I have a funny sort of job. I spend a decent part of each week either finding excuses to watch how people use technology or asking people slightly nosy questions about the tools they use. I do a lot of structured research, but this kind of lightweight, ongoing fieldwork just hums along in the background, gathering signals about what’s going on. In so much as I have a system for it, I call it “noticing”, and every now and then I notice something that starts to look like a trend. 

Recently I’ve noticed quite a few senior people in tech are starting to worry that LLMs are making them less capable. No one has any hard evidence for this – it’s just a feeling, or a suspicion – but it’s coming up in conversation more and more. Someone who, a month ago, was raving about how Claude had transformed their workflow said the other day that it was starting to make them feel “weird” and “overwhelmed”. Another person said they were forgetting how to do things, and a third said they thought they were losing skills. Other people have mentioned in passing that they've started spending more time intentionally offline - reading books and doing physical tasks - because they wanted to feel their brain work and one or two even used the word “stupid”. None of these people are particularly engaged in the “AI discourse”; they hadn’t heard of cognitive surrender or read the “Your Brain on ChatGPT” paper, they just explained that they felt different. And I think this feeling is going to be important. If the people with the purchasing power in organisations start to feel stupid, they’re unlikely to want to keep buying the technology. 

Feeling Good 

The best technologies make us better – they safely extend our capabilities so we can complete tasks we couldn’t otherwise do. Calculators make it easier to add up, binoculars make it easier to see things, telephones make it easier for us to talk to our friends. I’m short sighted and I wear contact lenses, which – after 30 years - still feel like a freakin’ miracle. 

Good digital technologies do a similar thing: when they work, they make us feel smarter, happier, more capable. Some of the good things I’ve done this morning include: send some messages to people I love, used a search engine to find useful information, set a calendar reminder so I don’t forget something important, and checked my bank balance on the go. I’ve also done some scrolling and posting on Bluesky, which is a bit more complicated – a mix of dopamine and disaster – but I still feel like I’ve learnt something, had a laugh, done some low-key showing off and got a couple of niggles off my chest. 

Our ability to qualitatively assess physical products – Does it work properly? Is it practical to carry around? Will it break if I drop it?  – does not so easily transfer to digital tools, which have different affordances. While calculators, binoculars, and telephones are all complex, they are also the sum of their visible parts - what you see is what you get. You can roll a pair of binoculars in your hands, if you were very determined you could probably trace where the major components came from and how they were made, and it’s pretty easy to tell whether or not they’re working. Digital products, however, combine the seen and the unseen. Data, algorithms, application updates, and buttons made from pixels are all on the edge of tangibility. They are not repairable by third parties, we can’t take the back off and have a poke around, or send them back if they break down within warranty. We don’t know how they’re made, where the data comes from, who was involved in its creation. Sometimes we might get freaked out if a tool is creepy or weird, but like it or not some people have a surprisingly high threshold for creepy and weird (after all, there’s a market for Ring doorbells and Meta glasses). What seems like a flaw for some people is a feature for others. 

If you spend a lot of time on LinkedIn, you might also be forgiven for thinking that productivity gains are the main driver of digital adoption, but I think there’s something more complicated – more emotional – going on, particularly at the early stage of a technology’s life cycle. 

Way before anyone starts making macroeconomic predictions about the impacts of a technology’s uptake, a product needs some early adopters. These early adopters include the people in your life who still watch the Apple keynote, might be easily drawn on how wild token use is getting, and are probably getting back into coding via Claude; they like a mixture of utility (that works) and novelty (that’s neat). It’s kind of like buying a car: they know all the sensible things it needs to do, but are also there for the visceral thrill. For workplace technology adoption, this is an important segment because a decent number of these people are CTOs who will make decisions that oblige large numbers of to people use their preferred software on a daily basis.* 

And my theory is that there’s a more emotional, vibes based component to genAI adoption that wasn’t in the room when the C-suite signed off on Office 365. After all, there is a significant number of basic things generative AI can’t do  – “You’re telling me this tool that can’t count the number of Rs in strawberry is going to obliterate the white collar job by the end of next year?” – but which don’t seem to stop people from buying huge numbers of enterprise licenses.

And I have no way of knowing this for sure, but I suspect that using generative AI has slightly altered some early adopters’ brain chemistry: since 2022, the balance of factors being weighed in consideration of Utility—Novelty has kept tipping towards novelty. While I’m sure most enterprise AI purchasers have completed a procurement matrix and a rock-solid business case, I’m willing to bet that the little twinge of superhero potential – that feeling of being smarter, more capable – when they fire up an application is a determining factor in making the purchase. 

It’s also clear that a lot of people who started out as developers are thrilled to get their hands dirty again with a bit of vibe coding – that using genAI makes them feel more in control of a tech stack that’s got away from them, perhaps even makes them feel younger and more dynamic – which in turn changes how they see their own job and the future of their organisation. I get a similar sense when talking to experienced people in other job functions: over the last 18 months I’ve worked with senior marketers, operations specialists, charity CEOs, and policy professionals who all feel kind of thrilled by the extension of their personal capabilities that LLMs afford.** 

LLMs are most useful for people who know what they are doing – experienced people who can spot errors quickly, know the obvious pitfalls, and feel frustrated they can’t get more junior staff to do exactly the thing they asked for. They are made for the kind of people who think “if only there were 10 of me” on a regular basis and who might secretly wish they were still doing the job they got promoted from 15 years ago. And I suspect this feeling of increased capability, of feeling better harder, better, faster, stronger, is what drives the buzz of adoption through the senior tier of organisations – softening up otherwise rational senior leaders to accept more implausible McKinsey charts and preposterous claims from tech CEOs. 

But if LLMs are going to endure for this group, then they need to keep delivering a thrill (or illusion) of competence. Bolder senior leadership teams who are vibe coding their way through a dozen different midlife crises are already creating a future in which other people are redundant – blowing up the skills pipeline without a thought for long-term consequences – but if the technology also starts to make them feel stupid, it seems likely they’ll start to pull back. So if a Claude subscription is the cutprice alternative to a midlife Lamborghini purchase, then it needs to keep delivering the thrills to stay useful. After all, no one buys a sports car because they want to feel stupid. 

———————

*I’ve had several conversations recently with people at big corporates who are assessed on how frequently they use AI tools in their working day – no one is monitoring their work for quality of outputs or uplifts in performance, just on frequency of AI use. Last week, I ran workshops with about 60 CFOs and finance directors – they almost all had access to Copilot, but only person was confident to say they knew how to make the best use of it. None of them had had training. 

**If you’re interested in tracking developers’ sentiment about LLMs, I highly recommend delving into the METR research on self-reported impacts on productivity on tech workers, which surfaces more on the emotional aspects of developers’ adoption of LLMs