Fish Food 684: Are we having the wrong conversation on AI and jobs?
On the AI jobs apocalypse, Google Stitch, why social media is polarising but AI is not, Claude computer use, and the infinity machine.
This week’s provocation: The hidden impact of AI on jobs and work
Given all the hype you’d be forgiven for thinking that we’re on the cusp of an AI-driven apocalypse in the jobs market. But are we? Last year Dario Amodei, CEO of Anthropic, described what was called at the time a ‘white-collar bloodbath’. Up to half of all entry-level white-collar jobs, he said, could be wiped out by AI within one to five years, and unemployment could potentially increase to 10-20%.
More recently Anthropic released a study on the labour market impacts of AI that included a graphic which was widely circulated at the time, but which was also widely misread. The chart plotted, for various occupations, the share of job tasks that LLMs could theoretically perform (in blue) against a measure of how much AI is actually being used in those roles today based on their own usage data (in red). The dramatic gap between the two across many categories led many to suggest that adoption only has to catch up with capability for AI to be able to handle the bulk of white collar work.
But the study’s actual finding pointed in a somewhat different direction. Whilst AI is far from reaching its theoretical potential, there is no systematic rise in unemployment among the most ‘exposed’ workers. The one early signal worth watching is a slowdown in hiring of younger workers into exposed roles, but even that finding is tentative. And the gap between what AI could do and what it is doing is likely to be far stickier and full of bumps in the road than a simple diffusion curve would suggest. AI will undoubtedly be transformative, but the timeline and mechanics that many commentators seem fixated on is wrong. Not only that, it potentially distracts from another real risk that nobody seems to be talking about.
A while back Ian Leslie made the point that there seems to be an inverse relationship between confidence in automation predictions and proximity to the actual work. He describes why it’s the hidden complexities in occupations that so often get in the way of simplistic predictions of automation. And he uses a fascinating example (originally from Dan Hanson): truck drivers. Almost a decade ago there was a lot of discussion on how trucking would be the first great automation bloodbath. Long distances, straight lines, repetitive routes. But truck drivers do a lot more than simply drive trucks. They secure loads, act as agents for the trucking company, deal with authorities and customers, handle situations when things go wrong - all things that require judgement, human interaction or physical effort. The point being that this pattern repeats across almost every occupation - the further you are from the actual work, the simpler it looks.
Looking closer to home, in agency and consulting work the tasks that look most like replaceable process from the outside are often the ones that are most saturated with contextual judgement on the inside. Most of what makes someone good at their job has never been written down, and likely never could be. Michael Polanyi called this the tacit dimension, the kind of understanding that is not easy to codify because it lives in experience, intuition, and human pattern recognition rather than in documents or processes. It's the strategist who can feel when a brief is asking the wrong question, or the account director who knows which battles to pick with a client and which to leave alone. LLMs are trained on what has been made explicit. They have no access to this implicit layer. This makes them brittle in ways that matter. They can execute a process but they lack what the Greeks called Phronesis, the practical wisdom to know what the appropriate action is when the situation is messy, ambiguous, or doesn't match any previous pattern. This is the difference between knowing the rules and knowing when to break them. And it's precisely this kind of knowledge that separates competent from good and good from exceptional in most white collar work.
Daniel Oks gives us another interesting angle on this question, focusing on why it was that ATMs didn’t directly replace bank tellers. It’s an example often used by those keen to show that technological-driven change doesn’t directly impact on jobs but which misses the second half of the story which is that the iPhone did. The author makes an argument for complementarity, saying that:
‘…labor substitution is about comparative advantage, not absolute advantage: the relevant question for labor impacts is not whether AI can do the tasks that humans can do, but rather whether the aggregate output of humans working with AI is inferior to what AI can produce alone’.
Oks suggests there will be a gap between the potential of technology and its impact on economic life, and human complementarity with AI will last much longer than many seem to think. It is, he says, only when we see the construction of entirely new paradigms that the full power of a technology can be realised. The ATM might have substituted some tasks but the iPhone made a wider set of tasks irrelevant.
One of my favourite examples of this phenomenon is electricity. In the 19th century factories were powered by steam, with the entire factory being designed around a huge steam engine that powered every work station from the centre. Everything had to be close to the engine, connected by belts and pulleys, spread across multiple floors to accommodate the power transmission system. When electricity came along the factory owners ripped out the steam engine and replaced it with a big electric dynamo. But they still had the same factory design, the same belts and pulleys, and the same workflow, and so productivity barely improved. It took 30 years for the real productivity impact of electricity to be realised. It was only when factories were completely redesigned so that instead of one big electric motor powering everything, a unit drive system featured hundreds of small electric motors, one on each machine. Suddenly you could redesign production around the flow of materials not the flow of power and mass production was born.
It’s unlikely that it will take 30 years for the true productivity benefits of AI to be realised of course but the electricity example is useful for a number of reasons. Like AI, it is a general purpose technology which enables many other innovations. Like AI, the technology required a complete rethink in the logic of how the system worked to realise the full impact. Factory owners waited until existing assets depreciated before investing in newly redesigned factories because of the economics involved. The new paradigm required a new generation of expertise like factory architects and electrical engineers that could work out the details in the context of many kinds of new industrial facilities in many different localities. Conventional productivity indexes at the time failed to account for qualitative enhancements from adopting the new system. So some of the early gains (things like better lighting, safer working conditions, more flexibility) were real but didn't show up in the numbers.
