making – think again. In Austria, Argentina and more recently Germany, the Netherlands and France, the young vote has drifted to the right. What you get is a drift towards both the far left and far right, as they tend to hold more extreme views, with certainty, before they hit the brick wall of real life. In Scotland it was used to boost the SNP vote, and there was a shift toward populism – extreme nationalism, a charismatic leader, authoritarian attitudes, control of the media and madcap policies like self-ID.

Let them enjoy those halcyon years. Get drunk, take drugs, be fanatical. We had two boys. They were insufferably disrespectful and wild. Despite the craziness, the police, ambulance, impromptu parties, they’ve turned into likeable adults. There is no need to make them even more insufferable at that age on politics.

Memo from CEO of Shopify.... Reflexive AI usage is now a baseline expectation at Shopify

 Team,

We are entering a time where more merchants and entrepreneurs could be created than any other in history. We often talk about bringing down the complexity curve to allow more people to choose this as a career. Each step along the entrepreneurial path is rife with decisions requiring skill, judgement and knowledge. Having AI alongside the journey and increasingly doing not just the consultation, but also doing the work for our merchants is a mindblowing step function change here.

Our task here at Shopify is to make our software unquestionably the best canvas on which to develop the best businesses of the future. We do this by keeping everyone cutting edge and bringing all the best tools to bear so our merchants can be more successful than they themselves used to imagine. For that we need to be absolutely ahead.

Reflexive AI usage is now a baseline expectation at Shopify

Maybe you are already there and find this memo puzzling. In that case you already use AI as a thought partner, deep researcher, critic, tutor, or pair programmer. I use it all the time, but even I feel I'm only scratching the surface. It’s the most rapid shift to how work is done that I’ve seen in my career and I’ve been pretty clear about my enthusiasm for it: you've heard me talk about AI in weekly videos, podcasts, town halls, and… Summit! Last summer I used agents to create my talk, and presented about that. I did this as a call to action and invitation for everyone to tinker with AI, to dispel any scepticism or confusion that this matters at all levels. Many of you took up the call, and all of us who did have been in absolute awe of the new capabilities and tools that AI can deliver to augment our skills, crafts, and fill in our gaps.

What we have learned so far is that using AI well is a skill that needs to be carefully learned by… using it a lot. It’s just too unlike everything else. The call to tinker with it was the right one, but it was too much of a suggestion. This is what I want to change here today. We also learned that, as opposed to most tools, AI acts as a multiplier.  We are all lucky to work with some amazing colleagues, the kind who contribute 10X of what was previously thought possible. It’s my favorite thing about this company. And what’s even more amazing is that, for the first time, we see the tools become 10X themselves. I’ve seen many of these people approach implausible tasks, ones we wouldn’t even have chosen to tackle before, with reflexive and brilliant usage of AI to get 100X the work done.

In my On Leadership memo years ago, I described Shopify as a red queen race based on the Alice in Wonderland story—you have to keep running just to stay still. In a company growing 20-40% year over year, you must improve by at least that every year just to re-qualify. This goes for me as well as everyone else.

This sounds daunting, but given the nature of the tools, this doesn’t even sound terribly ambitious to me anymore. It’s also exactly the kind of environment that our top performers tell us they want. Learning together, surrounded by people who also are on their own journey of personal growth and working on worthwhile, meaningful, and hard problems is precisely the environment Shopify was created to provide. This represents both an opportunity and a requirement, deeply connected to our core values of Be a Constant Learner and Thrive on Change. These aren't just aspirational phrases—they're fundamental expectations that come with being a part of this world-class team. This is what we founders wanted, and this is what we built.

What This Means

Using AI effectively is now a fundamental expectation of everyone at Shopify. It's a tool of all trades today, and will only grow in importance. Frankly, I don't think it's feasible to opt out of learning the skill of applying AI in your craft; you are welcome to try, but I want to be honest I cannot see this working out today, and definitely not tomorrow. Stagnation is almost certain, and stagnation is slow-motion failure. If you're not climbing, you're sliding.

