Chapter 26: Geometric AI Curation — Recommendation, Discovery, and the Filter Bubble

RUNNING EXAMPLE — Priya at the Recommender

Priya has moved on from TrialMatch. The firm that acquired her clinical-matching work is a large streaming platform, and she is now lead scientist on the “Taste Engine” — the system that decides which songs, which films, which essays surface on each user’s home screen. On her first week, she pulls the engagement dashboards. Click-through is strong. Watch-time is strong. The A/B test logs show that every tweak to the recommender increases the scalar metric the product team watches. And yet: three months in, a pattern becomes unignorable. Users who begin the quarter listening to a wide spread of genres end the quarter listening to one. Users who begin reading across politics, fiction, and science end reading one political valence. The system is not broken. The system is doing exactly what its loss function asked. It is asking the wrong loss function. Priya’s predecessor left a note in the codebase: # TODO: figure out what "good recommendation" even means. She opens a blank notebook and begins again, this time with a manifold in mind.

This chapter applies the geometric framework to the engineering discipline that, more than any other, determines what art the twenty-first century sees, hears, and reads. Recommendation systems — the algorithms behind Spotify Discover, Netflix’s ranked rows, Amazon’s “books you may like,” TikTok’s For You page, and a thousand lesser engines — are the dominant aesthetic curators of the contemporary attention economy. Their scalar objective functions are precisely the target of this book’s critique. When we argued, in Chapter 2, that the four-star rating destroys aesthetic information, we were speaking of individual judgment. The recommender generalizes that destruction to billions of judgments compressed into a single optimization target. The filter bubble is not a sociological accident. It is the predictable geometry of greedy ascent on a scalar-valued objective over a high-dimensional aesthetic manifold. This chapter makes that prediction precise, empirically grounded where we have evidence, and prescriptively sharp where the framework has something to say that current practice does not.

We also acknowledge, at the outset, what we do not have: an end-to-end evaluation of a geometric recommender against a scalar-engagement recommender in production. No such study exists. What we have is the framework, the empirical findings of Chapters 17, 20, and 21, and a set of falsifiable predictions about what a geometric alternative would measurably change.

26.1 The Shape of the Problem

Consider the canonical recommendation pipeline. A user $u$ has a history of interactions $H_u = \{(w_1, r_1), (w_2, r_2), \ldots\}$ where each $w_i$ is a work (a track, a film, a book) and each $r_i$ is a response (a play, a rating, a completion fraction, a share). A learned embedding $\phi: \mathcal{W} \to \mathbb{R}^d$ places each work in a latent space. A user representation $u \in \mathbb{R}^d$ is derived, typically as a weighted average of the $\phi(w_i)$ in the user’s history. For a candidate work $w$, the system computes a scalar score $s(u, w) = \langle u, \phi(w) \rangle$ — a dot product, a cosine, a learned ranking head — and ranks candidates by $s$.

The loss function during training is, almost universally, a proxy for engagement. In early matrix-factorization systems (Koren et al. 2009), the objective was to minimize squared error on observed ratings. In contemporary deep-learning recommenders (the two-tower architectures of YouTube, the transformer-based rankers of TikTok), the objective is to maximize the predicted probability of a positive interaction — a click, a watch, a share, a save. The scalar $s(u, w)$ is the learned approximation to the expected engagement.

This is, in our framework’s vocabulary, a greedy local search on the aesthetic manifold. At each recommendation step, the system takes the user’s current location and moves in the direction of steepest ascent on $s$. There is no memory of the manifold’s global structure. There is no planner that considers where the trajectory will end up after a thousand such steps. There is no term in the loss that penalizes convergence to a small region. The system is gradient descent on engagement, nothing more.

