When I wrote about the three dimensions of expertise last year, I was answering a practical question: why does GenAI work well for some people and produce such confident sounding bullshit for others? In that article, I arrive at situated, domain, and technological expertise. In order to use AI successfully, you need a solid grounding in at least one area, and at least a developing expertise in all three.
Those initial ideas are starting to be borne out in studies which highlight the importance of domain expertise. But in the months since, reading more widely for the PhD and the next book, I’ve started to feel that the model describes something without really explaining it. The three dimensions tell you what expertise is made of, but they don’t tell you what AI actually does to it.
I’ve also found that situated expertise – in my model the kind of lived experience, shared understanding, and relational meaning that comes from “time on the job” – is still one of the missing components in research into GenAI and expertise. But situated expertise itself is not a new idea, and so in this post I’ll spend a bit more time exploring the “squishy bits” of expertise that fall outside of domain and technological.
A thirty-year-old argument about plumbers
My “situated expertise” dimension draws on the social linguistics of people like James Paul Gee, and was arrived at via my studies of identity and language in my PhD. But I “borrowed” it from Gee, and not the equally diverse field of expertise and cognition. So, I went looking for older literature on expertise partly because I wanted to know whether the situated strand I’d added was novel or whether I was reinventing a wheel. It turned out, of course, that the wheel is very well worn. But more interestingly, the people who established those ideas were arguing about something I hadn’t even thought to ask.
In the 1990s, when researchers were trying to build expert systems, the first serious attempt to put human expertise into an AI model, they got into a long fight about where expertise actually “lives”. One camp said it’s in the expert’s head: facts and rules you could in principle extract and program. The other camp said no, expertise is social: it lives in the community and the context, not in any individual skull at all. Robert Hoffman and his colleagues, trying to settle it, used an example I latched onto immediately, since my dad is a (now retired) plumber.
Picture a plumber and a customer in a kitchen, talking through a problem. The plumber gets called away to an emergency. Where did the expertise go? Obviously it walked out with the plumber; his knowledge of plumbing is attached to him, as the authors put it, much as his kidneys and his toenails are. And yet “plumber” is also a role conferred by a whole apparatus of training, licensing, trade history and people willing to call him an expert. Strip that away and the same knowledge in the same head means nothing to anyone.
Their conclusion was that you can’t locate expertise in either place alone. The smallest unit that makes sense is the “expert-in-context.”
This is the tension I’d been circling, approached from the position of machine learning and data science. I’d added situated expertise as a third strand to round out the picture. But the plumber example suggests the picture has a deeper fault line running through it, which cuts across all three of my dimensions: part of expertise is portable, and part of it is conferred.
The portable part is what you carry out the door when you leave: your discrimination, your sense of what’s good, your ability to act. The conferred part is the standing a community gives you, the accountability that makes your judgement count for anyone other than yourself.
I think that distinction, more than the three dimensions on their own, is what tells us what AI is doing.
AI gives and takes
GenAI counterfeits the portable part of expertise – the part that is “carried away in the head” of the expert. It produces work that reads as though it came from someone with discrimination and a sense of what’s good, whether or not anyone with those qualities was involved. What it cannot do is settle in the conferred part. Not because the conferred part is off-limits to AI: people clearly do trust AI, whether they should or not. The problem is what happens after. Conferral is extended on the expectation that someone will be answerable for the result, and answerability is one thing that is hard to pin onto an AI model (something which the AI companies are counting on).
There’s an echo of this in a recent piece by Arvind Narayanan and Sayash Kapoor on why AI hasn’t replaced software engineers. They describe software work as a “decide, execute, deliver” sandwich (looks like a burger to me).
The execute layer – actually writing the code – is the part AI compresses dramatically. But the buns push back: deciding what to build in the first place, and being accountable for what you deliver. Their evidence is either funny or terrifying. When researchers looked at a hundred thousand developers using AI agents, the volume of code written went up eightfold, while the number of things actually released went up by about a third. The machine flooded the middle and the human bottlenecks at either end barely shifted.