AI must be part of your GSD Prototype phase. The prototype phase of any GSD project should be dominated by AI exploration. Prototypes are meant for learning and creating information. AI dramatically accelerates this process. You can learn to produce something that other team mates can look at, use, and reason about in a fraction of the time it used to take.

We will add AI usage questions to our performance and peer review questionnaire. Learning to use AI well is an unobvious skill. My sense is that a lot of people give up after writing a prompt and not getting the ideal thing back immediately. Learning to prompt and load context is important, and getting peers to provide feedback on how this is going will be valuable.

Learning is self directed, but share what you learned. You have access to as much of the cutting edge AI tools as possible.There is chat.shopify.io, which we had for years now. Developers have proxy, Copilot, Cursor, Claude code, all pre-tooled and ready to go. We’ll learn and adapt together as a team. We’ll be sharing Ws (and Ls!) with each other as we experiment with new AI capabilities, and we’ll dedicate time to AI integration in our monthly business reviews and product development cycles. Slack and Vault have lots of places where people share prompts that they developed, like #revenue-ai-use-cases and #ai-centaurs.

Before asking for more Headcount and resources, teams must demonstrate why they cannot get what they want done using AI. What would this area look like if autonomous AI agents were already part of the team? This question can lead to really fun discussions and projects.

Everyone means everyone. This applies to all of us—including me and the executive team.

The Path Forward

AI will totally change Shopify, our work, and the rest of our lives. We're all in on this! I couldn't think of a better place to be part of this truly unprecedented change than being here. You don't just get a front-row seat, but are surrounded by a whole company learning and pushing things forward together.

Our job is to figure out what entrepreneurship looks like in a world where AI is universally available. And I intend for us to do the best possible job of that, and to do that I need everyone’s help. I already laid out a lot of the AI projects in the themes this year- our roadmap is clear, and our product will better match our mission. What we need to succeed is our collective sum total skill and ambition at applying our craft, multiplied by AI, for the benefit of our merchants.

-tobi

CEO Shopify

Fiverr CEO''s stunning letter to all employees..... AI is coming for your jobs. Heck, it’s coming for my job too.

Hey team,

I’ve always believed in radical candor and despise those who sugar-coat reality to avoid stating the unpleasant truth. The very basis for radical candor is care. You care enough about your friends and colleagues to tell them the truth because you want them to be able to understand it, grow, and succeed.

So here is the unpleasant truth: AI is coming for your jobs. Heck, it’s coming for my job too. This is a wake-up call.

It does not matter if you are a programmer, designer, product manager, data scientist, lawyer, customer support rep, salesperson, or a finance person – AI is coming for you.

You must understand that what was once considered ‘easy tasks’ will no longer exist; what was considered ‘hard tasks’ will be the new easy, and what was considered ‘impossible tasks’ will be the new hard. If you do not become an exceptional talent at what you do, a master, you will face the need for a career change in a matter of months. I am not trying to scare you. I am not talking about your job at Fiverr. I am talking about your ability to stay in your profession in the industry.

Are we all doomed? Not all of us, but those who will not wake up and understand the new reality fast, are, unfortunately, doomed.

What can we do? First of all, take a moment and let this sink in. Drink a glass of water. Scream hard in front of the mirror if it helps you. Now relax. Panic hasn’t solved problems for anyone. Let’s talk about what would help you become an exceptional talent in your field:

Study, research, and master the latest AI solutions in your field. Try multiple solutions and figure out what gives you super-powers. By superpowers, I mean the ability to generate more outcomes per unit of time with better quality per delivery. Programmers: code (Cursor…). Customer support: tickets (Intercom, Fin, SentiSum…), Lawyers: contracts (Lexis+ AI, Legora…), etc.

Find the most knowledgeable people on our team who can help you become more familiar with the latest and greatest in AI.

Time is the most valuable asset we have — if you’re working like it’s 2024, you’re doing it wrong! You are expected and needed to do more, faster, and more efficiently now.