The empirical consequence, well-documented across platforms, is convergence. The filter bubble (Pariser 2011) is one name for it. The rabbit hole (Ribeiro et al. 2020, on YouTube radicalization) is another. The genre trap (Anderson et al. 2020, on Spotify) is a third. In each case, the pattern is identical: user trajectories on the learned embedding concentrate over time. Variance contracts. The diversity of the consumed corpus declines. The effect is not equally strong for all users — committed omnivores resist it — but averaged across the platform, the direction is unmistakable.

Our framework identifies three failures of flatland in this picture, parallel to the three we identified for scalar aesthetic judgment in Chapter 2.

First failure: scalar irrecoverability. The predicted engagement $s(u, w)$ is a single number. From it, no downstream process can recover whether the predicted engagement is driven by genre affinity, by trajectory-geometry resonance with the user’s preferred pacing, by spectral-divergence match, or by some other channel. The loss is blind to channel identity. When the system recommends a work and the user engages, the reward does not propagate to the right channel — it propagates to whatever combination the embedding happened to encode, weighted by whatever the ranker happened to learn. In our book experiments (Chapter 20), we found that the four structural channels — spectral divergences, internal coherence, trajectory geometry, and genre-axis Lasso loadings — make distinct, non-redundant contributions to predicted reader rating. A scalar ranker collapses them.

Second failure: greedy locality. Gradient ascent on a scalar does not maintain diversity on a manifold. It finds the nearest basin of high predicted engagement and descends into it. On an aesthetic manifold with genre-sized basins and sub-genre sub-basins, the system’s trajectory is deterministic once the initial condition is set: the user falls into the nearest genre basin, then the nearest sub-basin within it, then the nearest cluster within that. What exploration exists is typically implemented as $\epsilon$-greedy random perturbation — a scalar-blind mutation that samples uniformly from the candidate pool and has no notion of directed exploration along the manifold’s geodesics.

Third failure: confounded objective. This is the failure we have direct empirical evidence for. In Chapter 21 we reported that, in the FMA-medium music dataset with MERT-v1-330M embeddings ($n = 24{,}801$ tracks, artist-disjoint CV), 91% of the hand-feature predictive $R^2$ on log-listens was genre confound: the within-genre residualized effect was $R = 0.043$ (8.4σ) compared to the raw $R = 0.302$. In the books pipeline (Chapter 20, $n = 4{,}998$), the corresponding confound was 85%. When a recommender optimizes predicted engagement without residualizing genre, it is, with high probability, optimizing a genre-classifier in disguise. The “personalization” is real but its resolution is coarse: it is personalization of which genre cluster you belong to, not of what within-cluster variation you respond to. The filter bubble is a genre bubble wearing a personalization mask.

These three failures compound. Scalar irrecoverability means the system cannot tell you why you liked what you liked. Greedy locality means the system will bring you more of the same on the easiest-to-detect similarity axis. Confounded objective means that axis is almost certainly genre, not aesthetic content. The result is genre-collapse disguised as taste.

26.2 The Engagement Manifold Is Not the Aesthetic Manifold

The critical distinction, which current recommender research largely elides, is between two manifolds that inhabit the same embedding space but differ in their geometry.

Let $\mathcal{M}_{\text{eng}}$ denote the manifold on which predicted engagement $s$ is smooth — that is, the learned geometry under which nearby points have similar engagement probability. Let $\mathcal{M}_{\text{aes}}$ denote the aesthetic manifold of Chapter 5 — the geometry on which structural-aesthetic content (spectral, coherence, trajectory, and the modality-specific axes of Chapters 20–25) is smooth. Current recommenders learn $\mathcal{M}_{\text{eng}}$ and operate on it as though it were $\mathcal{M}_{\text{aes}}$. They are not the same manifold.

We can say more. The engagement manifold is a projection of the aesthetic manifold along the dimensions that correlate with user click probability, weighted by the sampling density of the training data. If the training data is dominated by a handful of high-engagement genres, the projection is dominated by those genres. If the click signal is noisier on within-genre aesthetic variation (as our confound results suggest it is), the engagement manifold will smooth over that variation. The aesthetic manifold has structure that the engagement manifold has compressed away.

