Average Is the Ceiling for AI. It’s the Floor for a Real Teacher.

Every student, client and workshop participant I meet these days carries a small, tireless genius in their pocket.

Ask it almost anything in your field, and within a few seconds it returns a clean, confident, well-formatted step-by-step answer.
It never admits uncertainty. It never mentions the time that same tactic later collapsed in a real campaign.

That second part, the part about the collapse, is now my responsibility. To be a teacher in the age of AI means standing next to something that never tires, never feels awkward, and never has to sit with a client and explain why the numbers went the wrong way.

I have been doing this work for thirty years, in university lecture halls, technical schools and workshop rooms across four continents. A large share of my old job involved showing people “how you do this.”

That portion has largely disappeared. No one needs me to walk them through setting up a segmentation rule or drafting a cold outreach sequence. Their pocket genius will produce it before I finish the diagram on the board.

So what remains for the human standing at the front of the room?
Quite a bit. Just not the part most people still expect.

Congratulations, You’ve Got a Genius Intern Who’s Never Been Fired

In March 2025, Harvard Business Review published “Strategy in an Era of Abundant Expertise”, co-written with three Microsoft executives and Harvard’s Karim Lakhani. Their core observation is straightforward. A company used to be a collection of hard-won, costly expertise. That expertise is now cheap, fast and available inside a chat window. They are right, and the reframe is useful. Knowledge that once lived only inside a consultant’s head or a professor’s decades of notes is now open to anyone with an account.

What the reframe often leaves out is this: cheap and instant does not equal correct. It equals popular. Those two words describe very different things, and confusing them is where most of the current conversation goes off track.

The Problem Is What The Genius Actually Learned From

Ask almost any large language model about email marketing ROI, and there is a good chance it will give you the number that has circulated in articles for years: forty-two euros returned for every euro spent.

It sounds precise. It sounds authoritative.
Try to follow the citation trail, and you end up in a loop of references that never land on a real dataset.

I have refused to use that figure in anything I publish for quite some time because that average means very little to your use case.
I have seen too many budgets built on a number no one can actually verify, and that alone is reason enough.

That statistic is a perfect example. It is not accurate. It is widely repeated. A model trained to predict the next most likely word from what appears most often on the internet will deliver it with complete confidence, because from the model’s point of view “most repeated” and “most true” look the same. They are not the same, and recognising the difference has become roughly half of my real work.

Microsoft Research and Carnegie Mellon examined this directly. In a 2025 survey of 319 knowledge workers covering 936 real AI-assisted tasks, they found that higher confidence in the AI’s output was linked to less critical thinking, while higher confidence in one’s own knowledge was linked to more of it.

In plain terms: the more you trust the machine, the less you question its answers. The more you actually know the field, the more errors you catch. That finding points clearly to where the value of a real expert now sits.

My Actual Job: Curator of Scar Tissue

This is how I now understand my role when I teach at universities or run workshops from Lisbon to Bangkok or Barcelona. I am the curator. Facts themselves are free, so collecting and repeating facts stopped being the core job years ago. Selecting what is actually worth doing, filtered through thirty years of watching tactics succeed, fail and sometimes explode in a client meeting, is the job that remains.

There is a name for the kind of knowledge that comes from doing something badly and then working out why. Philosopher Michael Polanyi called it tacit knowledge: the understanding you possess but cannot fully write down, the kind that only appears when you are in the middle of the situation.

A model trained on text has read every manual about riding a bicycle. It has never fallen off one. My scars are the falling-off part, and they are the element a model cannot generate, because it has never lived through a single live deployment, budget cut, algorithm shift or angry client conversation.

I learned the expensive way that rapid growth nearly shut down two of my agencies.
The growth itself was not the problem. What failed was everything underneath that wasn’t built to carry that load.

I can describe the lesson in a clean paragraph, and an AI can rephrase that paragraph elegantly. What it cannot do is tell you, with the specific weight of having lived it, which early warning signs to watch in your own business before the same cracks appear. That is not information. That is judgement earned the hard way, and it is exactly what a curator supplies that a search result never will.

Turns Out, Even The Best AI Tutor Needed A Human To Design It

This is my favourite piece of evidence in the entire discussion, because it demonstrates the opposite of what it first appears to say.

Harvard ran a proper randomised controlled trial in 2025, published in Scientific Reports, comparing a custom AI tutor with Harvard’s own highly regarded active-learning physics classes.

The result drew attention: students using the AI tutor learned more than twice as much, in less time, than those in the classroom.
At first glance, that looks like the end of the case for human teachers. Read a little further, and the opposite conclusion emerges. 😉

The tutor, called PS2 Pal, was not an unmodified general model. Physics lecturer Greg Kestin and his colleagues supplied the correct answers in advance so it would not invent them. They required it to reveal only one step at a time and never the full solution. They limited every response to a few sentences because, left alone, the system tends to produce long answers. According to the Harvard Gazette’s own reporting, Kestin argues the tutor succeeded where others failed precisely because of the careful guardrails a human expert placed around it. See it now? 🤔

In short, the AI tutor that outperformed the classroom only worked because a curator designed the boundaries. Remove those boundaries and the outcome changes.

A 2025 study in PNAS involving nearly a thousand Turkish high-school students found that unrestricted ChatGPT access left students scoring 17% lower on their exam once the tool was removed, compared with students who never used it. They had treated it as a crutch rather than a tutor. The paper’s title captures the point: generative AI can harm learning. The same technology produced opposite results, and in both cases the decisive factor was the human who set the conditions.

Genius, But Only If You’re Already Behind

A consistent pattern appears across rigorous studies of AI at work, and it matches almost exactly what I observe in my own workshop rooms.

