The Algorithm Doesn't Know You. It Knows the Version of You That Clicks.

Think about the last time you discovered something you genuinely loved — a band, a writer, a strange documentary about competitive forklift racing — and you found it not because an algorithm suggested it but because a friend pressed it on you, or you stumbled into it by accident, or you were bored in a way that platforms don't allow anymore. That experience of stumbling is almost gone. And what replaced it feels so much like taste that most people don't notice the substitution.

A 2025 paper published in Frontiers in Psychology introduced a term for what's happening: the algorithmic self^[2]. The researchers were interested in a specific and uncomfortable question — not whether recommendation systems reflect who we are, but whether, over time, they begin to construct who we are. Their argument is that the longer a person interacts with a recommendation engine, the more their expressed preferences converge with the preferences the platform has found most reliably productive. Productive meaning: clicks, watches, scrolls, shares, return visits. The platform is not trying to know you. It is trying to keep you. The distinction turns out to matter enormously.

This is not a new intuition. Anyone who has spent an afternoon on YouTube and ended up somewhere they didn't intend to go has felt the gravitational pull. But the Frontiers in Psychology paper does something more than describe the drift — it tries to trace the mechanism. Recommendation systems generate a behavioral profile from your clicks^[4] and watch-times and pauses. That profile then shapes what you're shown next. What you're shown shapes what you click. What you click updates the profile. The loop runs thousands of times before most users are aware it exists. By the time you notice you keep watching a certain kind of content, the platform has already built an architecture around that observation — and your future choices are being made inside a structure you didn't design.

The unsettling part isn't that this happens. It's that it feels like self-knowledge. People describe their recommendations as accurate, as the platform finally getting them, as proof that the system is smart. The accuracy is real. But it is the accuracy of a mirror that has been slowly tilting — and by the time it has moved, you can no longer remember where you were standing.

What a Behavioral Profile Actually Is

A behavioral profile is not a portrait of a person. It is a statistical summary of a person's engagement history within a particular interface, under particular emotional conditions, at particular times of day, against a particular menu of options. It captures something real — you did click those things, you did watch that long — but it is also radically incomplete in a way the system cannot see and the user rarely thinks to question. It doesn't know what you watched under social pressure. It doesn't know what you abandoned because your phone died. It doesn't know that you were anxious the week you watched twelve home renovation videos, or that you'd normally never care about flipping houses. It knows that you watched them. That signal gets weighted. Those recommendations follow.

Recommendation systems are fundamentally designed around engagement^[1], not preference satisfaction in any deeper sense. This is an important distinction. There are things you prefer that you would not necessarily click on under the design conditions of a fast-moving interface. Long essays. Slow cinema. Music that needs three listens before it opens up. Ideas that don't resolve. These things exist in your taste but they are often outcompeted, in the algorithmic sense, by content that produces a quicker and more measurable response. The system learns that you prefer what you click. What you click is shaped by what you're shown. And what you're shown is selected by an optimization process indifferent to your flourishing and very attentive to your attention.

“The platform is not trying to know you. It is trying to keep you. The distinction turns out to matter enormously.”

Researchers in human-computer interaction have been documenting what they call preference erosion — the gradual hollowing out of the range of things a person is willing to engage with, as repeated algorithmic exposure narrows the field. This isn't just about content getting more extreme, which is the version of this story that gets the most attention. It's subtler: people start to feel impatient with things the algorithm wouldn't have shown them. Slower, harder, less immediately rewarding material begins to feel like it's failing them, rather than like it's asking something of them. The algorithm sets a pace, and then that pace becomes the baseline. What doesn't meet it feels broken.

The Self You Build vs. The Self the Platform Builds

Identity is not stable. No serious theory of selfhood claims otherwise. Who we are shifts with context, relationship, age, circumstance, and the slow pressure of what we're exposed to. The question the algorithmic self research raises is not whether technology shapes identity — everything does — but whether recommendation systems shape it in a particular and particularly distorting direction. The claim is that they compress identity toward a smaller, more reactive, more engagement-optimized version of itself. You don't become more yourself over time on these platforms. You become more legible to the platform. Those are not the same thing.

The Frontiers paper focuses on what it calls preference-signal feedback loops, and the core finding is worth sitting with: users who had been active on recommendation-driven platforms for longer periods showed reduced ability, in structured interviews, to articulate preferences that weren't already reflected in their consumption history. When asked what they would want to discover, or what they wished they knew more about, longer-term heavy users were significantly more likely to describe something they were already watching or already clicking on. Their imagination of what they might want had narrowed toward what the algorithm was already giving them. The preferences weren't just shaped — they had become circular.

This is the part that makes the algorithmic self concept genuinely troubling rather than merely interesting. The recursive loop doesn't just change what you consume. It changes how you think about what you want. It installs a version of desire that is pre-optimized for the interface. And once that's in place, the interface stops needing to work very hard to keep you. You've already internalized its logic.