Proposition 26.1 (Engagement–Aesthetic Gap). Let $\phi_{\text{eng}}: \mathcal{W} \to \mathbb{R}^{d}$ be an embedding learned on engagement signal and $\phi_{\text{aes}}: \mathcal{W} \to \mathbb{R}^{d}$ be an embedding learned on aesthetic-structural signal (in the MERT or LaBSE sense of Chapter 17). The correlation between within-embedding distances, $\operatorname{corr}(\|\phi_{\text{eng}}(w_i) - \phi_{\text{eng}}(w_j)\|, \|\phi_{\text{aes}}(w_i) - \phi_{\text{aes}}(w_j)\|)$, is bounded above by the residualized within-genre aesthetic signal and is therefore small whenever genre confound is large.

We argue rather than prove, because the proposition is a claim about empirical objects (two embeddings learned under different objectives) rather than a theorem about intrinsic structure. The argument: engagement distances are dominated by whatever predicts clicks, and clicks are dominated by genre (Chapters 20–21, 85–91% genre confound). Therefore engagement distances are dominated by genre distances. Aesthetic distances, residualized for genre, retain only the 9–15% non-confounded signal. The two distance functions share only that fraction of structure.

The prediction is testable. An experiment that ran an engagement-trained two-tower recommender and an aesthetic-structural recommender on the same corpus, and that measured the between-embedding correlation on held-out works, would provide the direct test. We do not have that experiment. We predict the correlation would fall in the range $0.3$–$0.5$ — high enough to look similar on inspection, low enough that they recommend materially different works for the same user.

26.3 Filter Bubbles as Basin Convergence

The filter bubble, in geometric language, is the phenomenon in which repeated greedy ascent on $s(u, w)$ drives the user representation $u$ into a small neighborhood of the engagement manifold and keeps it there. We make this precise.

Definition 26.1 (User Trajectory). The user trajectory is the sequence $(u_t)_{t=1}^T$ where $u_t$ is the user representation after $t$ recommendation–interaction cycles. In contemporary systems, $u_t$ is typically updated as a running average or attention-weighted pool of recently consumed $\phi(w_i)$, so $u_t$ evolves stochastically as a function of which recommended works the user engaged with.

Definition 26.2 (Convergence Basin). A convergence basin $B \subset \mathcal{M}_{\text{eng}}$ is a region such that once $u_t \in B$, the probability that $u_{t+k} \in B$ for all $k \leq K$ exceeds some threshold $\rho$, for a horizon $K$ and probability $\rho$ characteristic of the basin. Large, easily-detected genres (pop, rock, electronic in the music domain; romance, mystery, fantasy in books) form the dominant basins on engagement manifolds learned from mainstream platforms.

Proposition 26.2 (Basin Convergence Under Greedy Ascent). Under a pure greedy scalar ranker, if a user representation $u_t$ enters a convergence basin $B$ with positive probability, the expected residence time in $B$ grows without bound as the catalog size $|\mathcal{W}|$ grows, provided $B$ contains sufficiently many unplayed works.

The intuition: larger catalogs mean more candidates within any basin, and the greedy ranker will always find one of them before it finds something outside the basin. The exit probability per step is inverse in basin density. Platforms with growing catalogs see strengthening filter bubbles for a structural reason.

In our music empirical work, we observed basin structure directly. Rock ($n = 7{,}088$), Electronic ($n = 6{,}284$), and Hip-Hop ($n = 2{,}190$) within-genre Ridge regressions yielded $R = 0.139$, $0.143$, and $0.141$ respectively — measurable within-genre aesthetic signal. But these within-genre effects are small; the cross-genre distances on an engagement-trained embedding are much larger. A user who enters the rock basin faces a landscape in which the nearest cross-genre work is farther (in engagement distance) than the thousandth within-genre recommendation. Greedy ascent will exhaust the basin before it escapes it, and the basin is for practical purposes unexhaustible.