I don’t work in software, but I recognise the shape of the burger-sandwich. It’s every educator who’s used AI to generate a term’s worth of resources in an afternoon and then spent three afternoons working out which ones were any good before daring to take them anywhere near students. It’s the entire reason I have been saying for three years that GenAI doesn’t actually save anyone any time. Generating the lessons is the “execute” layer. The judging and the back-and-forth over quality and purpose are the parts that were never really the burden, and AI doesn’t relieve them; if anything it makes the burden heavier, because now there’s so much more volume to judge.
These two separate literatures – one from 1990s expertise researchers and one from 2026 AI economists – use the same word for a controlling force that doesn’t change: accountability. Hoffman’s model pictures the expert as a kite, held steady by an external line labelled, among other things, “accountability mechanisms and performance expectations.” Narayanan and Kapoor put accountability at the deliver end of their sandwich as the thing capability improvements don’t entirely re-route away from humans.
So, the conferred part of expertise isn’t just about being believed. If it were, then AI would fit the bill (evidenced by how many people are willing to stake their career reputations on hallucinated reports). It’s about being answerable. You can route an enormous amount of work through a machine, but you cannot route the answerability through it. Even if the plumber’s reviews are good, even if you believe the plumber is an expert, and even if the plumber can speak their expertise (whistling through their teeth and shaking their head while calculating the fee), the conferral of that expertise is still contingent on them actually fixing the sink.
Trust issues
There’s yet another complicating factor with AI, though. Back when expert systems were shiny and new, researchers tested how much people trusted machine advice versus human advice. People would sometimes agree with the computer more, but they consistently trusted it less, and even when they were told outright that the system outperformed the human expert, they still wouldn’t extend it the confidence they gave a person. People in the 1990s had an instinctive discount on machine authority: a little voice that said, it’s only a computer, make sure you check its work.
The fluency of GenAI dissolves that instinct. The old systems wore their awkward, clunky, robotic machine-ness on the surface, and that clunkiness kept people’s guard up. A modern language model produces prose that carries the surface signal of a thoughtful, expert, human mind, notwithstanding the perennial tells that come from overuse or a heavy-handed, cavalier approach to churning out text with AI. The danger was never that AI is frequently wrong. It’s that it used to be wrong in a register that made us suspicious, and now it’s wrong in a register that makes us nod along.
In the original post on the three dimensions of expertise I argued that you need expertise to spot errors and hallucinations. Now, I’d now put it more sharply: the thing under threat isn’t only the knowledge that lets you catch a hallucination. It’s the instinct to look in the first place. Fluency doesn’t just slip more errors past the expert; it erodes the reflex the whole three-dimensional model depends on that says “this output is a claim to be checked, not an answer to be trusted.”
Where this leaves the three dimensions
The three dimensions still describe what expertise is composed of. I suppose what I’m reaching toward is a layer underneath it. Put crudely: domain and technological expertise together are mostly where the portable part comes from: the discrimination, the ability to tell good from bad, to know where the machine breaks. Situated expertise is mostly where the conferred part comes from: the social standing, the relational aspect, the accountability to a community. Which means situated expertise might be the one that matters most in an age of fluent machines, precisely because it’s the one AI can’t fake. You can prompt your way to something that looks like domain knowledge. You can leverage technological expertise to give a veneer of understanding. But you cannot prompt your way to being the person a community has decided to trust and hold responsible.
I think there’s something in this but I’m not quite sure what. I don’t know if the earlier Venn diagram is the right setup (though everyone loves a good Venn diagram), but I’m also not convinced that the domain, technological, and situated dimensions are easily separable and defined. I think that portability is mostly a feature of the domain and technological aspects, and conferral mostly a feature of situated expertise, but both could be true of the other.
A nice, clean, tidy diagram might look something like this:
But what about this technology, or expertise, or anything else ever fits into a nice tidy diagram?
Back to the drawing board…
I’d love to hear people’s thoughts. Please drop a comment wherever you found this article, comment directly on the blog by scrolling to the end, or get in touch via the contact form to let me know what you think:
Leave a Reply