It gave me a playlist

I was nearing the weekly reset for one of my personal subscription accounts and had something like 22% of my weekly usage allocation remaining. That’s the kind of constraint that makes a person curious rather than careful, and so I gave a Fable agent permission to wow me. I told it to consider my recent projects, personal notes, past and upcoming events, and whatever else it felt was relevant in my personal memory vault, and I told it to deliver whatever it judged would create the most value for me: software, documents, research, anything at all. I told it to work without regard to efficiency and without coming back to me for guidance. And I told it to be creative and to consider things that had not occurred to me, and things we had never discussed in any previous session. “Autonomous berserk mode” was the phrase I used in my instructions. It felt like the right name for a carefree experiment in which the only judgement I supplied was whatever the model could infer from my stored history.

It made me a Spotify playlist. Then it proofread an essay I had already written.

Both were done competently. The playlist was fine.

In June I published a piece on this site about the imagination gap, and I argued that the gap widens as the models get stronger. I still think that is right. But as I’ve reflected on the piece since then, I’ve realised that I was describing two gaps and calling them one, and the distinction has turned out to matter more than the original argument did.

The gap almost everyone means by “the AI skill gap” is the operating gap: how well you drive the thing. Which model, which words, which settings, how to structure the context so the output comes back usable. That gap is closing, and I think it will keep closing.

Two mechanisms are closing it, and neither of them is you. The first is model competence: the current systems infer what you meant, decompose the job, and ask the clarifying question you should have anticipated. The second is harness quality: the scaffolding around the model now assembles the context, selects the tools, and recovers from its own mistakes. The specifying work that carried a job title in 2023 has been absorbed into the product, and when the product does the specifying, the distance between a skilled operator and a novice narrows with every release.

The best measured evidence points the same way. A Harvard–BCG field experiment run on GPT-4 in 2023 found that work quality scores rose more for the lower-skilled participants than for the highest-skilled on tasks inside the model’s competence. The bottom of an already-elite population lifted nearly three times as much as the top, and the gap between them compressed, as a finding. That was in 2023, with a model now generations out of date. I don’t think the direction since is hard to guess.

How much further the compression runs is where the evidence ends and my anecdotes begin. I watch people produce a lot of work with these tools, and the outputs are not diverging the way the anxiety says they should. Off the back of what I see, I think that the trend continues into the current frontier of AI when it’s applied to real, open-ended professional work.

But here is where it’s easy to conflate concepts, and where I think the common rhetoric has it backwards. The operating gap is closing because it is the kind of gap that can be closed for you: model competence and harness quality are mechanisms that we expect to improve with each new release, and every improvement narrows the output gap between a good prompt and a bad one. Imagination is not that kind of skill. There is no mechanism on anyone’s roadmap that makes you more likely to have something worth asking for. Folding the two skills into one “AI skill gap” is a logical error. You will not be more creative, or a better judge of what comes back, just because the operating gap closes.

The market disagrees with me, and it disagrees in public. PwC’s 2026 Global AI Jobs Barometer, built on more than a billion job advertisements, puts the wage premium for workers with AI skills at 62 per cent and rising, with postings for AI-specific roles growing far faster than the wider market. That is a widening, priced and paid. But look at what is being priced: “AI skills”, as a single bundle — the exact conflation. A premium on the bundle tells you nothing about which skill inside it the money is chasing.

There is a second objection that is stronger still, and it is not about skill at all. Section’s proficiency report found 69 per cent of organisations taking action on AI agents while 16 per cent of workers had actually used one. The person with agentic tooling, connected systems and an allowance that permits an hour-long autonomous run is not playing the same game as the person with a consumer chatbot and a policy forbidding them to do anything useful with it. That gap is real, and getting bigger, and so is the frustration of the workers caught on the wrong side of it. But it is a procurement gap and a permission gap, not evidence of a divergence of skill.

And the same Jobs Barometer piece later unbundles “AI skills” for me. It splits AI-exposed work into two tracks: “professionalised” roles, where the machine absorbs the routine and human judgement becomes the job, and “democratised” roles, where the machine simplifies the task so that non-experts can do it. The professionalised track shows roughly double the job growth and the faster salary rises, and entry-level positions exposed to AI are now seven times more likely to demand what used to be senior skills, such as judgement and leadership. The market is not, in the end, paying for operating skill. It’s paying for the things that operating skills used to stand in front of.

If the market knows this, why do the advertisements still ask for the bundle? My guess is that an ad naming judgement alone would fill the room with people who have it and will not point it at AI-driven work. The label is doing the screening. Which means the bundle is defensible as a hiring shorthand. It is only an error when you adopt it as a personal development plan.

So back to the playlist.

I will confess that for a moment, I adopted the easiest conclusion: that the machine is dull. But on further reflection, I don’t think that’s right, and the position I’ve arrived at has stayed at the front of my mind in my work ever since.

Firstly, there is the risk posture. A system trained hard against taking large unrequested actions on someone’s behalf will, handed an unbounded licence, reach for the safest defensible act available to it. A playlist harms nobody. Proofreading something I had already written is help that cannot possibly be unwanted. What I measured may have been nerve rather than imagination.

