A few weeks ago, a producer from an international radio broadcaster got in touch. They were looking for an AI expert, and he told me he had found me through AI. We got as far as planning an interview before they moved on to another story, which is how these things usually go. A few weeks later, a young engineer in the region sent me a cold email asking for feedback on a grant application. It opened with the title of my website, word for word, and an old number from a resume I had forgotten to take down (it is gone now). Two very different people had reached me in the same way, and neither of them had found me by asking a person.

This recounting is not meant to point accusation at either of them. They used the tools available to them. I want instead to ask what it means for the rest of us when being found depends on what a machine can see, and on what it happens to say.

When I first came to Nairobi in 2010, the digital divide was mostly about access, about who could get online at all (remember cyber cafes?!). Those questions have not gone away (see for example, Endless Global’s laptop program). But I think a second divide is forming around legibility. AI tools that search on someone’s behalf mostly read the open web, and they read it best in English, from personal websites, institutional bios and conference pages. There is even a name now for trying to show up well in AI answers: Generative Engine Optimization (GEO).

Why me and not my friend?

If someone asked me who is doing important AI work in Africa, I would name a good friend of mine who co-authored Kenya’s AI policy. She is a popular speaker who does far more public events than I do, and we have co-authored papers together. So when the radio producer said AI had pointed him to ME, I was really surprised. I rarely write about AI so I still don’t understand why it surfaced me instead of her.

Since then, I have started running a small experiment to understand these dynamics better, and what they mean for efforts to diversify who counts as a knowledge producer and an expert. Every month I ask a few AI tools (logged out) the same questions about experts on AI and related topics in Africa. One question asks for a Kenya-based expert who could comment on AI and its impact on society. In the first round, ChatGPT’s top recommendation for that question was actually my friend, and I did not appear on any of the tools’ lists. (More on this experiment later in the series.)

No stable answer

I don’t know exactly what the producer typed or which tool he used, so I can’t say why it found me but I think that’s the point. Across four questions about experts, the three tools I tested named 98 different people, and only 14 of them appeared in more than one tool’s answer. In other words, if you ask a slightly different question, or the same question in a different tool or on a different day, and you seem to get a different set of experts! In research this would be called a reproducibility problem: you cannot run the same search twice and get the same result.

To be fair, web search was never perfectly stable either. Google results vary by location and search history, so a search from Nairobi may not match one from New York. But a page of results used to give you a list of sources you could open, compare and judge for yourself. You could see that some links were out of date, some were not really experts and some were not in Kenya at all, and you could read between the lines. In comparison, an AI tool gives you a confident list of names with a one-line reason for each, and often no sources (when I tested this, ChatGPT showed none for its lists of experts). Each answer sounds definitive, but it is one of many possible answers, and nothing on the screen indicates this. The radio producer had no way of knowing that a similar question, asked again, might have given him my friend instead.

Put together, I think two things are going on. Legibility decides who the tools can see at all, and among the people they can see, who actually gets named can come close to chance. The confident way the answers are presented hides both. The first part worries me most in places like East Africa where reputations travel through community meetings and WhatsApp groups instead of search results. In my experience, those on the ground in these less visible online spaces are often where the real experts are. For example, I’m thinking of the people who keep alive a library of Kenyan liberation history in Nairobi. Would any AI tool ever name them as experts on Kenyan knowledge production?

An echo chamber for experts

The answers may change from one search to the next, but they keep drawing from the same pool of well-documented people. If producers and funders find experts through AI, then the same people get quoted, and each quote becomes one more page for the next search to find.

In academia we’ve seen a version of this loop before. Most fields have their canon, that is to say, a short list of names every student in that field is expected to read and cite. In Anthropology and other fields, the canon has been critiqued as privileged certain people’s work over others, especially white men in Europe and North America and the problem is partly that, every new citation makes the canonical names even more central. Scholars have pushed back explicitly for years, rewriting syllabi, diversifying reading lists and making a deliberate practice of citing people who were left out. In a 2019 piece for the LSE Africa blog, I extended this work to argue that changing who we cite is not enough if the infrastructures that circulate scholarship keep re-entrenching the same voices. I wonder if AI search is fast becoming one of those infrastructures as well.

The echo chamber critique is familiar from social media, where recommendation algorithms keep showing people perspectives that confirm their own. I have not seen it made as explicitly about AI search and expertise, but the mechanism feels similar. I am wondering who the experts will become, propped up in part by these tools.

What to do with this

For funders, journalists and convenors, I think it means importantly treating AI answers as just a starting point. Ask the same question more than once, in more than one tool, and notice how much changes. Ideally then ask your human contacts and network too! Interrogate to understand who is missing. And anyone who wants to find the people their networks miss has to go looking, in places and languages that search engines don’t index well. Finding the right people is still real work.

This is the first of a few pieces I’m working on about what finding people looks like now. I haven’t seen much work on this yet, so if you know of others working on it, I’d love to hear about it in the comments. And if you fund, commission or invite experts, how are you finding the people your usual searches miss? Who would you recommend that a search would probably never surface?