Can an algorithm recommend art? The limits of digital curation

An algorithm can recommend art, but only in a narrow sense: it is very good at showing you more of what you have already liked, and very poor at the thing real curation does best, which is to put something unexpected in front of you and make the case for it. A recommendation engine matches patterns; a curator makes a judgement. Those are not the same task, and confusing them is where digital curation quietly lets you down.

What can an algorithm actually do well?

Algorithms are genuinely good at sorting scale, and that is no small thing. Faced with tens of thousands of works, a recommendation system can group them by colour, medium, subject, price and the behaviour of people whose taste resembles yours, then hand you a shortlist in milliseconds. If you have spent an evening saving coastal paintings, it will find you more coastal paintings, and some of them will be lovely. For the early, exploratory stage of looking, this is a real convenience. It spares you the infinite scroll and gives browsing a shape.

The honest framing is that an algorithm is a sorting tool, not a taste. It tells you what is statistically near to your past behaviour. That is useful for narrowing a field, in much the same way a directory filter is useful: a way to move quickly from everything to a manageable few.

Where does digital curation fall short?

Digital curation falls short precisely where human curation earns its name: in context, judgement and risk. A curator does not just retrieve work that resembles other work; they decide why a particular piece deserves your attention now, what it sits beside, and what it asks of you. They will hang an unsettling painting next to a calm one on purpose. They will show you an artist you would never have searched for, because they believe the encounter is worth having. None of that is reducible to a similarity score.

The deeper problem is the feedback loop. Most systems optimise for engagement, which means they reward what you already respond to and quietly bury what you do not. Over weeks, your recommendations narrow rather than widen. You end up in a comfortable corridor of near-identical work, convinced you are exploring when you are really being shown your own reflection. The surprise that changes how you see, the whole point of looking at art, is the first casualty.

An algorithm can tell you what is like the art you already love. Only a person can tell you why you should love something you have never met.

Why does the surprise matter so much?

The surprise matters because taste is built, not retrieved. Almost everyone who cares about art can name a work that meant nothing to them at first and everything to them later, and that shift usually happened because someone, a curator, a gallerist, a friend, insisted they sit with it. A system tuned to your existing preferences will never insist. It has no reason to show you the difficult thing, because the difficult thing performs badly on the metrics it is built to chase.

This is why a curated space still does something a feed cannot. Walking into a programme such as Baltic Centre for Contemporary Art in Gateshead, you encounter choices made by people willing to risk your indifference for the chance of your astonishment. An artist-run space like Bloc Projects in Sheffield does the same at the emerging end, showing early-career work precisely because it is unproven. Neither is optimising for your click.

How should you use recommendation tools, then?

Use them as a first draft, never a final answer. A recommendation engine is excellent for generating candidates: a list of names, a sense of what is out there in a medium or a price range, a way into a field you do not yet know. Treat its output the way you would treat a search result, as something to investigate rather than trust. The moment it starts feeling like the work is choosing itself for you, widen the input deliberately. Follow artists who are unlike the ones you already save. Look at the galleries those artists show with. Visit a studio complex such as Bow Arts in east London, where what you see is decided by who happens to be working there, not by what you clicked last week.

The practical discipline is to let the algorithm narrow and let a human, including you, decide. Software is unbeatable at reducing ten thousand options to fifty. It is hopeless at telling you which of the fifty will still matter to you in a year.

So can an algorithm recommend art?

It can recommend art the way a map can recommend a city: it will get you to the well-trodden places efficiently and never once tell you where to actually go. Digital curation is a fine servant and a poor master. Lean on it to handle scale, to surface names, to make the first cut. Then close the laptop and let a curator, a gallery wall, or your own restless eye do the part no model can: choosing, in front of the real thing, what is worth keeping. The screen is where looking starts. It has never been where it ends.