Secondly, and I think this is the real one: it has no stake. Human imagination runs on dissatisfaction and ambition. You picture a different way of doing something, or notice a gap in the workflow, because what presently exists irritates you, or bores you, or makes you look worse than you are. The model is never irritated, bored, or behind on anything. It has nothing riding on Monday’s meeting, or next quarter’s financial results. It does not want, and wanting is where the new thing starts.

Lastly, the training pulls against it. These systems are tuned toward output that a broad population of raters finds acceptable. The average of many people’s preferences is not anybody’s particular vision, and it never will be.

There is also the plain fact that my brief had no edge to it. An unbounded instruction is not a gift, and this is the most serious objection to the request I made of my agent. Would you be able to impress an acquaintance if they sat before you and told you to “do something amazing” in one attempt, with no further instructions? I’ve supported enough friends acting in amateur stage productions to place my bet against you. I stripped every edge out of my instruction on purpose, and then treated the result as a diagnosis.

I accept all of it, but you should notice that the conclusion survives every one of those explanations. Whether the cause is nerve, or the absence of a stake, or the pull of the training, or my own empty instruction, the result points the same way: the model can expand an idea but struggles to originate one. Give it a target and it will do things with that target that would not have occurred to me, quickly and often better than I would have done them. Give it no target and it will find the nearest safe thing that resembles help. Initiation is the step it does not take.

And I should say what origination means here, because in a professional setting it is not the romantic kind. Creativity for its own sake often produces a Jackson Pollock, and a Jackson Pollock is usually not what was commissioned by an employer or client. The origination that matters is a judgement: deciding that a particular thing, out of everything that could now be done, is worth the resources, including where the answer is experimental, or something that did not previously exist. Deciding what is worth doing is the creative act.

I don’t read the result as unflattering to anyone involved, because the experiment did the one thing it was equipped to do. “Do whatever you think is valuable” is empty. I gave the agent no dissatisfaction to work from, or ambition to work with, because I had not articulated either. I ran the experiment because my usage allocation was about to expire, not because I had a considered goal. So as a test of the machine’s imagination, the result deserves no weight, and I give it none. But as a demonstration of where the imagination in this partnership has to come from, it could hardly have been cleaner: with the judgement withheld, what was left was a playlist.

The ask is the artefact. What comes back is mostly a rendering of what was put in.

I’m being careful here, because I have argued the opposite-sounding thing on this site before. I have written that most of what people produce is reproduction (previously synthesised material, absorbed phrases, inherited positions, formats we have seen work), and that unique thought is rare and expensive, so we ration it. I’m not now going to turn around and claim imagination as the special human faculty the machines cannot reach. That would be both contradictory and untrue.

The accurate claim is that imagination is scarce in people too. The gap was never between people and machines. Instead, it runs between people who originate and people who reproduce, as it always has. What changed is the price. Reproduction used to cost money and hours, so it carried a signal, and a person could build a career on being reliably competent at it. Now it costs nothing, and it is worth roughly that same amount.

So what is left?

At one end of the spectrum, there is the option to reinvent your approach to the work. Same role, with the same outputs, but a better method: the report you used to spend a day on now takes ninety minutes. That is worth having, but it has a ceiling, which is set by the limits of your current job description.

At the other end of the spectrum, there is the option to reinvent the work itself, and this is a road most people never take. Not for want of capability, but for want of a permission that nobody is coming to give. The obstacle is the static tie to your title and function, and to the specific set of outputs you are measured on. All of it was drawn up for a world in which the doing was expensive, and a career’s worth of attachment to your sense of self has a lot of inertia.

On that note, I’ll say two uncomfortable truths: First, that the tie is being cut either way. The tools exist, and the roles are being redrawn. The scope of what counts as “your job” is likely to have already been affected, with or without your consent. And second, that the fear around all of this is not a character flaw. What you do and how you do it is part of who you are, and “imagine your work anew” is no small request when you’ve described yourself with reference to your work for twenty years. However, the alternative is that this will be done to you on a timetable that isn’t your own.

The way through, I think, is to relocate the attachment. In June I made a distinction between people who do the work and people who pursue the goal. In a snapshot, the two are indistinguishable, but only one of them is still standing when the form of the work changes. If what you are attached to is the goal (the thing your work exists to achieve), the form can change underneath you without taking anything that matters. Hold the goal fixed and let the form move. And once it can move, practise moving it: as these systems carry longer and longer pieces of work, the constraint is no longer how precisely you can specify a task, but how large a thing you can conceive of handing over. Most of us have never practised a big ask, because nothing before now could carry one: a junior needed a tight brief, and software needed exact instructions, so decades of delegation taught us to think in fragments. Now the machine can carry far more than we think to give it, and that unused capacity is what the old habit costs.

The world your role was drawn for has ended, and the drawing is being done again whether or not you pick up a pen. The people who redefine the work will not be the ones who prompt best. They will be the ones who can still be dissatisfied on purpose, and who can describe what they want instead.

The playlist is still in my library. I played a handful of the songs once, and added none to my Liked Songs. But I have thought about it often, which is, in the end, probably the most value it could have delivered.

If you aren’t the one doing it, somebody else might imagine your work for you. For now, it won’t be the AI.