The sociological consequences — political polarization, niche addiction, the collapse of shared cultural referents — are not our primary concern. Our concern is the geometric diagnosis: the filter bubble is not a bug in the recommender. It is the intended behavior of a system whose objective does not penalize it.

26.4 What the Geometric Framework Prescribes

We now sketch, prescriptively, what a recommender built on the aesthetic-manifold framework of this book would look like. We emphasize: this is speculative engineering. The components exist (manifold learning, diversity regularization, multi-objective ranking have each been studied), but the integration we propose has not been built and evaluated at scale.

Prescription 1: Multi-channel targeting, not scalar scoring. Instead of a single score $s(u, w)$, compute a vector of channel-specific scores $(s_{\text{spec}}, s_{\text{coh}}, s_{\text{traj}}, s_{\text{genre}}, s_{\text{nov}})$ corresponding to spectral-divergence match, coherence match, trajectory-geometry match, genre affinity, and novelty. The final ranking is a multi-objective optimization over these, with channel weights $\lambda_k$ that may be user-controlled, editorially shaped, or learned per-user-per-session. The point is that the channel decomposition remains in the pipeline. When the system recommends, it can say, at least internally: this is a trajectory-geometry match, this is a genre match, this is a novelty probe. The system’s explanation budget grows, and its failure modes become diagnosable.

Prescription 2: Diversity-on-the-manifold as a first-class regularizer. Current recommenders that claim “diversity” typically implement it as maximal marginal relevance (Carbonell & Goldstein 1998): rank candidates by $s(u, w) - \lambda \max_{w' \in \text{shown}} \langle \phi(w), \phi(w') \rangle$. This is diversity in embedding space. Our prescription replaces it with diversity on the aesthetic manifold — specifically, diversity on the channels of Chapters 20–21 that have been shown to carry cross-modality-invariant aesthetic signal (Chapter 12). Concretely: penalize recommendations whose trajectory-geometry features cluster tightly, even if their genre and surface embeddings differ. And — this is the novel move — penalize recommendations whose genre channel is too narrow even when the trajectory-geometry channel is diverse, because the framework predicts that genre coverage carries information beyond trajectory geometry.

Prescription 3: Geodesic exploration. The $\epsilon$-greedy random exploration of typical bandit-based recommenders is manifold-blind. A geodesic-exploration scheme would instead: identify the user’s current manifold location $u_t$; identify a target region $T$ distant from $u_t$ on some chosen channel (e.g., “classical music,” which has $R = -0.013$ in our data and is therefore in a distinct engagement basin); compute a geodesic path from $u_t$ to $T$ on the aesthetic manifold; and recommend works sampled along that path, not at its endpoint. The user is not abruptly jumped from techno to Schoenberg — they are walked there through works whose aesthetic distance from their current position grows gradually. This is the algorithmic analog of what a good DJ or a good bookseller does.

Prescription 4: Time-indexed personalization with corpus priors. The personalization term in the objective should be time-indexed: $u_t$ decays toward a corpus prior $u_0$ at a configurable rate. Users who do not interact with the system for a time drift back toward the general taste distribution; the filter bubble is leaky by construction. The decay rate can be user-configurable (a “reset my recommendations” slider) or editorially set (higher decay for news, lower for music preference).

Prescription 5: Evaluate on manifold-coverage, not just CTR. The most consequential prescription is evaluative. A recommender should be evaluated not only on click-through or watch-time but on the manifold coverage of its users’ consumption over time. A natural metric: the entropy, or equivalently the effective support size, of the distribution of user positions on the learned aesthetic manifold after $T$ interaction cycles. A system that concentrates its users (low entropy, small effective support) should be dispreferred to one that spreads them, even at some scalar-engagement cost.

