The marketing says AI can do anything. The doom-mongering says AI will do everything. Both are wrong in the same interesting way: they ignore the specific, stubborn list of things these tools are genuinely bad at — a list that happens to read like a job description for the arts.
It has no taste
Ask an AI to choose the best of five ideas and it will choose the most average one, eloquently. These systems work by predicting what usually comes next — they are machines for finding the middle of the road, and the middle of the road is precisely where art dies. They can generate options, imitate styles, polish sentences. What they cannot do is want one thing more than another, or recognise the strange idea that shouldn't work but does. Taste is a track record of risks; a prediction engine can't have one.
It doesn't know when it's wrong
AI doesn't lie, exactly — lying requires knowing the truth. It generates plausible text, and most of the time plausible overlaps with true. When it doesn't, you get a confident, fluent, well-formatted mistake: an invented statistic, a rehearsal venue that closed in 2019, a quotation nobody said. This is why the rule in every sensible organisation is that a human signs off everything that leaves the building. The fluency is the trap: it reads like the work of someone careful, and it was written by no one at all.
It can't be in the room
The core product of the performing arts is not content. It's several hundred people breathing in the same space while something happens that will never happen exactly that way again. AI can generate a script, after a fashion; it cannot hold a pause while an audience decides whether to laugh. It cannot notice the house is cold tonight and lift the second act. Liveness — presence, risk, the possibility of failure in real time — is not a feature these systems lack this year and will gain next year. It's a different category of thing.
It can't hold a relationship
A donor gives because of a decade of trust. An artist accepts a smaller fee because the venue believed in them early. A participant returns because someone remembered their name. This is the actual infrastructure of the arts, and none of it can be generated, because a relationship is precisely the thing that can't be faked at scale. AI can draft the thank-you letter; it cannot be the reason the letter means something.
It can't take responsibility
When the programme goes to print with the wrong dates, someone owns that. Accountability — the ability to be answerable — belongs to people. Every organisation experimenting with AI rediscovers this within a month, usually via a small embarrassment, and the wise ones write it into their policy.
Now the uncomfortable part
Honesty cuts both ways, so here it is: AI is very good at adequate. Adequate marketing copy, adequate stock imagery, adequate background music, adequate coverage of the middle ground. If your organisation's output is distinguishable from adequate — if there's a reason it's yours — the machines are not close. If some of what you produce has quietly drifted into the generic, AI doesn't threaten that work so much as expose it. The mediocre middle is the contested territory.
Which makes the strategy unusually clear. Let the machines have the drudgery — there's plenty to give them — and spend the recovered hours moving your work further from the middle: stranger, warmer, more particular, more alive. Distinctiveness was always the job. It's just never been this literally true.
The image above was AI-generated — as ever, we'll always tell you.