We Learn to Use AI by Watching Others, and That Is Why Adoption Is Uneven

Bandura published Social Foundations of Thought and Action in 1986. The core idea was that people do not learn only by doing. They learn by watching others, noticing what happens to them, and deciding whether to try the same thing. Bandura called this observational learning, and he placed it inside a broader model of reciprocal determinism where personal factors, behavior, and environment shape each other continuously. I read the theory page for my comps preparation earlier this week, and I kept noticing something. The same mechanism that Bandura described for learning in general applies almost perfectly to how people adopt AI tools inside organizations. The adoption pattern is not random. It follows the observation structure.

Compeau and Higgins (1995) brought social cognitive theory into IS by studying computer self-efficacy, the belief that you can use a computer to accomplish a task. Their contribution was showing that self-efficacy shapes effort, persistence, learning, and effective performance. But the part that matters for this argument is their grounding in the full social cognitive model. Self-efficacy does not come from nowhere. Bandura (1977) identified four sources: enactive mastery (doing it yourself and succeeding), vicarious experience (watching someone else do it), verbal persuasion (being told you can do it), and physiological state (how your body reacts). Vicarious experience is the observational learning channel. When you watch a peer succeed with a tool, your own self-efficacy for that tool rises. When you watch a peer fail, it drops. This is not a minor effect. Bandura argued that vicarious experience is the second strongest source of self-efficacy after enactive mastery. For a new technology where most people have not yet had their own mastery experiences, vicarious experience may be the dominant source.

I wrote about AI self-efficacy as the new digital divide. The argument was that belief in your ability to use AI effectively predicts outcomes more than access to the tools does. But that post focused on the individual level. The question I kept coming back to was why AI adoption clusters so visibly by team and department inside the same organization. Two teams sit on the same floor. Same company. Same AI tools available. One team integrates AI into daily work. The other barely touches it. The standard explanations, training budgets and IT support, are usually the same for both. The difference is observational.

One person in the adopting team tried ChatGPT for a routine task. It worked well. They showed the output to a colleague. The colleague tried something similar. It worked for them too. Now two people are using it. A third person watches both, sees the pattern, and joins. Within weeks the entire team has normalized AI use. The mechanism is observational learning. Each visible success lowers the perceived risk for the next person and raises their self-efficacy. No formal training required. No mandate from leadership. Just one person acting as an unplanned model.

The team that does not adopt AI probably has no visible model. Someone might have tried it, got a mediocre result, and never brought it up. Or no one tried it at all because no one saw anyone else try it. The absence of observation is itself a signal. It tells everyone in the team that AI is not something people in this group do. The behavior, the environment, and the personal belief system all reinforce each other in a stable low adoption equilibrium. Social cognitive theory predicts this stability. Reciprocal determinism means that once a pattern is locked in, changing any one element requires changing the others too.

This explains why the AI champion model works. Organizations that designate or support an AI champion are not just funding an experiment. They are creating a visible model for observational learning. The champion uses AI in visible ways, talks about the results, and makes the behavior observable. The effect is not just that the champion produces useful outputs. The effect is that other people watch and think, if they can do it, maybe I can too. That is vicarious experience increasing self-efficacy. When no champion exists, no vicarious experience is available, and adoption depends entirely on enactive mastery, which most people will not attempt without first seeing someone else succeed.

The same mechanism explains why I think remote work has slowed AI adoption in some organizations. Observational learning depends on visibility. You cannot watch a colleague succeed with an AI tool if you never see their screen, never hear them mention it in a hallway conversation, and never notice that their output quality improved. In an office, these signals are ambient. You absorb them without trying. On Slack or Teams, the signal must be deliberately sent and noticed. A colleague posts an AI-generated summary in a channel. You might see it. You might scroll past. The observation is weaker, less frequent, and easier to miss. Remote work does not prevent AI adoption, but it removes the passive observational channel that social cognitive theory says is critical for spreading new behaviors. Organizations that want AI adoption in remote settings need to make observation deliberate: show and tell sessions, shared prompt libraries, visible output sharing. The ambient model will not work because the ambient is gone.

The organizational modeling effect is the same story at a higher level. Social cognitive theory distinguishes between individual modeling (one person watches another) and organizational modeling (what happens when leadership demonstrates behavior). When executives use AI openly, describe how they use it, and share outcomes, they create vicarious experience across the entire organization. The effect is broader but shallower, because the model is farther from the observer's own context. Watching the CEO use AI for strategic analysis is less directly transferable than watching your deskmate use AI for the same Excel report you both dread. But it still matters. It signals that the behavior is legitimate, that the organization values it, and that trying it will not be seen as deviant.

When leadership does not model AI use, the opposite signal propagates. No one in authority is seen using these tools. The implicit message is that AI is not serious enough for executive attention. Observational learning works in both directions. People learn what not to do by watching what their leaders do not do.

I think the practical implication for organizations is straightforward but uncomfortable. You cannot close the AI adoption gap by buying more licenses or running more workshops. Those address access and declarative knowledge. They do not address the observational learning structure. The gap between teams that have visible AI champions and teams that do not will persist because the mechanism that closes it is social, not technical. If I were advising a company on AI adoption, I would stop counting active users on the enterprise AI dashboard and start mapping where the visible models are. Which teams have someone whose AI use is observable to others? Which teams have none? The adoption pattern will follow the observation pattern. Social cognitive theory predicted this in 1986. We just have not built our AI strategy around it yet.