Do Rationalists Dream of Electric Art?

Abstract

This paper argues that as the marginal cost of cognitive production falls, scarcity relocates from production to legitimation. Art makes that shift visible because its value depends less on output scarcity than on institutions of judgment, public interpretation, and social ratification. The question is therefore not simply whether alignment-adjacent communities discuss art, but what kind of value model becomes legible when art enters view.

In a hand-verified corpus of 6,279 posts from September 2022 through March 2026, art appears in only 34 posts (0.54%), with the sharpest scarcity in the technical core. When art does appear, literature and terminal framing dominate, while outright dismissal is rare. A 50,378-post retrieval substrate and a historical arts-critical comparison corpus for 2023-2025 (15,448 documents; 34,509 art-only passages) then extend the project from a scarcity note to a comparative discourse-legitimation paper. Current comparative results support one scalable claim and one bounded one. AI relation is the strongest surviving axis: alignment-adjacent passages more often treat AI as constitutive medium or practice, while arts-critical passages more often frame AI as tool, opportunity, labor dispute, or governance problem. Differences in object of valuation also appear substantive, but remain review-gated and are used directionally rather than as flat headline percentages. The broader claim is that alignment-adjacent and arts-critical discourse are rival regimes of judgment under conditions of cheapening intelligence.

Introduction

Generative AI has made one historical shift unusually easy to see. As the marginal price of many cognitive outputs falls, scarcity does not disappear. It relocates. The bottleneck moves away from making more text, images, music, or design variants and toward deciding what counts, what circulates, what gets institutionalized, and what remains publicly legible as worth preserving. The central cultural question of AI is therefore not only what can be generated, but what remains legible as valuable once generation becomes cheaper.

Art is the right stress test for that shift because its value is least reducible to productivity or coordination efficiency. Artistic labor does not disappear when models get better, but the social conditions under which art is recognized, filtered, funded, and remembered change sharply. A regime that can talk fluently about AI safety, model capabilities, and governance while only weakly rendering art legible is not merely neglecting one topic among many. It is revealing something about what it thinks culture is for.

In Which Art Endures?, we measured the institutional side of this problem: how scarce endorsement, relationship capital, and network position shape the allocation of arts funding. DRDEA asks the discourse-side analogue. How do communities most actively shaping AI strategy, governance, and long-run futures make art legible as valuable? The initial answer was a scarcity finding. The stronger answer is sociohistorical: alignment-adjacent discourse and arts-critical discourse encode rival regimes of judgment under cheapening intelligence.

Cheap Intelligence and the Return of Legitimation

The strongest version of the argument is not that production scarcity vanishes. It is that production scarcity becomes a weaker explanation for value in some domains, while selection scarcity becomes a stronger one. When text, image, and music generation get cheaper, value shifts toward infrastructure, ownership, distribution, filtering, and the institutions that convert raw abundance into durable public judgment.

Art makes this mechanism unusually visible because art has never depended only on the cost of making another object. It has always depended on thicker social processes: criticism, curation, museums, magazines, collectors, publics, pedagogy, and the ability to distinguish what is merely produced from what is worth keeping. In that sense, art is not a sentimental exception to AI economics. It is the clearest domain in which legitimation remains scarce even when output becomes abundant. Calling art a stress test does not reduce it to an external proxy; it recognizes that arts institutions have long managed the gap between abundant production and scarce public ratification.

DRDEA is therefore a paper about discourse legitimation under relocated scarcity. Its claim is not that art enjoys mystical exemption from economics, or that artistic labor disappears. The claim is narrower and stronger: as the marginal price of many cognitive outputs falls, discourse and institutions become more decisive filters of value. Art is the domain where that change is easiest to diagnose because it concentrates questions of taste, labor, memory, publics, and legitimacy.

Research Object

The object of study is no longer just art frequency in alignment discourse. It is discourse legitimation: the set of frames, authorities, and evaluative assumptions through which art becomes intelligible inside alignment-adjacent communities. In practical terms, the paper asks four linked questions.

1. How rarely does art appear in alignment-adjacent discourse, and where is that rarity most acute?
2. When art appears, which art forms and value frames become legible?
3. How much larger is the retrievable substrate once the corpus is expanded beyond the initial hand-audited window?
4. How should a comparative analysis be structured if the goal is to compare alignment-adjacent discourse to arts-critical public discourse without collapsing into generic sentiment analysis?

The comparative object is therefore not “do these corpora feel different?” It is whether alignment-adjacent and arts-critical discourse instantiate different evaluative regimes: different answers to what art is for, what counts as value, where judgment sits, and how AI enters the scene. That is the bridge to WAE. WAE studies legitimation in allocation. DRDEA studies legitimation in discourse.

Two Discourse Regimes

The corpora in this paper are not just two piles of text. They are historically formed institutions of judgment. Alignment-adjacent discourse is a digitally native, technocratic field shaped by engineering, analytic philosophy, economics, forecasting, and internet rationalism. Its dominant problems are risk, optimization, coordination, governance, and intervention design. Arts-critical discourse is institutionally layered and historically thick. Its dominant problems are form, labor, publics, institutions, memory, and contested judgment. Treating these as rival regimes of judgment does more explanatory work than treating them as style clusters.

