The joyless learning trap, and why engagement is not a 'nice-to-have'

There's a word that makes L&D professionals visibly uncomfortable.

Not compliance. Not mandatory. Not even e-learning, though that one gets close.

The word is joy.

Mention it in a business context and watch what happens. People smile cautiously, as if they suspect a trap. Senior stakeholders shift in their seats. Someone says, "that's a lovely idea" in a tone that makes clear they mean the opposite. Joy is not, in the grammar of most organisations, a serious word. It belongs in the same category as wellbeing, flourishing, and psychological safety- things that are nominally valued and structurally deprioritised, filed under "soft" and quietly moved down the agenda whenever real priorities assert themselves. I want to argue that this is a catastrophic misreading of the evidence. And that as AI transforms the landscape of learning, getting this right matters more, not less.

Start with the neuroscience, because the neuroscience is unambiguous in a way that business conversations about engagement rarely are.

Barbara Fredrickson's broaden-and-build theory- one of the most replicated findings in positive psychology- shows that positive emotional states literally expand cognitive capacity. When a learner is in a state of genuine engagement, curiosity, or even mild amusement, their thought-action repertoire broadens. They see more connections, generate more options, take in more information. When they are in a state of anxiety, boredom, or the particular low-grade dread of mandatory training, the opposite happens. Attention narrows. The brain prioritises self-protection over learning, and new information fails to encode.

The University of Warwick found that happy employees are 12-13% more productive than their unhappy counterparts- not because they work harder, but because their brains process information more efficiently. The Yale Center for Emotional Intelligence found that employees who reported higher joy at work showed triple the creativity and significantly better memory retention. These are not marginal effects. They are substantial, replicated, and directly relevant to the question of whether learning sticks.

The anterior cingulate cortex- the region that processes social exclusion and rejection — does not distinguish between being left off a meeting invite and being the only person in a training room who doesn't understand what's being taught. Both register as threat. And under threat, the brain does not learn. This means that the design of the learning environment- the psychological safety of the room, the quality of the facilitation, the degree to which the learner feels seen and respected rather than processed- is not a soft consideration. It is a prerequisite for the content to work at all.

Here's where AI enters the picture in a way that I think is underexplored.

The dominant conversation about AI and learning is about content. What can AI generate? How accurate is it? How fast? How much does it cost? These are reasonable questions, and I ask them too. But they all sit on the production side of the equation, and production is not where learning breaks down; learning breaks down in the experience of the learner. More specifically: it breaks down in the moment when someone encounters a piece of content- a module, a course, a video, a scenario- and feels, consciously or not, that it is not for them. That it was made for a generic version of someone in their role, not for them specifically. That it does not know what they already know, or what they're struggling with, or what they need right now. That moment of non-recognition is where engagement collapses. And once engagement collapses, completion becomes performance- going through the motions of learning without the neurological conditions for any of it to land.

What AI makes possible, for the first time, is learning that knows who it's talking to. Not in the superficial "Hi, [Name]!" sense of early personalisation attempts. In the substantive sense: content that is calibrated to the learner's existing knowledge, delivered at the moment they actually need it, responsive to what they've engaged with before and what they haven't, capable of adjusting based on signal from the interaction itself.

This is not yet widely available. The systems that do it well are early-stage and, as I'm sure we've all encountered, imperfect. But the direction is clear, and the implications are significant: for the first time in the history of large-scale learning, it may be possible to deliver something approaching the experience that Benjamin Bloom identified as producing the most powerful learning outcomes- the kind that comes from one-to-one tutoring, where content is responsive, feedback is immediate, and the learner never loses the sense that someone is paying attention to them specifically. Bloom called this the "two sigma problem"- the finding that one-to-one tutoring improved outcomes by two standard deviations compared to conventional instruction. He spent his career trying to achieve tutoring-like results at scale. AI may be the first technology with a realistic chance of getting there. But here's the condition: it only works if the design of the learning system is built around the learner's experience, rather than around content production efficiency.

This is where the joy argument and the AI argument converge, and why I think they belong in the same conversation. Joy- real, biologically grounded, neurologically significant joy- is not a feature of learning content, but of of learning experience. It comes from feeling genuinely seen. From encountering an explanation that lands in exactly the way you needed it to. From the moment of understanding that cognitive scientists call insight- the "aha" that happens when new information connects to existing understanding and both reorganise around a new pattern. From the social experience of learning alongside people you trust, laughing at the absurdity of getting something wrong, being willing to try again.

None of this can be engineered directly. You cannot write "insert joy here" into a prompt and get it from a language model. But you can design for the conditions that make it possible.

Those conditions are:

• Genuine relevance to this learner in this moment.

• A sense of agency- the learner is choosing to engage rather than being processed through content.

• Appropriate challenge- hard enough to produce the small cognitive effort that makes things stick, not so hard that it triggers anxiety.

• And social connection, even mediated, even asynchronous- the sense that the learning is happening in relation to other people, not in isolation.

These are the design principles that have always produced good learning. AI does not change them, but rather changes is the scale at which they can be applied.

I want to be direct about the risk on the other side of this, because it's real and I don't think it's talked about enough. The organisations that use AI primarily as a content production tool- churning out modules faster, filling learning catalogues at unprecedented speed, measuring success in volume and completion- will produce a learning landscape that is, at scale, more joyless than what we have now. More content, more efficiently delivered to people who are less engaged with it than ever, measured by metrics that have never told us what we needed to know. The organisations that use AI as a learner-facing technology- as a way of getting closer to what each individual needs, when they need it, in a form that respects their intelligence and responds to their context- have an opportunity to build learning experiences that are more engaging, more effective, and more human than anything that has existed at scale before.

The difference between these two futures is about what you believe learning is for.

If you believe it's for compliance, for coverage, for the production of evidence that training has occurred, then AI is a production tool and a better completion dashboard is the goal.

If you believe it's for behaviour change- that the only meaningful measure of a learning programme is whether anyone does something differently- then AI is something altogether more interesting. It's the first real infrastructure for learning that responds to people rather than processing them.

The joyless trap is not inevitable. But it requires choosing, clearly and repeatedly, to design for the learner's experience rather than for the production metrics that are easiest to report. That choice is available right now. It will become harder to make, not easier, as AI accelerates the pressure to show output. Make it early, make it loudly, and build the measurement systems that make it defensible. The neurological case for joy in learning is solid and the business case follows directly from it. The AI case makes it urgent.

Alice Veitch is a Learning & Development Leader with a background in behavioural science. She has led learning strategy across EMEIA and is a respected disruptor in the industry.