Nobody Learns AI Alone

Worth establishing up front: the idea that technology adoption is socially determined has been formalized in the literature since 2000 and empirically supported ever since.

TAM2 extended the original Technology Acceptance Model by adding social influence constructs — subjective norm, image, and voluntariness — and accounted for roughly 60% of the variance in perceived usefulness across four longitudinal field studies. UTAUT went further, synthesizing eight prior acceptance models into four core determinants, of which social influence is one, and explained about 70% of the variance in intention to use — substantially above the 17–41% range of the models it consolidated.

So when I say adoption is social, I'm not making a contrarian claim. I'm restating something that has been one of the best-supported findings in information systems research for a quarter century.

Which raises the more interesting question: why do enterprise AI programs keep getting built as though it weren't true?

Enterprise AI adoption is almost universally designed as an information problem. People would use these tools if they knew how, so the intervention is to distribute the knowledge — training, documentation, prompt libraries, recorded sessions, an internal wiki.

That model produces the pattern everyone recognizes: strong training completion, weak sustained usage.

There's now direct evidence on this for generative AI specifically. A UTAUT study of 300 employees across large and small South Korean enterprises found that effort expectancy and social influence significantly drove behavioral intention to use generative AI systems — with social influence, including support from supervisors and peers, strongly driving adoption. Performance expectancy and facilitating conditions showed no significant effect.

That last construct deserves attention. Facilitating conditions is, roughly, the organizational support infrastructure — the resources, training, and assistance an organization provides. In that study it did not significantly predict adoption of generative AI. Social influence did.

If you are running an enablement program, you are investing in the construct that didn't move and largely ignoring the one that did.

The acceptance models treat social influence as a single construct. In practice it's several different mechanisms doing different work, and organizations tend to fund one and ignore the rest.

Casual conversation sets the range. Someone mentions offhand what they used it for that morning. It's a throwaway remark, and it does something no documentation does: it establishes that this is a thing a person like you does, for work like yours, without ceremony. Most of what I know about the practical range of these tools came from remarks nobody would have thought to write down.

Passive observation supplies the model. Watching a peer work through something gives you the shape of the interaction — how they framed the request, where they pushed back, what they did when it went wrong. That's procedural knowledge: hard to write down, easy to absorb by watching. And watching a peer is a different signal than watching a vendor or an executive. A peer establishes that this is achievable from your position.

Being challenged builds calibration. The skeptics are the most valuable people here and the ones enablement programs most consistently fail to engage. Someone articulating exactly why they don't trust the tool for their work is handing you a failure mode you haven't hit yet. Knowing where a system is unreliable is the core competency, and you cannot encounter enough failure alone, fast enough, to assemble it from personal experience. It circulates socially or not at all.

Explaining consolidates. You don't fully know something until you've had to explain it to someone else. Explaining forces your reasoning into the open, and you find out mid-sentence which parts of your understanding were load-bearing and which were pattern-matching you'd never examined.

That last one has a substantial evidence base, and an important condition attached. A meta-analysis of 39 studies found the learning benefit of teaching was g = 0.48 — nearly a medium effect — when people studied with the expectation of teaching. When they studied without that expectation and then taught, the effect was statistically indistinguishable from zero.

Which is a sharper finding than the folk version. It isn't the act of explaining alone that does the work. It's knowing in advance that you'll have to explain, which changes how you engage with the material in the first place. Practically: telling someone they'll be demoing this in two weeks is itself the intervention, before the demo ever happens.

Here is what the acceptance literature does not give you, and why I think this matters more for AI than for previous technologies.

TAM2 and UTAUT predict behavioral intention and usage. They are models of whether people will adopt. In TAM2, social influence largely operates through compliance — subjective norm is defined as the perception that people who matter to you think you should use the system, and its direct effect attenuates as experience increases and in voluntary settings.

That's social pressure producing uptake. It is not what I'm describing.

My claim is about capability, not compliance. The social layer gives you range — a sense of what this looks like from positions other than your own, which is what judgment is made of. Instruction gives you the general case. Solo experimentation gives you your own case. Only exposure to other people gives you the distribution.

And for these tools specifically, that distribution is the whole ballgame. Calibration — knowing where the system is reliable, where it isn't, and what to verify — is the difference between a workforce that can use AI and one that can be misled by it at scale.

No acceptance model measures that. They measure whether people use it, not whether they got good at it. Which means an organization can score well on every construct in UTAUT and still have produced a workforce with uniformly miscalibrated trust.

Adoption isn't a decision someone makes after being persuaded. It's what happens when a tool becomes socially present in someone's working life — when it's in the ambient conversation, when colleagues reference it casually, when a person has seen enough peers use it that not using it starts to feel like a choice rather than a default.

Programs try to produce that state directly, through campaigns and mandates and enablement. Social presence isn't something you can announce into being. It accumulates from a thousand small exposures, most too minor to be anyone's initiative.

So the useful question isn't “how do we drive adoption.” It's “what are we doing that suppresses the social layer, and what would support it.”

That means caring about whether people have any venue to show unfinished work. Whether reporting a failure carries a cost. Whether the skeptical are treated as a source of information or an obstacle to conversion. Whether anyone knows in advance that they'll be explaining this to someone — because the evidence says that expectation is doing more work than the explanation itself.

The informal layer is already doing most of this. It's happening in DMs, in hallway conversations, in the two minutes before a meeting starts, in someone screen-sharing what they built over the weekend.

It's invisible to measurement, which is why it gets no investment, which is why organizations keep funding the mechanism the evidence says works least well and wondering why the training didn't take.

The question was never whether AI adoption is social. That was answered before most of us started working. The question is whether you build for what the evidence already shows — and whether you're willing to measure the thing that matters, which is not whether people used it, but whether they got good.