The Manifesto

Future Learning Manifesto

A point of view on learning in the age of AI

AI is about to give every L&D team the ability to produce more content than their people could ever consume. That is not the same thing as giving them the ability to change anything.

The Problem We Brought With Us

For decades, the learning industry has had a favourite answer to the question 'is this working?' That answer is the completion rate. Seventy-eight percent finished the module. A hundred percent of the team has been through the course. Job done.

Here is what a completion rate has never told anyone: whether a single person did anything differently the following Thursday. That is the only question that matters. We have been measuring the wrong thing, calling it data, and building more of the same on top of it.

AI does not fix this. It accelerates it. When producing content costs nothing, the pressure to produce more of it is irresistible- and if the measurement infrastructure isn't there to tell you whether it's working, you end up with an enormous, expensive, beautifully designed pile of content that isn't changing anything. Faster than before.

This is not a technology problem. It is a clarity problem. Until learning functions are honest about what they are actually trying to change- and ruthless about measuring whether they do- AI is just a more efficient way of producing the wrong thing.

THREE THINGS WORTH ARGUING ABOUT

Great learning is not information transfer. It is behaviour change.

This sounds obvious until you look at how most learning content is designed, commissioned, and measured- at which point it becomes clear that the industry has known this for thirty years and largely ignored it. Bjork's desirable difficulties research, Ebbinghaus's forgetting curve, self-determination theory: all of it points the same direction. The conditions that feel like learning are frequently the conditions least likely to produce it. Smooth consumption. Easy recall. The warm glow of familiarity. These look excellent in a completion dashboard and leave almost no trace in behaviour six weeks later.

Which brings us to why behaviour change is harder to produce than information transfer, and why so much well-intentioned learning fails to cross that line. Fredrickson's broaden-and-build theory shows that positive emotional states literally expand cognitive capacity- a learner who is genuinely engaged processes information differently, makes more connections, and retains more of what they encounter. A learner who is anxious, bored, or simply going through the motions does the opposite. The anterior cingulate cortex doesn't distinguish between social exclusion and being the only person in a training room who doesn't understand what's being taught. Both register as threat. Under threat, the brain protects itself. It does not change. This is why designing for behaviour change and designing for engagement are not two separate considerations- they are the same consideration. You cannot produce the first without attending seriously to the second.

1

One of the more convenient myths in L&D is that quality is a matter of taste- that "great learning" is too contextual, too human, too slippery to define with any precision. This is, to put it plainly, an excuse for not doing the definitional work. Great learning has observable characteristics: it earns attention before it asks for effort; it creates genuine understanding rather than the performance of it; it changes what people do, not just what they know; and it makes people want to keep going rather than relieved to be done.

These characteristics can be defined precisely enough that a team can hit the standard without you in the room. They can be used to evaluate AI-generated content with the same rigour as human-written content. The moment "great" is defined clearly, it becomes possible to defend it at scale-and to notice, quickly and honestly, when it isn't being met. That's not a constraint on creativity. It's what makes creativity trustworthy.

Educational quality is not subjective- and "I'll know it when I see it" is not a quality standard.

2

The most important design decision is what AI touches- and what it doesn't.

AI is genuinely excellent at production: drafting, iterating, translating, scaling, localising. What it cannot do is the diagnostic work that has to happen before a single brief is written- understanding what actually needs to change, in whom, and why. That distinction sounds straightforward until you try to draw the line in practice, at which point most organisations discover they've handed AI the wrong half of the problem.

The diagnostic work is not the same as the ability to look. Watching a team and understanding what you're seeing are entirely different skills, separated not by intelligence but by years of work on yourself- enough experience in group settings to recognise when your interpretation of a situation reflects the room rather than your own history, enough bias training to distinguish genuine pattern recognition from your assumptions wearing the costume of insight. Without that self-knowledge, the diagnostic stage produces not a clear brief but a mirror: AI then faithfully builds content that addresses the problem as it was perceived, rather than as it is.