Become a prompt engineer. Google is dead. LLM and GenAI are the new basics, and if you’re not using them as experts, your value will decrease before you know what hit you.

Get involved in making the organization more efficient using AI tools and technologies. It does not make sense to hire more people before we learn how to do more with what we have.

Understand the company strategy well and contribute to helping it achieve its goals. Don’t wait to be invited to a meeting where we ask each participant for ideas – there will be no such meeting. Instead, pitch your ideas proactively.

Stop waiting for the world or your place of work to hand you opportunities to learn and grow—create those opportunities yourself. I vow to help anyone who wants to help themselves.

If you don’t like what I wrote; If you think I’m full of shit, or just an asshole who’s trying to scare you – be my guest and disregard this message. I love all of you and wish you nothing but good things, but I honestly don’t think that a promising professional future awaits you if you disregard reality.

If, on the other hand, you understand deep inside that I’m right and want all of us to be on the winning side of history, join me in a conversation about where we go from here as a company and as individual professionals. We have a magnificent company and a bright future ahead of us. We just need to wake up and understand that it won’t be pretty or easy. It will be hard and demanding, but damn well worth it.

This message is food for thought. I have asked Shelly to free up time on my calendar in the next few weeks so that those of you who wish to sit with me and discuss our future can do so. I look forward to seeing you.

Yours,

Micha

 

UDL is is an ideologicaly-driven sham

Universal Design for Learning (UDL) claims to be revolutionary but it is an ideologically-driven sham. It does far more harm than good. Folding in some of the worst learning theory into one 'my way or the highway' approach, it literally makes no sense. It is a barely disguised expansion of discredited Learning Styles theory, with costly and ineffective prescriptions.

They claim its framework promises inclusivity, adaptability, and improved access for all learners. Yet beneath this superficial rhetoric lies a troubling lack of theoretical grounding. It is more of an ideological banner than a well-developed pedagogical method.

At its core, UDL asserts that variability in learning is not an exception but the norm, and that curricula should be designed to offer multiple means of engagement, representation and expression. This sounds progressive, but it begs the question: where is the evidence that this approach is universally effective? Unlike more established pedagogical frameworks that are underpinned by decades of empirical research, UDL lacks the scaffolding of a cohesive theoretical tradition. It cherry-picks ideas from neuroscience, educational psychology and accessibility studies without unifying them into a coherent explanatory model. Consequently, it has become a collection of loosely connected guidelines.

UDL is undeniably ideological in nature, that is its appeal. Rooted in a utopian vision of educational equality, it is executed so simplistically that it does great harm. It devastates actual individual needs by stupidly replicating things across many modalities. By focusing so heavily on inclusion, UDL ignores the realities of constrained resources, teacher capacity, and the complex priorities of educational institutions. It is a simplistic bromide that puts replication over meeting real needs in real organisations. It literally abandons pedagogy on the altar of ideology.

The ideological bent of UDL sidelines alternative approaches that need specialisation or targeted interventions. For example, tailored support for students with specific disabilities often requires highly individualised methods that go beyond UDL's dilute and generic calls for flexibility. In an obsessive attempt to be ‘universal’, it dilutes its applicability to those who really do need personal, specialised attention the most. Ironically, this universalism can neglects the very individuals it aims to serve.

UDL’s slavish emphasis on learner choice and agency becomes an ideological trap in practice. The belief that all students should have multiple ways to engage with content and demonstrate understanding assumes they have a level of maturity, self-awareness, and autonomy that many learners, particularly younger ones. This is quite simply mad. 

An overemphasis on learner freedom leaves less confident or less experienced students adrift, struggling to navigate an educational landscape designed to cater to everyone and, paradoxically, to no one in particular.

Universal Design for Learning aspires to reshape education into a more inclusive, flexible, and equitable system. But its lack of theoretical rigor and reliance on ideological principles leave it vulnerable to critique, means it should be dismissed as an impractical, overhyped framework. It has little to do with the real-world complexities of education, which is why it only thrives on the fringes of Higher Education.

