This July newsletter is late (published 17 Aug 2026) because I’ve taken a break from all things work-related over the summer while we moved into our new house, several months later than planned. I won’t get into the details but will say that the UK conveyancing system is like role-playing a Monty Python sketch; bewildering, frustrating, and comedic.
Just before stepping back I finished a paper with Euson Yeung on AI and clinical reasoning, submitted a chapter synopsis on the future of AI and health professions education, and completed the final edit of the book I’ve been writing with Benita Olivier, on AI and doctoral research.
All three of those pieces drifted towards a single question, which I can’t seem to move past. All of them have as their central claim the idea that you need real expertise to judge what AI produces, and that this expertise gets built by doing the work AI is offering to take off your hands. I still think that’s right. But what it doesn’t answer is the question sitting beneath it, which is, what we are actually adding to the system once the machine does the work better than we do? If the answer is that we merely check its output, that’s not going to survive many more model releases.
Reid Hoffman published a piece last week that arrives somewhere very close to where Benita and I ended up in the book, having come at it from a completely different direction. He’d been talking to Sougwen Chung, an artist who trains robotic arms on a decade of their own drawing data and then guides them live through an EEG sensor while they work. Chung has made themselves more machine-legible than almost anyone alive, but they still draw a line beyond which the machine cannot go; they describe this as wanting to be machine readable but not machine executable. The system can know everything about how they work, but not enough to do the work without them.
Their reason for this line is worth thinking about. Chung doesn’t argue that the machine will produce worse drawings, which is the argument you see in a lot of professional contexts i.e. the machine isn’t good enough (and will never be good enough) because of empathy, human relationships, creativity, and so on. This argument describes an essential human element that will always be necessary for the work. Instead, Chung asks why they would build something that takes away the thing they love most, noting that they draw more now than they ever have.
What Hoffman does with this is the part I keep returning to. The standard advice we’re seeing now is that we should use AI to automate the tedious parts of our work and keep the interesting parts, but Hoffman points out - and I agree - that tedium and skill formation overlap far more than that advice allows. Reps are boring but they’re also where pattern recognition and judgement come from. Think about professional athletes and musicians who build mastery through the “boring” reps. Reid’s example is the junior associate reading four hundred documents, who is bored, and who is at the same time building the instinct that later lets her see what matters in a deal. Hand that “boring” reading to a machine and the instinct is now being formed in the model rather than in her, except the model belongs to someone else.
Chung’s line in the sand is determined by two questions you can ask of any task: Is this where my learning happens, and if the system does this instead of me, do I get worse? This is almost exactly where Benita and I arrived in our book on AI and doctoral research and I’d be interested to know where you’ve drawn that line in your own work.
Something from me
Most of my time over the last 2 months has been spent working on Path, which I made publicly available in July. It’s a free tool for supporting professional development, as well as a more traditional CPD portfolio. You start by choosing a standard or framework you’re working towards, capture evidence as you go, and Path shows you where your foundation is strong and where the gaps are, so that the portfolio accumulates out of work you were doing anyway instead of being assembled in a panic the week before a deadline.
It’s still an alpha build, and access is invite-only through a waiting list for the moment; I’d rather make sure it works properly for a few people first before I open it up for individual sign-ups. A few people who lead professional organisations have been in touch to discuss whether Path might suit their members or their staff, which has been the most useful feedback I’ve had.
Everything else from the past couple of months is on the Recently added page.
Worth reading
Three pieces looking at the same problem from different angles:
- Ke, Jin, Ong and colleagues: AI-induced never-skilling in medical education (Nature Medicine). The authors separate never-skilling from deskilling, where deskilling means losing a capability you once had, while never-skilling means that the capability never forms, because AI was there when it would have been built. If you’re in the business of training practitioners, this changes what you’re protecting and when you have to protect it.
- Reid Hoffman: Become machine-readable, not executable. The piece I mentioned above above, worth reading for the new-hire test if nothing else. For any task you’re thinking of offloading, ask whether you’d be comfortable if someone joining your team never did it. If you wouldn’t, the task is load-bearing for expertise, and is probably something you want to protect.
- Corbin, Bearman, Fawns and colleagues: Hybrid reading practices: how GenAI summarisers can support rather than replace reading (Studies in Higher Education). A counterpoint, and one of the reasons none of this amounts to an argument for abstinence. Summarisers can genuinely support close reading, and the difference lies in whether they’re built into a considered practice or reached for by default.
Quotes that resonated with me
Every rep you take builds something that accrues to you. Every rep you hand over builds something too, but it accrues to whoever owns the model, gets trained on everyone’s work rather than yours, and arrives in your competitor’s hands in the same release… The model’s improvement is not your personal improvement.
Reid Hoffman, Become machine-readable, not executable
Why would I want to build a system that takes away the thing I love the most from me?
Sougwen Chung, quoted in the same piece
Our moral imagination is the raw material these systems learn from. That makes up how they will understand our world.
Chloe Lubinski, Anthropic
Something for the commute
Hoffman’s full conversation with Sougwen Chung on Possible: possible.fm/podcasts/sougwen. They describe the realisation that digitising their drawings had thrown away the time in them, the pauses and the order in which decisions were made, and then building a tool to put that back. It’s a good illustration of how much of the work lives in things the finished output doesn’t record. Again, this feels similar to the through-line of the book on AI and doctoral research I just finished with Benita; the final output can’t be the work because the work includes so much that isn’t captured in the output.