← Thinking out loud 7 November 2026

Facilitator or substitute? Using AI in peer learning without losing the people

A follow-on to my post on AI as a facilitator for peer learning groups, prompted by research on workplace teams published in Harvard Business Review in October 2026. The research is about teams, not learners. Applying it to learning is my own step.

In my last post I asked whether AI could be the facilitator that peer learning groups can't get for themselves. Something that helps people prepare before a session, asks the challenging question peers are too polite to ask during it, and captures what was learned afterwards. I still think it can.

But I left a question hanging, and new research has made it harder to ignore. What happens when the AI stops facilitating the group and starts replacing it?

Put simply: when a learner is stuck, who do they go to? And if it's no longer each other, what is the rest of the group losing?

Relationslipping

In October, researchers from BetterUp and Stanford published work in Harvard Business Review on what happens when people turn to AI instead of a colleague. They surveyed 1,033 US desk workers and gave the pattern a name that has stuck with me:

Relationslipping: working relationships weakening as people substitute AI for each other.

In teams where this was the norm, coordination and trust were both more than 20% lower.

The mechanism matters for learning. Groups run on a shared sense of who knows what. Each time someone skips a conversation for a quick answer from a model, that map stops being updated. When people hand their thinking to AI, it becomes invisible to everyone else. Swap "team" for "cohort" and you have a description of a peer group slowly slipping apart, while every individual in it feels more productive.

A caution before going further: this is a survey, it shows association rather than cause, and the authors say themselves the causal effect hasn't been measured yet. The findings are a prompt, not a verdict.

Why an AI default?

It would be easy to write this as a warning and stop. The honest problem is that every individual choice to go to AI is reasonable.

The research found people with low psychological safety were more than twice as likely to turn to AI instead of a teammate. One said, more or less, that it saved them from looking like an idiot. The article also cites research in Science finding that leading models affirmed users' actions 49% more often than people did. So the alternative to a peer is a partner that is always available, never judges and tends to agree. Of course people choose it. The authors call this a collective action problem: everyone acts sensibly, and the group ends up worse off.

That is what makes this so tricky for anyone designing learning. The AI that could facilitate a group, prompting preparation, challenge and reflection, is the same AI a learner can use to skip the group altogether. Nothing in the technology decides which one you get. The design does. And to see what's lost when the design gets it wrong, it helps to look at a form of learning that depends entirely on peers.

Action Learning as the test case

For me that's Action Learning. I try to build Action Learning Sets into the leadership development programmes I design, because they bridge the gap most programmes struggle with: between the learning event and applying it in practice. Participants go back to work, try something, and bring what happened back to the set.

Reg Revans' set is a small group of peers, each bringing a real problem they own, helping each other through questions rather than advice. His shorthand, L = P + Q, puts learning down to programmed knowledge plus questioning insight. AI is extraordinarily good at P. The set exists for Q.

You can tell a model to only ask questions, and it will do a convincing impression of a good set member. What it can't do is ask a question shaped by its own job and mistakes, learn anything itself from working on your problem, or hold you to the action you promised last month. That last one is the bridge back to practice. Take it away and the set stops doing the job I put it there for.

There's a subtler risk too. Action Learning works on raw problems, where the presenter often finds out what the problem really is while being questioned. If people work their issue through with AI first, they arrive at the set with something tidy, half-solved and already defended. The set reviews a conclusion instead of exploring a problem.

Revans called set members comrades in adversity. That phrase matters more now than when he wrote it. A sparring partner sharpens you. A comrade is in it with you: they have their own problem on the table, they take the same risks, and they notice when you don't come back.

AI can be a good sparring partner. It can't be a comrade.

Using AI without losing the people

So how do we get the facilitator without the substitute? I don't think there's a clean answer, but there are design choices that tip the balance. The principle behind all of them: AI should send people towards each other, not away.

  • Let AI work the edges and keep the middle human. AI helps people prepare before a session and reflect after it. The thinking together stays with people.
  • The group first, AI afterwards. Take the raw problem to peers before polishing it with a model, not the other way round.
  • Bring AI's answers into the room as material, not conclusions. "Here's what AI suggested, and why I'm not convinced" is a good way to open a set.
  • Make AI use visible, so people can still see each other's thinking.
  • Build the safety that makes asking a peer feel as easy as asking a model. This was always good practice. Now it's the thing standing between a group and its replacement.

None of this is free. Each of these choices is slower than letting everyone go straight to the model, and someone will have to defend that to a sponsor who wants efficiency. That's the hard part, and I don't think we should pretend otherwise.

Here's the test I'd now apply to any AI-enabled learning design. Does it make the people in the group need each other more, or less? A facilitator does the first. A substitute does the second, and that's where relationslipping starts. The tool can be the same. The design decides whether your learners end up with a sparring partner, or keep their comrades.


Reference: Niederhoffer, K., Rapuano, K., Liebscher, A., Robichaux, A. and Hancock, J.T. (2026) 'Team dynamics weaken when employees turn to AI instead of colleagues', Harvard Business Review, 15 October.

Have a learning challenge, an opportunity or an idea you’re still shaping? I’m happy to talk it through.

Get in touch