One could well rename it a Universal Disaster for Education.

Raising in Arizona with AI: groundbreaking journey in AI ducation

Arizona has taken a bold leap in education by approving a pioneering model: a charter school where artificial intelligence takes the lead in teaching. Set to launch next year, Unbound Academy will be a fully online institution catering to students in grades four through eight. 

Traditional teaching will be replaced with two hours of daily academic instruction tailored entirely by AI. Leveraging platforms like IXL and Khan Academy, the AI will adapt lessons to each student's pace and abilities, analysing emotional cues, task completion times, and responses to create a personalised, frustration-free learning experience designed to maximise engagement and progress. That's the claim at least. But there have been many attempts at online schooling.

For the rest of the school day, students will take life-skills workshops on topics like financial literacy, entrepreneurship, and critical thinking, guided not by teachers but by facilitators known as 'guides'. This innovative approach stems from Unbound’s success with its private school in Texas, where it claims to have doubled academic outcomes with fewer instructional hours. Inspired by Elon Musk, Unbound aims to refine this model for potential expansion into states such as Arkansas and Utah.

They see this as a much-needed shift to address inefficiencies in traditional education while preparing students for an evolving world. Backed by influential supporters like the Bill and Melinda Gates Foundation, Unbound’s AI-driven school could mark the dawn of a transformative era in education—or spark ongoing debate over its impact on accessibility and quality. Let's see...

Why Latin should NOT be in the school curriculum

The venerable Mary Beard invited me to debate the ‘Future of Latin’ at the British Museum some time ago and every few years, normally in line with the electoral cycle, the issue arises again. So how did it go and what were my arguments?

Sell-out

A sell out, with 350 paying Latinists, who turned up to hear Mary chair a debate which pitched David Aaronovitch and I against Peter Jones and Natalie Haynes. As we walked on stage and introduced ourselves (I found that I was the only person who didn’t go to Oxford), everybody seemed to know each other (except me). This is exactly the sort of challenge I like, as although it was a sell-out, I’m not in the habit of selling-out on my beliefs and principles.

Neither contrarian nor philistine

To be clear, I was not there as a contrarian or philistine, as I’ve been in love with the classics since I was a boy. My first secondary school lay astride the Antonine Wall and for over 40 year I’ve been to every corner of the Roman, Greek and Egyptian Empires, from Scotland to Syria. I cycled Hadrian’s wall, still go to Greece every year and never miss an opportunity to visit sites, especially on the Peloponnese. and I go to Egypt, as I do almost every year, for another dose of Egyptology in December, and have done since 1989. 

I am, however, also a rationalist and realist, and my 30 plus years of experience in the learning game have made me deeply suspicious of the position of Latin, among many other things, in our culture and school curriculum. 

As Bertrand Russell said:

“I was made to learn Latin and Greek, but I resented it, being of opinion that it was silly to learn a language that was no longer spoken. I believe that all the little good I got from years of classical studies I could have got in adult life in a month.”

I was there to argue that it should NOT be taught in schools at all. These were, and are, my arguments. Note that I am not against Latin the language as an object of study, I am against it being taught in schools. The arguments for each of these propositions were presented by my opponents and audience members. These were my responses.

1. LATIN does NOT help you learn other languages

Why scratch your ear by going over the top of your head? Isn’t it obviously easier to just get on and learn Spanish, Italian and French, rather than the convoluted route through Latin. Researchers Thorndike, Sherwin, Haag & Stern all think so. In the Sherwin meta-study 'Research and the teaching of English, “the study of Latin does not necessarily increase the ability to learn another language… No consistent experimental evidence in support of this contention was found.” Learners have limited time, that time is clearly better spent on the target language itself. 