Harvard Business School and Boston Consulting Group ran a major field experiment with 758 consultants, giving them GPT-4 for actual client assignments. Inside the model’s area of strength, consultants completed 12.2% more tasks, 25.1% faster, with quality rated more than 30% higher. The researchers then gave them a complex strategic task deliberately placed outside that comfort zone. AI assistance reduced correctness on that task by 19 percentage points. Same tool, same people, opposite effect, depending entirely on whether the work sat inside or outside what the researchers call the jagged frontier.

Stanford and MIT researchers observed a parallel pattern among more than five thousand customer-support agents using an AI copilot. Newer agents improved by roughly a third. The most experienced agents showed almost no gain. The researchers’ explanation is the one that fully convinced me of the curator framing.

In effect the AI had learned to imitate the average competent agent. The already excellent agents had little new to teach the model, because the model had already absorbed the average of everyone else’s best practices, and the average is, by definition, not where the strongest people already operate.

Pause on that finding, because it contains the entire argument. A generic AI tool’s upper limit is roughly the average of what has already been published about a field. If you are already above average, which is usually why someone is paying you to teach or advise, generic AI guidance can only pull you toward the middle. It cannot move you further ahead. Only a human who is already ahead can do that, and that is the position I now occupy. 😉

What Happens When Nobody Curates

None of this is abstract speculation. There is a concrete cost to skipping the curator role, and researchers have begun to call it cognitive debt.

MIT’s Media Lab attached EEG sensors to 54 participants while they wrote essays, some without help, some with a search engine, and some with a large language model. The LLM group displayed the weakest brain connectivity of the three and had difficulty recalling or quoting their own writing shortly afterward. The study is small and still a preprint, so treat the precise numbers with caution, but the direction matches everything else on this list.

Anthropic ran a cleaner version of the same idea with working developers in early 2026. Fifty-two engineers learned an unfamiliar Python library, half with AI support and half without. Everyone then took a quiz with the AI turned off. The AI-assisted group averaged 50%, compared with 67% for the group that worked through the material themselves, a gap of nearly two letter grades.

The largest deficit appeared in debugging, the very skill most needed when the AI is wrong. One encouraging detail in the same study: developers who used AI mainly to ask conceptual questions rather than to generate and paste complete answers performed as well as the group that had no AI. The tool itself was not the problem. Using it to avoid thinking was.

That is the curator’s real function reduced to a single idea. Not “keep AI away from learners.” Design the boundaries that turn a step-by-step answer machine into a real thinking partner, and know when to require a student to struggle for a while before the answer appears.

What This Looks Like In My Classroom Now

I am not the only one adapting. The University of Auckland has spent the last two years rebuilding assessment around structured interactive oral exams, conversations in which a student must defend what they actually understand about their own submitted work.

A live conversation cannot be outsourced to a chatbot open in another tab. It is a direct solution, and similar approaches are spreading through universities worldwide for the same reason my own practice is changing.

Here is roughly the shape of what I now do in my rooms, from Universidade Lusófona to Harbour.Space in Bangkok and Barcelona, to name a few.

Where the pocket genius wins	Where the curator still wins
Generating a first-draft step-by-step guide in seconds	Telling you which three steps in that guide will quietly damage your open rate
Summarising every public opinion on a tactic	Telling you which of those opinions I have personally watched fail, and why
Producing a competent, average answer instantly	Pushing you past average, because average is exactly what the model already knows
Never getting tired of your questions	Noticing when you are asking the wrong question entirely

I now spend far less time explaining how to build a segmentation rule and far more time examining live case studies of campaigns I have actually watched collapse.

I assess less by asking students to produce a document, since a document is trivial to generate, and more by asking them to defend a decision out loud, in real time, in front of me. I teach the “why” behind a tactic more than the “how,” because the how is now free, instant and sometimes confidently wrong.

None of this means I have turned against AI in the classroom. I use it constantly, and I teach students to use it constantly as well. What I have stopped doing is pretending that my value was ever mainly the how-to. That was always the easier 80% of the work. AI simply made that fact visible.

The Bit You Don’t Want To Hear

Here is the part of the argument that might make you uncomfortable, and I prefer to state it rather than avoid it.

If your entire teaching practice consists of a polished delivery of steps that are already public, the pocket genius is coming for that work, and it is already faster and cheaper than you. The World Economic Forum’s Future of Jobs Report 2025 quantified the scale of the shift: 39% of the core skills employers value are expected to change by 2030. That is a large portion of most current curricula.

Stanford’s Digital Economy Lab reported something that should concern every trainer whose business depends on new graduates entering the workforce. Employment for 22-to-25-year-olds in the most AI-exposed roles has lagged well behind less-exposed peers since 2024, driven mainly by companies hiring fewer juniors rather than by large-scale firing.

Fewer junior hires means a thinner pipeline of people learning a craft the traditional way, on the job, under experienced supervision. If the apprenticeship model shrinks, the value of anyone still able and willing to pass on real, tested, scarred experience only increases.

So teaching is not disappearing. It is being reduced to the parts that were always the real point, while the parts that were mainly information delivery are being automated, correctly, out of existence. I am glad to see those parts go. I never particularly enjoyed serving as a human printout.

Curate Or Get Curated Out

The pocket genius will continue to improve at producing step-by-step guides. It will never have sat in a boardroom explaining to an angry client why the campaign missed its target, and it will never carry the particular dread of watching a launch fail in real time. That dread, strangely enough, is now among the most valuable things I have to offer.

Your students, workshop participants and clients no longer need you simply to know things. Search engines already knew things. They need you to have been wrong about something that mattered, to have survived it, and to have understood exactly why.

That is not a line on a résumé. That is the entire curriculum now.

This article was researched and drafted with AI assistance but thoroughly fact-checked and rewritten by Rui Nunes before publication. Yes, I really wrote it the best I can. Does it show? 😅