How Spotify, Grindr and Uber Make Their Predictions Come True
The systems measuring us have started deciding what is allowed to happen next.
Today: We still talk about algorithms as if they were overconfident librarians with venture capital: recommending a song, a date or a product, occasionally misunderstanding us in ways that are irritating but basically corrigible. That description is now quaint.
The systems claiming to read our preferences increasingly control the conditions under which those preferences can be expressed at all. They decide which song reaches an ear, which person reaches your screen, which price reaches your checkout and which job reaches a worker. Then they observe the response and call it demand, desire, value or risk.
This matters because the model no longer has to be right. Once it controls distribution, selection or price, it can make its prediction expensive to disobey and profitable to become. We adapt. The system records the adaptation. The loop tightens.
The strange new bargain is simple: become easier to model, and you may become easier to choose. Spotify offers a clean place to begin.
Spotify has found a wonderfully contemporary way to charge artists for being chosen by Spotify.
It is called Discovery Mode. Artists and labels identify songs they want prioritised, Spotify feeds that preference into the algorithms powering personalised playlists and the tracks become more likely to be recommended. There is no upfront fee. Spotify instead takes a 30% commission on recording royalties generated by streams in Discovery Mode contexts. No envelope. No radio programmer. No suspicious lunch. The shakedown has excellent UX1.
Spotify says songs entered into the programme receive, on average, 106% more monthly listeners, 186% more playlist additions and 181% more saves. Those are Spotify’s figures, not an independent causal audit, but the proposition is clear: surrender part of the revenue and the platform increases the chance that the platform chooses you2.
The listener may genuinely like the song. That is what makes the system so easy to mistake for neutral.
The issue is that Spotify intervenes in distribution first. Listeners respond. Their behaviour returns as evidence of preference. The platform helps produce the behaviour, then receives it back as information about what people wanted.
The model goes first, then cites the world it helped make. That is the strange loop.
We already know what happens when a measure becomes a target: people reorganise themselves around it, and the metric stops describing reality because it has started producing it. That is Goodhart’s law, now enjoying a lucrative second life in every dashboard on earth. But the newer systems go further. They do not merely score what happened, but decide what gets the chance to happen in the first place.
Researchers call one part of this “performative power”: an algorithmic system’s ability to change the population it measures. Personalisation increases that power; competition reduces it. A powerful platform can steer behaviour towards what is more profitable3.
Even “performative power” understates the shift, because steering behaviour is not the same as controlling the conditions under which behaviour can happen.
A metric used to be a camera: it recorded the world. Then it became an engine: people changed themselves in response to it. Now it is the gate: money, visibility and opportunity pass through the model before reality gets a chance to prove it wrong.
That is proxy power.
The Feed Creates Its Own Proof
The system does not simply measure behaviour but allocates the conditions that produce it. People adapt to those conditions, the adaptation becomes fresh data, and the model cites the world it helped create as evidence that it understood us all along.
Culture is where proxy power becomes easiest to see, because the industry has already spent years mistaking the dashboard for the audience.
A controlled 2025 audit of TikTok found that bots expressing particular interests were strongly funnelled towards matching content, usually within the first 200 videos. As amplification increased, engagement with previously unseen hashtags declined. It was a sock-puppet audit, not a universal map of every feed, but it caught the mechanism: a few signals become an environment; the environment produces more signals; the narrowing looks like preference4.
A separate 2025 study based on 27 interviews found musicians and creators negotiating between platform demands, visibility, artistic identity and career sustainability. Some described becoming algorithmically pigeonholed: visibility made anything less recognisable harder to introduce5.
This is not marketing. The creator adjusts the work for the feed. The feed distributes the adjusted work. The audience responds. The response returns as proof of desire. Soon the song needs a fifteen-second recognisable section, the author needs a crying-to-camera BookTok moment, and every profession once conducted off-camera discovers that it has, regrettably, become television.
The preference is not fake. It is simply not independent of the machine claiming to discover it. The recommender is no longer the shelf. It has moved into the studio and begun clearing its throat during the first draft.
Grindr Is Learning to Decide Who You Want
Grindr told investors in 2025 that its 14.9 million monthly active users send more than 130 billion chats a year. Its 2026 roadmap puts gAI across discovery, search, inbox ranking and reconnection. Features promise to find “more of your type”, prioritise your “strongest opportunities”, estimate response likelihood and compatibility and resurface previous connections with AI-written summaries6.
The app is no longer merely displaying nearby men but also proposing a theory of your desire, ranking the evidence and advising you which encounter deserves to exist next. Finally, a gay man can outsource remembering why he stopped replying.
The person is adapting too.
A 2026 qualitative study surveyed 113 online daters across 15 countries who had used AI-generated material in profiles or messages seen by people they later met. Participants often used AI to fabricate attractive dating personalities, frequently without disclosure. The researchers called this “AI-as-self”: the potential partner first encounters a model’s idea of an appealing person, presented as the person7.
The issue is larger than whether ChatGPT wrote your Hinge prompt about Sundays (we get it, you like a Sunday roast). The environment establishes which performances receive attention. Users infer the grammar and then AI helps them speak it. The matches become evidence that the performance was desirable.
You become more selectable before you become more knowable.
Romance now has casting, optimisation and post-production. Apparently meeting the person was holding up the pipeline.
Your Ceiling, Your Floor, the Platform’s Margin
In 2025, the US Federal Trade Commission reported that surveillance-pricing intermediaries it examined could use granular consumer information to determine individualised prices and discounts, or show higher-priced products based on search and purchase activity. The firms worked with at least 250 clients. The FTC offered a hypothetical: someone profiled as a new parent might be shown expensive baby thermometers first8.
The market has always wanted to know what you will pay. It can now watch your mouse hesitate over the button.
Turn the model around and the same logic appears in wages. Legal scholar Veena Dubal calls it algorithmic wage discrimination: using granular data and changing formulas to create unpredictable, variable and potentially personalised pay for broadly similar work9.
An Oxford-led study examined more than 1.5 million trips made by 258 UK Uber drivers between 2016 and 2024. Following dynamic pricing, researchers found that inflation-adjusted hourly earnings fell from more than £22 to just over £19 before operating costs, while Uber’s median take rose from roughly 25 to 29% and exceeded half the fare on some trips. Uber disputes the report’s figures and says its pricing is transparent10.
The platform sits between both participants with information neither possesses. It can estimate the customer’s ceiling, estimate the worker’s floor and control the offer shown to each. The gap between those two numbers is where the platform gets rich.
The Model Does Not Need to Be Right
Across culture, dating and economics, the proxy has moved upstream. The song must pass through the recommendation model, the person through the selection model and the transaction through the pricing model.
Access is increasingly offered on condition of machine legibility. The problem is not simply that models misunderstand people. Human beings do that for free, before lunch.
Once a system controls the conditions under which you can be seen, selected, funded, loved or priced, accuracy becomes optional. It only has to make becoming more like its misunderstanding materially advantageous.
The philosopher Eleonora Viganò has proposed a “right to be an exception to predictions”: people should not be treated as wholly determined by their past choices or by the group into which a system has placed them11.
A prediction should not be allowed to close the futures in which it would have been wrong.
Recommendation and personalisation are useful. That is precisely why this power matters. The line is crossed when prediction becomes permission—when the model becomes the only route to attention, intimacy or economic participation.
The danger is not that the model gets us wrong, but that it owns the gate, then offers us access in exchange for helping its mistake become true.
The freedom worth defending is the right to become someone it did not see coming.