In fact, Latin can make learning a new language MORE difficult. In Search of the Benefits of Latin by Haag and Stern (2003), who followed up on Thorndike’s work nearly a century earlier, in the Journal of Educational Psychology is the key paper. They took two groups of German students, one who studied French, the other Latin as their second language. Both groups were given a course in Spanish and the results measured. When the results were analysed by a Spanish assessor (who didn’t know who had taken French or Latin) the French students made significantly fewer grammatical errors than the Latin students. As predicted the Latin students wrongly transferred the rules of Latin to Spanish. For example “misconstructions in verbs emerged to be either highly reminiscent of or identical to Latin verbs”. The French group turned out to be much better prepared to cope with Spanish grammar.

Psychologically the Latin students had suffered from negative transfer using false friends in their new language. The problem with understanding Latin is that you need to pay close attention to word endings; case markers on nouns and time markers on verbs. But in English and Romance languages word order and prepositions are more important. Endings play a minor role. The fact that the grammatical similarities between modern Romance languages are much greater than that between Latin and modern Romance languages, means that the defenders of Latin are flogging a dead horse. Thorndike was right – transfer of the wrong kind occurs.

2. LATIN does NOT have an edge in improving cognitive skills

This argument is greatly loved by the parents of ‘gifted children’ although I rarely come across a middle-class parent whose child is not gifted. For gifted, read ‘pushed’ (not a bad thing but very different). Again Haag and Stern (2000), in a review of the literature found that Thorndike, “did not find any differences in science and maths in students who learned Latin at school and those who did not”. Two groups of comparable students, where one studied Latin, the other English, were assessed after two years, “No differences were found in either verbal or non-verbal IQ or grades in German or Maths”. This again had been predicted by Thorndike decades before, namely that transfer needs common ground in the source and target.

3. LATIN does NOT give significant advantages in using English

English is a Germanic language – we are largely speaking in Old English rooted language. The TOP 100 words are Old English (sorry three are Old Norse – THEY, THEIR, THEM). As for one audience member’s argument that it is necessary as all children need to be able to understand etymology, I disagree. Is there anything more annoying than the dinner-party bore who stops you and explains the root of a word, as if it made any difference to your argument or its contemporary meaning. If anything a good course in Old English would be better. Do you really have to go through years of Latin so that you understand the roots of Homo sapiens?

4. LATIN does NOT guides us correct use of English grammar

Stephen Pinker, Harvard’s world renowned expert in psycholinguistics backs this up in The Language Instinct, “Latin declensional paradigms are not the best way to convey the inherent beauty of grammar”. He recommends computer programming and universal grammar on the grounds that they are “about living minds and not dead tongues”.

Pinker has a go at the Latin language mavens who want to pointlessly foist Latinate rules of grammar into English. As Pinker explains, this snobbery took root in 18th century London, when Latin was used as a mark of social class (still true today) and Latin grammar rules were crudely pasted into books on English grammar, for example, ‘don’t split infinitives’ and ‘don’t end a sentence with a preposition’. Latin simply doesn’t allow you to split an infinitive and to stupidly insist that it’s wrong in English, is as stupid as making us all wear togas.

English does NOT have Latin grammar. English grammar fundamentally Germanic. Latin highly inflected, role of word in a sentence indicated by its endings. English relies primarily on word order to determine grammatical relationships. Latin nouns gendered English are not.

5. LATIN is necessary for SCIENCE, LAW and MEDICINE

One girl in the audience, from Merseyside was adamant that Lawyers needed to have studied Latin. Incidentally, if you’ve heard the argument that Latin helps medical students learn and understand the considerable amount of medical vocabulary that has to be learned in medical schools. This also turns out to be false as shown in Pampush and Petto (2010). The Latin vocabulary in law, medicine and biology is there but one can simply learn the words, without knowing their etymology - most lawyers, doctors and scientists do exactly that. Why spend years learning Latin to help you in a small selection of vocabulary? The pay-off is, as Russell pointed out, is tiny.

6. LATIN brings the joy of ideas and literature

One contribution from the audience I did like was the idea that Latin bridges us to rich tradition of thought and literature. He mentioned Roman literature but also the Magna Carta, Bacon and Newton. First, anyone studying history will not lose out by working with translations of the Magna Carta or Bacon. And does anyone really need to read Newton’s tortuous Latin, other than scholars in the history of maths? I think not. 