Personalization as a Form of Enclosure

“You don't become more yourself over time on these platforms. You become more legible to the platform. Those are not the same thing.”

There is a land-use concept called enclosure — the historical process by which common land was converted into private property, eliminating shared access and the informal economies that depended on it. The concept has been borrowed usefully in digital theory^[3] to describe how platforms convert open information spaces into managed, gatekept environments. The algorithmic self research suggests enclosure operates on something even more intimate: the open commons of your own curiosity. Before recommendation systems, the experience of discovering media was messier, more social, more accidental, and navigated through a much wider field of possibility. Platforms enclosed that field. They built fences called personalization. And inside those fences, they told you the pasture was perfectly designed for you.

The enclosure metaphor is useful because it captures the power asymmetry that personalization language tends to obscure. Personalization sounds like a gift. It sounds like the platform working for you. But the platform is optimizing for engagement, which is its interest, not yours. The two overlap enough that the distinction often gets lost. You do, in some measurable sense, engage more with content selected by a recommendation engine than with content you found yourself. But engagement is not satisfaction. It's not growth. It's not even pleasure in any durable sense. It is a behavioral signal that the system has successfully captured your attention for another increment. The platform calls this serving you well. You might call it something else if you'd had time to notice.

What Gets Lost When the Algorithm Wins

There are categories of cultural and intellectual experience that recommendation systems structurally cannot surface well, not because they are obscure but because they don't produce clean engagement signals. Difficulty is one. Some of the most important things a person can read, watch, or listen to require sustained attention, previous knowledge, and a willingness to be confused before being illuminated. Recommendation systems are bad at these not because they can't identify them, but because they can't confirm engagement quickly enough. The signal comes too late, or is too quiet, to be weighted properly.

Contradiction is another. Real taste is contradictory. You might love minimalist electronic music and also bluegrass. You might read literary fiction and also extremely pulpy crime novels. You might care about architecture and also competitive cooking. A behavioral profile can technically hold all of this, but recommendation systems are not designed to honor contradiction — they are designed to find the strongest signal and amplify it. Over time, the weaker signals in your profile atrophy. You start getting more of what you clicked most, not more of everything you actually are. The multidimensional person becomes a simplified render.

This has a cultural cost that is harder to measure than engagement rates but probably more consequential. Shared cultural reference — the thing that lets strangers discover they both love something obscure — has always depended on some friction in how people find media. When everyone is in a personalized feed, fewer people are in the same one. The common spaces where unexpected cultural transmission happens — the overheard conversation, the friend's random recommendation, the bookstore shelf you weren't looking for — have been partially replaced by systems that are very good at giving you what you already indicated you like, and not especially interested in the rest.

The Hard Question You Have to Ask Yourself

“The common spaces where unexpected cultural transmission happens have been partially replaced by systems that are very good at giving you what you already indicated you like, and not especially interested in the rest.”

None of this is an argument for returning to some imagined pre-algorithm golden age of cultural self-discovery. That era had its own gatekeepers, its own distortions, its own power asymmetries. Record labels and publishers and broadcast networks were not optimizing for your authentic self either. The question isn't whether mediation existed before recommendation systems — it always did. The question is what this particular form of mediation does differently, and what habits it installs in people who use it long enough.

What it does differently, at minimum, is this: it personalizes the enclosure. It makes the narrowing feel like accuracy. It rewards you, in small dopaminergic increments, for staying inside a profile the platform built from your previous behavior. It trains you to experience its output as self-knowledge. And it does all of this invisibly, without announcement, through the ordinary accumulation of recommendations that each seem reasonable and individually benign.

The Frontiers researchers stopped short of recommending specific interventions — their paper is more diagnostic than prescriptive — but the implication of their work is clear enough. If the algorithmic self is a real phenomenon, then the habits that might interrupt it are roughly the opposite of what recommendation systems reward: seeking out things you wouldn't click on, spending time in spaces the algorithm can't see, accepting the friction of not knowing what you want yet and looking for it somewhere a platform isn't tracking the search. These are not glamorous practices. They don't optimize anything. They are just the old, slow work of finding out who you actually are — which turns out to be much harder when a very attentive system is constantly offering you a more clickable version.

References

1. Beyond Algorethics: Addressing the Ethical and Anthropological Challenges of AI Recommender Systems (arxiv.org)
   Establishes that AI recommender systems are fundamentally designed around engagement optimization rather than deeper preference satisfaction.
2. The algorithmic self: how AI is reshaping human identity, introspection, and agency (doi.org)
   Introduces the concept of 'the algorithmic self' and argues that recommendation systems construct identity over time rather than merely reflect it.
3. Digital enclosure (en.wikipedia.org)
   Provides the digital enclosure framework describing how platforms convert open information spaces into managed, gatekept environments.
4. Understanding Social Media Recommendation Algorithms (knightcolumbia.org)
   Explains how recommendation algorithms generate behavioral profiles from user clicks and watch-times to shape subsequent content shown to users.