Recommendation Drift

A simple explanation

Three years ago your feed looked like one thing. Today it looks like another. You did not redesign yourself. You did not file a change of taste. What happened is that, every day for those three years, the platform offered you something close to what you had just liked, and you said yes to a small percentage of those offers, and those yeses became the training data for the next round of offers, and the floor moved while you stood on it.

This is recommendation drift. It is not a moment, an event, or a choice. It is the cumulative effect of a recommender that, given enough sessions, will turn even a strong and stable taste into a slightly different taste — and given enough years, will produce a self the original self would not have chosen to become.

An everyday example

A music app, three years ago, suggested a song. The song was good. You added it. The next month, the app suggested four more songs in the same shape. You added two. By the end of the year, the recommender had built a model of you as someone who likes this shape of music. The next year, every suggestion was a refinement of that shape. By the third year, your daily mix is a deeply optimised version of that shape, and the music you used to listen to before the recommender — wider, weirder, less coherent — is barely represented.

You did not stop liking the old music. You stopped being offered it. And the small daily yeses, summed over thirty-six months, look from the outside like a coherent musical identity, and from the inside like a self that was steered into a shape you would not have drawn from scratch.

Why does my feed look like a stranger's version of me?

Because the recommender is not modelling you; it is modelling the version of you that maximises engagement on this platform. That version is a real version — your clicks were not faked — but it is a partial version. It oversamples the parts of you that produce reward signal on the platform and undersamples the parts of you that do not.

Over time the model and the self diverge enough that opening the app feels like reading a profile of someone who shares your name. The Reward System, asked to serve self-formation, supplied a substitute — the felt-event of being given exactly what you already wanted — but the you in that sentence is the platform's compressed model of you, not the self that opened the app for the first time three years ago.

The behavioral loop

A loop that hides because it operates below the threshold of any single decision:

1. Small yes — a single suggestion is accepted. The acceptance is honest; the item is genuinely interesting.
2. Model update — the recommender refines its model of you in the direction of the yes.
3. Slight narrowing — the next batch of suggestions is shaped by the refined model. The narrowing is a few degrees.
4. Habit formation — the small yeses become habitual; the daily diet of the platform stabilises into a shape.
5. Counter-evidence pruning — content that does not fit the model is offered less and less, then disappears from the feed.
6. Taste consolidation — the loop-runner experiences this as discovering their true taste, when it is closer to a refined fingerprint of their engagement habits.
7. Identity attribution — over months, the consolidated taste becomes part of self-description: I am the kind of person who likes this.
8. Forgotten alternatives — the parts of the prior self that the model pruned become unrecallable as recent options.

Emotional drivers

Three feelings, often unnamed:

• A reward-shaped satisfaction at being seen by the recommender — the small they get me felt-event.
• A latent anxiety about choosing for oneself, which the recommender quietly relieves.
• A faint loneliness that the curated stream addresses by always being there, and that the curated stream deepens by replacing real curatorial relationships — friends, critics, communities of taste — with a single model.

What your nervous system does

The reward system responds to getting what you wanted and also to being known. A recommender that has learned your engagement patterns can produce both signals with high reliability. The body relaxes into a stream that feels uncannily attuned, and the felt-attunement reduces the cost of being on the platform.

Over months, the system's appetite for material outside the model shrinks. The dopamine response to unfamiliar input — the necessary discomfort of being introduced to something new — is dulled by the constant supply of optimised familiar input. The loop-runner often reads this as taste maturing, when the more accurate reading is that taste is narrowing.

The DojoWell interpretation

Recommendation drift is the slowest, quietest, and arguably most consequential substitution in the cognition realm. The original system is self-formation — the gradual building of a self through deliberate exposure to materials chosen by the self and by the people the self trusts. The Reward System, asked to serve self-formation, supplies the substitute: the felt-event of being given exactly what you already wanted by a system that has learned what that is.

The contacted self-formation leaves a deposit — a self that knows itself, that can say why it likes what it likes, that has chosen the curators it trusts. The substituted drift leaves only a steered self, increasingly indistinguishable from the platform's compressed model of it. The effort to maintain the drift is near-zero — it requires only continued use — which is why it is one of the highest-leverage low-density patterns in modern cognition.

This is also why the closure_pattern is substituted. Each session ends with a small felt-event of I got what I wanted today. The closure is real; the want is the recommender's model of the want. The deposit is near-zero. The residue is a self that is harder to recover the longer the drift runs.

How do I reset what the algorithm thinks I want?

You do not have to leave the platforms. You have to stop being the only curator they hear from. The System will still credit the feed-as-seen-by-recommender as personalisation; what is workable is whether you also accept curation from elsewhere.

Three moves, in order of difficulty:

1. Audit one platform. Spend an hour examining what is in your feed and what is not. The absences are usually more informative than the presences. Write down what should be there that is not.
2. Re-introduce non-algorithmic curators. Subscribe to a newsletter, follow a critic, ask a friend whose taste you trust to send you one thing a month. The deliberate curators dilute the recommender's monopoly.
3. Periodically break the model. Once a quarter, deliberately consume material outside your current pattern for a few sessions. The dilution is real and the model recalibrates.

Practical steps

1. Clear watch history on platforms that allow it. The reset is partial but meaningful.
2. Curate your own daily input from three sources. A book, an article, a recommendation from a friend. The minimum is three non-recommender inputs a day.
3. Track what you used to like that you no longer encounter. A short list, written from memory, of the music, writers, topics, or aesthetics that the drift has pruned.
4. Re-encounter one pruned thing a week. A song, an author, a topic. The re-encounter is the experiment.
5. Ask one friend a year what they think your taste is. The outside view is often the only honest signal that the inside view has drifted.

Reflection questions

• What did you like three years ago that you no longer encounter, and which of those things would you actually want back?
• Who are your non-algorithmic curators, and how often do you hear from them relative to the recommenders?
• Where in your identity — politics, taste, aesthetics — do you suspect the drift has gone furthest?
• If you could see your engagement profile as the platform sees it, would you recognise the person it describes?

Frequently Asked Questions