And why not Greek? Wouldn’t you prefer the riches of Plato and Aristotle, Homer’s Iliad and Odyssey, along with the works of any one of Aeschylus, Sophocles, Euripides and Aristophanes to the largely derivative Roman philosophers (few can name any) and dramatists. Even in History, the Greeks Thucydides and Herodotus are a match for any Roman writer, including Tacitus. In politics, our democratic traditions are largely Greek. Even English law is not Roman (although in Scotland it is). Then there’s the politics and democratic traditions that are fundamentally Greek. Even in maths and science the mighty Pythagoras, Euclid and Archimedes trump the Romans.

7. LATIN’s history of exclusion

Of course, Latin was introduced to this country as the language of the clergy, it was not, as is sometimes supposed introduced by the Romans, as few Latinate words come from that era. As the language of the church it largely excluded the laity, as most remained illiterate and spoke various forms of English. It was then used as the gatekeeper for learning. This had some benefits, when Latin was sort of European Esperanto, but continued for centuries after that died and was long used as the gatekeeper at Oxbridge and other institutions. David Aaronovitch made the telling point that it is still a key ‘marketing’ differentiator for independent schools.

8. LATIN would NOT die if not taught in schools

There are plenty of scholars in subjects that are not taught at schools. The bottom line with any dead language, especially Latin, is that there’s little that is new and to be uncovered. Compare this to the vast amounts of Sumerian cuneiform tablets that still need to be both deciphered and excavated. In the end I agree with one of our greatest living Latin scholars Mary Beard, “the overall strength of the classics is not to be measured by exactly how many young people know Latin or Greek from school or University. It is better measured by asking how many believe that there should be people in the world who do know Latin and Greek.” This about sums up my position. 

I am not against the study of Latin or any other historic languages. This is largely a matter of proportionality for our Universities. By all means let a few study Latin. What I am against is too prominent a role for Latin in contemporary school curricula. Our young people have enough on their plate at 5-18, as the range of subjects expands to include a wider range of science subjects, IT and other vocational skills. A dead language, in the sense of no longer being spoken or used for scholarship, at this stage is merely the dead hand of educational history being played out by interested parties.

9. LATIN is NOT about choice

Several people argued that Latin is a matter of choice in schools. The curriculum is crowded enough with increasing demand every year as new subjects, such as computer studies, emerge. As we have seen from recent educational lunacy in policy, has resulted in the destruction of choice in the curriculum. Professor Alison Wolf, Gove’s lapdog, ignored advice from industry and education experts to crush 3100 vocational qualifications. That strange beast, the EBacc, which had Latin as a core choice, deliberately excluded all vocational subjects, even ICT. Gove's successors , such as Williamson did the same, with special temporary funding, creating a one-sided system that simply reinforces the old apartheid system we have in this country between academic and vocational learning. 

This is not to say that all education should be aimed at utility and employment, just that a contemporary curriculum in schools is always a trade-off, with new subjects and content always arriving, so that everything we taught in the past can’t just remain. We need to weed and feed a curriculum, which means some hard choices. My preference would be to focus more on modern languages themselves, where we have declining numbers.

10. LATIN is a political issue

It has now become a political pawn. A largely (not exclusively) right wing movement to impose private school norms (7%) on state schools (93%) became an electoral issue, hence Williamson's special funding.  Labour won and implemented their promised VAT policy on private schools and in reassessing budgets they also scrapped Williamson's artificial subsidy for Latin in state schools. We can argue for or against Latin in schools but this was democracy in action.

Conclusion

I rather liked this audience and I especially liked Peter Jones.  He was my opponent but put to bed those old tropes about Latin improving your ability to learn languages, improve intelligence or think logically. He was remarkably free from cant and any sense of snobbery. What he loved was Latin and his plea was for the beauty and intrinsic value of the language. With that I have no argument.

