Essay · Technology & Attention

Beyond the Echo Chamber

What retraining Spotify taught me about artificial intelligence — and ourselves

Kevin J. Parker

I spent an evening teaching Spotify how to surprise me.

That was not the intention when I sat down. I only wanted it to stop recommending the same music.

For months my listening had been quietly narrowing, and the irony was that the recommendations were, in one sense, excellent. Spotify had correctly identified what I liked — contemporary classical composition, world music, experimental soundscapes, music rooted in landscape rather than celebrity. Having learnt those preferences, it seemed increasingly determined to reassure itself that it had understood them. Every recommendation resembled the last. Every new discovery arrived sounding faintly familiar.

The music was becoming an echo of itself.

At first I blamed the algorithm, which has become our default explanation for almost everything that troubles us online. Feeds amplify outrage, harden existing beliefs, and serve us more of whatever we consumed yesterday. All true. But sitting there, skipping through another playlist of near-identical strings, a less comfortable possibility occurred to me.

What if the algorithm was not simply training me?

What if I had spent years training it?

So I stopped asking for genres and started issuing commissions. Rather than searching for “new contemporary classical”, I asked Spotify to imagine itself the artistic director of an international festival devoted to Deep Time — the Earth’s four-and-a-half-billion-year history — and to programme it.

The result was unlike anything the search box had ever produced. Spectral composers appeared beside Inuit vocal traditions. Icelandic orchestral works flowed into Central Asian ensembles, into music written out of glaciers, volcanism and ocean currents. Encouraged, I repeated the experiment with Water, with Wilderness, with Climate, with the Five Elements, with Shamanic Journeying.

Somewhere in those hours the experiment stopped being about music.

The ecology of attention

The phrase echo chamber has always struck me as oddly architectural — a room, a hard surface, a sound returning unchanged. After that evening I began to think the better metaphor was ecological.

Healthy forests persist through diversity, disturbance, migration and renewal. Monocultures look productive and are quietly fragile, because repetition has replaced resilience. A single pathogen, a single dry decade, and the whole plantation goes.

Curiosity may behave in much the same way.

Every skipped song, replayed album, shared article and ignored headline is a small ecological intervention in our own mental catchment. Recommendation systems gather these signals and cultivate, faithfully, the conditions we reward. They are not malicious. They are agricultural. They plant what grew last season.

The trouble is not that such systems maximise confidence. It is that confidence steadily displaces serendipity — and serendipity is not a luxury of the intellectual life but its engine. Horace Walpole coined the word in 1754 for discoveries made by accident and sagacity together: you must be looking for something else, and you must be alert enough to recognise what you actually found. A system optimised to eliminate the first condition will, in time, atrophy the second.

Our intellectual worlds contract incrementally. Not through conspiracy. Through efficiency.

If attention behaves like an ecosystem, then an echo chamber is simply a landscape that has not experienced enough migration.

From search engines to curators

For most of the internet’s history the relationship was clerical. We asked computers questions; they retrieved documents. The skill involved was the skill of the good filing clerk — knowing the right term, the right operator, the right narrowing.

Artificial intelligence has quietly changed the terms of that exchange. Increasingly we are not asking machines to retrieve. We are asking them to exercise judgement.

The distinction matters more than it first appears.

Searching begins with an object already held in mind. You know roughly what you want; the machine’s task is to find where it is kept. The quality of the result is limited, absolutely, by the quality of what you already knew to ask for.

Commissioning begins with an idea whose shape has not yet emerged. You supply a frame, a constraint, a question, a set of values — and you allow the connections to be made by something that has read more widely than you have.

When Spotify was asked to curate Deep Time rather than recommend contemporary classical, it stopped sorting by category and started interpreting relationships. The Inuit vocal tradition was not adjacent to the Icelandic orchestral piece in any genre taxonomy. It was adjacent in the frame I had supplied.

Good curation is not the opposite of structure. It is structure placed in the service of exploration.

Which suggests that the quality of our digital lives may depend less on the intelligence of the algorithms than on the intelligence of the conversations we learn to have with them.

The commissioning mind

The morning after, I looked differently at almost every platform I use.

Instead of treating recommendation systems as shop assistants — trained to notice what you bought and offer you something adjacent — it became possible to treat them as research librarians, museum curators, commissioning editors.

Imagine asking a video platform to assemble a seminar on drought from an atmospheric physicist, an Indigenous knowledge holder, an environmental historian, an economist, a philosopher and a restoration ecologist — and to sequence them so that they argue with one another.

Imagine asking a bookseller for three works that would most seriously challenge the assumptions of the last book you enjoyed, and to say why.

Imagine commissioning your own morning newspaper on a principle of curiosity rather than political preference: five stories you would not otherwise have found, from three countries you rarely think about.

None of this requires new technology. It requires a different posture: the willingness to arrive with a question rather than a keyword, and to be told something other than what you expected.

What commissioning cannot do

Here honesty is required, because the argument has an obvious failure mode — the suggestion that the fix for industrial-scale attention capture is simply that we all become better prompters.

It is not.

The commissioning stance works where there is a prompt. It works in a chat interface, a search field, a place where you can state an intention and have it honoured. It works badly or not at all in the environments where most of the damage is done. You cannot brief an infinite scroll. The feed does not ask what you are curious about; it observes what detains you, which is a different and much cruder question, and one you answer involuntarily. No amount of individual sophistication rewrites a ranking system optimised for engagement, and none of it substitutes for regulation, transparency obligations or genuine choice about how these systems are tuned.

There is a second limit, quieter and closer to home. A commissioned diet is still a diet, and the frame you supply is still your frame. Ask for Deep Time and you get the music of a mind already drawn to geology and ice. The machine will oblige you brilliantly, which is precisely the risk. Serendipity that arrives entirely on request is not quite serendipity. It is a wider cage, well appointed.

The honest claim, then, is a modest one. Commissioning does not dismantle the echo chamber. It reintroduces migration — new species arriving from outside the plantation, some of which take.

That is not nothing. In an information ecology tending relentlessly toward monoculture, it may be a great deal.

Optimising for curiosity

We have spent two decades building systems of extraordinary power and asking them, almost exclusively, to make things efficient: faster retrieval, higher engagement, better prediction of what we will do next because of what we did last.

Artificial intelligence invites a different ambition. Not to optimise efficiency, but to optimise curiosity — to build and to use these tools in ways that widen the frame rather than narrow it.

The shift is small and it is not automatic. It is the difference between accepting a recommendation and issuing a brief; between the question what else is like this? and the question what should I be listening for that I do not yet know exists?

I still cannot pronounce half of what Spotify gave me that evening. Several pieces I have returned to a dozen times since. One of them I have no idea how to categorise, which is now the highest compliment I can pay a piece of music.

Perhaps the defining literacy of this age will not be learning to use artificial intelligence at all.

It will be learning how to commission curiosity.

Kevin J. Parker writes on environment, technology and consciousness at kevinparker.com.au