So the design decision isn't simply "AI does production, humans do everything else." It's more specific than that: AI handles what can be specified in a prompt; humans are responsible for everything that has to be understood before the prompt exists. Getting that distinction right- and revisiting it honestly as models improve - is the most consequential choice in any AI-assisted learning workflow.

3

A 90-DAY GUIDE TO SHIFTING YOUR ORGANISATION'S APPROACH

Most organisations don't have a bad learning strategy. They have a strategy that made sense before AI changed the production equation, and nobody has stopped to ask whether it still does. The good news is that the shift doesn't require burning everything down. It requires a short period of honest interrogation, followed by some deliberate choices. Here's a structure that works.

DAYS 1–30

STOP BUILDING AND START LISTENING

Before touching a single piece of content or a single tool, spend a month asking one question of everything in your existing library: what was this designed to change, and do you have any evidence it did? Not completion data. Evidence of behaviour change.

Most organisations find, at this stage, that a significant proportion of their content either has no measurable outcome attached to it, or was never observed against the outcome it claimed to address. This is uncomfortable and important. It tells you where the real work is- which is rarely where the production pressure has been focused.

In parallel, resist the temptation to start experimenting with AI production tools. That comes later. Right now, the goal is to understand the shape of what you actually have before AI makes it trivially easy to produce more of it.

DAYS 31–60

DEFINE "GREAT" BEFORE YOU SCALE ANYTHING

This is the work most organisations skip, and it's why AI-assisted content production so often produces more content that looks like learning without functioning as it. Before you give any team- human or AI- a brief to produce something, you need to be able to answer three questions precisely: what should the learner be able to do differently after this that they couldn't before? How will you observe whether they do? And what does the content need to do to create the conditions for that change?

The answers to these questions constitute your quality standard. Write them down. Make them specific enough that two people, independently, would evaluate the same piece of content the same way against them. This is harder than it sounds, particularly the second question - organisations that have been measuring completions for years often find they've never actually decided what observable change looks like for their most important programmes.

Once you have this standard, you have something genuinely useful: a framework against which AI-generated content can be evaluated with the same rigour as anything a human produces. The standard doesn't care who wrote it. It only cares whether it works.

DAYS 61–90

RUN ONE HONEST EXPERIMENT

Choose one audience, one capability gap, and one piece of content. Use AI deliberately in the production- not to see how fast you can move, but to test where in the workflow it adds genuine value and where it produces something that looks right but isn't. The distinction matters and it won't be the same in every organisation or for every content type.

Measure it properly. Not completions- the actual behaviour you defined in month two. If you can't observe it, redesign the measurement before you ship the content. A programme you can't evaluate is not a programme. It's a gesture.

What you learn in this experiment sets the standard for everything that follows. The specific decisions- where AI drafts, where humans review, where the line sits between acceleration and compromise-should come from evidence, not instinct. The experiment gives you that evidence. It also gives you something more valuable: a shared language within the team for talking about quality in a world where content can be produced faster than it can be evaluated.

WHAT LEARNING BECOMES

Benjamin Bloom spent his career chasing an uncomfortable finding. One-to-one tutoring improved learning outcomes by two standard deviations compared to conventional instruction. Two sigmas. The difference between average and exceptional. He called it a problem because nobody could figure out how to achieve it at scale. Classrooms are not tutors. Content libraries are not tutors. For fifty years, the gap stayed open.

AI is the first technology that makes closing it plausible. Not because it generates content quickly, but because it can- when designed well- respond to the individual. To this learner, to this gap, to this moment. That is a genuinely different thing from anything the industry has been able to offer at scale before. But it only works if the foundation is right. AI produces content that is retroactive by nature- built from what has already been written, observed, understood. The most important learning problems are prospective: capabilities that don't yet exist, for situations nobody has encountered yet, in contexts changing faster than any library can track. That gap will always need human judgement to close. It always will.

So here is what learning actually becomes: better, or more. Better, if the people building it are honest about what they're trying to change, rigorous about whether they achieve it, and clear about which half of the work only they can do. More, if they aren’t- more content, more completions, more dashboards full of numbers that don't mean anything.

That choice is sharper now than it has ever been. AI has removed the last excuse for not making it deliberately.