We do not claim these prescriptions would yield higher short-term revenue. They may well yield lower. The framework’s argument is not a business argument. It is a claim about what recommendation ought to do, given what aesthetic judgment is. If aesthetic judgment is multi-channel, modality-sensitive, cross-lingually invariant at the structural level (Chapter 12, $\rho \approx 0.71$ across 6 language families), and confounded with coarse category at the 85–91% level (Chapters 20–21), then a recommender that optimizes a scalar engagement proxy is recommending categories, not works. It is doing something, but it is not doing curation.

26.5 Connection to Ethics Chapter 26

In Geometric Ethics Chapter 26, the parallel volume of this series, we analyzed AI systems whose scalar objectives produced moral misalignment. The alignment gap — the distance between the AI’s scalar-optimal geodesic and the human’s manifold-optimal geodesic — was shown to be nonzero on at least one moral dimension for any scalar reward function (Ethics Theorem 26.1–26.2). The argument in this chapter is the aesthetic cousin. An aesthetic reward function $s: \mathcal{M}_{\text{aes}} \to \mathbb{R}$ is non-injective: there exist aesthetically distinct works mapping to the same predicted engagement. The recommender, optimizing $s$, cannot distinguish them. The aesthetic information destroyed is, by Brouwer’s invariance of dimension, mathematically irrecoverable from $s$ alone.

The moral case is the aesthetic case is the same case. Scalar compression of high-dimensional judgment is an information-theoretic failure before it is a political or sociological one. The reason curation and ethics are both hard, and both currently done badly by optimization systems, is the same reason: real judgment has rank greater than one.

26.6 Honest Caveats and What Would Refute This Chapter

The prescriptive argument above rests on three load-bearing empirical claims, each with caveats we have stated and now restate.

First, the genre-confound claim (Chapters 20–21). In books, 85% of headline predictive $R^2$ is genre confound; in music, 91%. These numbers are point estimates from two specific corpora (Gutenberg↔Goodreads, FMA-medium), two specific encoders (LaBSE, MERT-v1-330M), and two specific CV schemes. Other corpora — commercial music catalogs, streaming-platform books catalogs, film — may show different fractions. If the genre confound in an actual streaming corpus were 20%, not 90%, the prescriptive argument of this chapter would weaken (though not collapse: even 20% is not zero).

Second, the cross-modality sign-flip claim (Chapter 17 and Chapter 21). Books and music respond in opposite directions to pair_sim_mean and step_mean, at $p < 10^{-28}$ robustness. This is the strongest evidence that aesthetic signal is modality-specific, which in turn is the strongest evidence that a single scalar recommender is necessarily lossy. Replication in film and visual art, which we do not yet have, would strengthen or weaken the generalization.

Third, the claim that geodesic-exploration would outperform $\epsilon$-greedy. This is conjectural. It is consistent with the framework. It has not been tested in production. An experiment with a negative result — geodesic exploration yielding no improvement in long-run user manifold coverage over random exploration — would falsify the specific prescription without falsifying the broader framework (exploration on some manifold is still better than exploration in surface embedding, but the choice of manifold would be empirically undetermined).

The largest gap is the one we have already flagged: no end-to-end evaluation exists. Chapter 17’s findings concern predictive power, not recommender behavior. Translating “geometric features carry aesthetic signal in held-out prediction” to “a recommender using geometric features produces better long-run user outcomes” requires a production experiment with live users. That experiment is the most important empirical gap this book identifies.

26.7 Bridge to Chapter 27

We have argued that AI curation, done with scalar engagement objectives, is a specific and diagnosable pathology — greedy ascent on the wrong manifold. We have sketched what a geometric alternative would look like and flagged what an evaluation would need. In Chapter 27 we turn to another domain in which aesthetic trajectories drift in time: fashion and product design. There, the “recommendation system” is called the market, and the filter bubbles have names like capsule collection and brand aesthetic. The same geometric vocabulary applies, with different dynamics.