The ATM was like swapping the steam engine for a dynamo. The iPhone was like the unit-drive factory, a completely new paradigm that made the old architecture irrelevant. Right now, most AI adoption looks a lot more like dropping a dynamo into a steam-powered factory than building something genuinely new. This is the difference between task automation within an existing paradigm versus paradigm replacement. AI will require us to rethink our entire relationship with human labour, but also to redesign the systems in which work happens, and that will take time. Even where AI could replace roles, organisations are messy, political, path-dependent, and slow. Jobs aren’t allocated rationally, status is tied to team size and resources, processes become full of exceptions, and internal politics generate complexities that are opaque and hard to navigate. Conway’s Law means that it is hard to rearchitect and redesign workflows at the system level, not just at the task or functional level.
So let’s finish with the risk that no-one seems to be talking about. The Engels Pause (coined by Robert Allen) describes the period during the Industrial Revolution (roughly 1790-1840) when productivity surged but wages flatlined. For fifty years the gains went to capital owners while workers saw nothing. Allen named the phenomenon after Friedrich Engels who described it in his book The Condition of the Working Class in England. So a more immediate risk with AI is arguably that gains from improved productivity flow to corporate profits rather than benefitting the workers themselves. And we’re already seeing echoes of that same pattern emerging. History has shown that during major technological shifts, wages tend to lag productivity, sometimes for decades. It is typically only when industries are redesigned around the new technology, when new roles and skills emerge, and when labour markets tighten, that workers finally start to share in the gains. In the meantime, instead of freeing up time for workers to refocus on higher-level and potentially more rewarding tasks, AI risks intensifying and expanding work. You still have a job, but you’re running faster on the treadmill for the benefit of someone else.
Labour substitution is harder and will take longer than many people think. The real questions we should be concerned with are about power, distribution and organisational design. Who benefits from the productivity gains? How is work redesigned in a way that benefits workers as well as businesses? And do workers have any real leverage in shaping how AI is integrated into their roles? If we want to avoid a modern-day Engels Pause, we need to spend less time worrying about whether AI can do our jobs and more time asking why the people doing those jobs aren't seeing the benefits.
Rewind and catch up:
Using AI to transform client relationships
Why is Corporate AI Innovation so Hard?
Photo by Merakist on Unsplash
If you do one thing this week…
Google have launched an AI tool (still in Beta) called Stitch which enables you to turn ideas into working UI designs for mobile and web, even if you’re not a designer. You can just talk to it to build and change a full UI. With the caveat that there’s a lot more to good design that just being able to produce designs, this looks like an extremely useful tool and is further evidence (if any were needed) of the democratisation of a huge range of tools and disciplines surrounding innovation. (HT Siddhi Mittal).
Links of the week
Some fascinating analysis in the FT by John Burn-Murdoch on how social media is populist and polarising but also how AI may be the opposite.
More signals that AI is now moving more towards infrastructure and not just capabilities. OpenAI released a new set of APIs that allow models to securely take actions in third‑party systems (things like updating databases, triggering workflows, or interacting with SaaS tools). In other words AI platforms are evolving into middleware layers that sit between users and software ecosystems, not just interfaces.
This is a useful guide from Anthropic on creating consistent brand assets within Claude. And Anthropic have launched Computer Use, which basically is the next level of a personal assistant that lives in your machine. It can ‘open apps, click, type, and see your screen on macOS’ and you can send messages from your phone to get your Mac to do things for you.
Meanwhile OpenAI really seems to be pivoting more towards enterprise adoption (HT Sean Betts)
Relevant to a lot of the work I do with clients around AI and critical thinking - research from Anthropic’s AI fluency index shows that simple techniques such as iterating and refining outputs, clarifying goals before engaging an AI, providing examples and specifying format are all indicators of strong AI fluency
And Wikipedia is cracking down on the use of AI in writing articles
I’ll be doing the next IPA Advanced Application of AI in Advertising face-to-face course in London on 30th April. Love to see you there.
And finally…
Demis Hassabis, co-founder of Google Deepmind, has always struck me as having a more diverse perspective and situational awareness than your average tech bro. This new book by Sebastian Mallaby (who had unprecedented access to Hassabis and Deepmind over 3 years) bills itself as a ‘revelation-packed portrait of a singular mind and a historic reckoning with the AI revolution’, and it looks to be genuinely interesting and a cut above your average AI-hype fest. There’s a good FT review of it here (HT Simon Andrews).
Weeknotes
This has been a quiet week (thanks to cancelled programmes in the Middle East) so I've used it to do some research and some writing, and to write some proposals. Next week I’ll be doing more of the same (thanks to cancelled programmes in the Middle East) but I’m also delivering a session for a small agency on integrating AI into their workflows which should be fun.
Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this edition please do like, share and pass it on.
If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.
My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.
Only dead fish swim with the stream.
I find the comparison of the implicit versus the explicit quite telling. For those of us in media, nuance and context are getting shunted aside - leading to a rush to judgment, the search for ever simpler answers and a perspective from nowhere. Thanks for your perspective!
Great piece Neil, thanks 👍