AGI is here...

As easy as one-two-three? No. Easy as one-three. OpenAI has skipped o2 to call their new ‘frontier’ model o3 (Altman hinted that this may have been due to Telco o2 issue). I like the idea that AI is undermining everything we were told makes good marketing - no read branding, messaging, just get it out there with benchmarks.

Social media was full of people saying Google had destroyed OpenAI. But their pre-Christmas release single was smashed off the Charts by OpenAI’s o3.

Transcending human intelligence

Not sure we've realised that we now have AGI. Arc-AGI solved with 87.7% (human threshold 85%) as well as other benchmarks... we will, of course, move the bar higher, find ways to avoid recognising the achievement... then pass it again... Arc Prize is a not-for-profit with an AGI benchmark. There are new skills so that the AI cannot memorise them.  

To give you some idea of the speed and scale of progress:

Software engineering

On real-world software tasks, evaluations on o3 are more than 20% better than o1 models at 71%. AI has already established itself as solid, useful and widely used coding tool. People have been letting AI code then iterating  iterate. This takes it to another level. One wonders how long the job of coder will survive at this rate.

Maths and science

Also superb at maths and PhD science questions at 88.7 %. A typical PhD gets around 70%. Frontier is toughest mathematical dataset – extremely hard problems. o3 has 25% accuracy which is a strong result.

GPQA Diamond

An interesting measure of how far we have come is GPQA Diamond. It's a clever test that compares a model against npvice Google search, human domain experts and the model. Experts get 81% right in their fields, highly skilled non-experts with 30 minutes per question and Google access get 22%. GPT-4 got 37% at the start of 2024. o1 got 78%. o3 is 87.7%. That is astonishing.

AGI

Artificial General Intelligence suffers from a problem of definition, as do most abstractions at this level. Does it mean:

Specific human reasoning skills (maths, science etc)

Average human competences

Greater than all of humanity

There is no singularity, there is a spectrum or constellation of possible targets here. What is clear is that these targets are being hit, not in one go, but one by one, sometimes in clusters. Maths and science are easy to measure but also fiendishly difficult to achieve, so this is a real milestone. Yet they focus on clear rationality. To be fair critics were telling us that AI would never get here, never mind get here so fast. We should celebrate this as many of the problems we face with climate, energy, healthcare and education may well be solved, not in the sense of final solutions but better solutions.

Problem solving in real life is messier and more of a challenge. That's why the agentic move is so interesting, as it tackles this set of human capabilities. We have brains that have evolved into a specific environment where we had to solve specific problems. This is where the dynmaic interrogative, dialogic nature of AI helps enormously. It has already made great strides in this direction.

Embodied AI, in the physical world of elevators, cars, cabs, trucks, drones, ships and submersibles has also taken great leaps. This is another set of targets that are being hit one by one. One could argue that neurological targets are also on our hotlist - Neuralink is a good example, where we enhance our neurological deficits.

AI may help us understand neuroscience and the brain, solve engineering problems to accelerate fusion, help with drug discovery and many other intractable problems. What we can be sure of is the increased impact of AI on productivity. The leaps in efficacy prove this.

Conclusion

This is a warning to people who claim that scaling is over. Sutskever was right – we have a way to go and other techniques are clearly delivering the goods. It is clear that AI is delivering faster than expected. The consequences of AGI are closer than expected with huge productivity gains on the horizon. That is the subject of the book I am currently writing.

Future issues

One issue needs discussion - compute costs. I think this will be solved. We saw a 250x decline in token costs in 20 months. So $20 for hard problems are likely to come down to cents.

One can now start to ask how the cost of compute compares to human costs in organisations for similar tasks and roles. The productivity game looks as though it will start with coding, where much of it can be automated.

On problems to solve:

• fusion

• medical research

• personalised tutors

• optimising political policies

• next-generation batteries

• cheap renewables

A final though on AI being self-generative. At what point does this technology start working on itself to solve the frontier problems and advance even quicker?