That difference is not merely stylistic. It reflects divergent institutional memories and different class positions inside the current AI transition. Alignment-adjacent venues are optimized to reason about systems. Arts-critical venues are optimized to reason about cultural objects, artistic labor, institutions, and publics. The comparison is therefore sociological before it is lexical.

The arts regime in this corpus is also more historically inhabited than a generic “arts blogs” label would suggest. It includes museum writing about permanent collections and exhibition architecture, criticism about biennial selection and canon formation, trade reporting on galleries and auctions, and news coverage of war, protest, repatriation, and heritage law. These outlet families inherit postwar museum expansion, magazine criticism, donor-board governance, biennial circuits, and current fights over restitution, copyright, and cultural policy. In the current build, arts discourse is routinely about how artworks are acquired, housed, defended, translated, and historicized. That is why it makes sense as a comparison set for legitimation rather than as a simple control group of people who like art more.

Evidence Tiers

The revised paper uses three evidence tiers. Each tier supports different claims. The numbers should not be pooled as if they describe the same empirical object.

Method

Corpus A: Alignment-Adjacent Discourse

The verified preliminary result is based on 6,279 posts from five alignment-adjacent venues: the Alignment Forum, LessWrong, the Effective Altruism Forum, 80,000 Hours, and Astral Codex Ten. The current hand-audited window runs from September 2022 through March 2026. For descriptive purposes, the corpus is split into three discourse regimes by proximity to technical alignment research: core alignment (1,981 posts), rationalist public sphere (1,813 posts), and EA / institutional discourse (2,485 posts).

Retrieval used contrastive embedding similarity rather than keywords. Each post was embedded with BGE-small-en-v1.5 and compared against art anchors and non-art anchors; a post was flagged when its maximum art-anchor similarity exceeded its maximum non-art-anchor similarity by at least 0.02.¹ This design is necessary because art appears in these corpora indirectly, analogically, and often without stable lexical markers. Semantic retrieval is therefore more appropriate than keyword search, even though the initial threshold was intentionally recall-heavy and precision-light.²

All 169 candidates in the preliminary run were hand-reviewed. That audit produced the 34-post verified art corpus used for the current frequency, framing, and art-form results. It also exposed the main failure mode of the initial classifier: philanthropy and donation posts in the EA / institutional regime often looked close to “arts funding” anchors without actually being about art. The gap between the initial candidate rate (2.69%) and the verified rate (0.54%) is therefore methodologically important in its own right. Tier A remains the only promoted prevalence claim.

Tier B: Expanded Retrieval Substrate

Since the initial verified run, the corpus has been expanded to 50,378 posts. The full embedding pass yields 1,616 candidate documents and 2,334 candidate-filtered passages. This larger substrate changes the paper architecturally rather than epidemiologically. It shows that DRDEA is no longer a small-window retrieval exercise and can sustain passage-level comparison, but it does not replace the verified prevalence estimate because the full pass has not been promoted through a second manual audit.

Corpus B: Arts-Critical Public Discourse

The comparative module uses a separate arts-side corpus built by outlet family rather than by traffic rank. The current historical build includes six outlets across five families: Hyperallergic (contemporary art news), Artforum and Art in America (criticism magazines), ARTnews (trade / market publication), Guggenheim (museum / institutional journal), and BOMB Magazine (independent review / essay). As of March 20, 2026, this backfilled `2023-01-01` through `2025-12-31` build contains 15,448 normalized arts documents and 34,509 art-only passages.

This is no longer merely an available corpus. It is a time-aligned historical comparison set for the planned V0 regime map. But some families remain thin: the museum / institutional and independent review / essay families are each represented by a single outlet, so the arts side should still be treated as a strong but incomplete slice of arts-critical public discourse rather than the whole field.

Comparative Module: DRDEA_CVG_V0

The comparative layer, DRDEA_CVG_V0, shifts the unit of analysis from document to passage because full essays mix summary, reportage, quotation, and evaluation. Passage-level windows are the right unit for identifying valuation grammar. The frozen codebook defines six axes, but this draft promotes only those that survive the current reliability gates. In practice that means one robust main-text axis (`ai_relation`) and one review-gated supporting axis (`object_of_valuation`). The full proxy-pair, seeded-scale, and low-margin review workflow remains in the appendix because it matters for interpretation, not for the core narrative.

Verified Preliminary Findings

Scarcity Is Real, But It Is Not the Whole Story

The verified preliminary rate is low: 34 art-relevant posts out of 6,279, or 0.54%. That result survives the precision audit and remains the paper’s empirical floor. But the more revealing pattern is the distribution of that scarcity. Art is almost absent from the technical core and far more visible in the broader rationalist layer.

The key descriptive fact is therefore not just “low frequency.” It is a 26x gap between core alignment and the rationalist public sphere. The places where AI is most explicitly theorized as a technical governance problem almost never discuss art at all. The places where adjacent communities reflect on human life more broadly are where art becomes intermittently legible.

Just as important, scarcity is not equivalent to contempt. The hand-audited corpus does not show a community repeatedly denigrating art. It shows a community that mostly routes attention elsewhere, then engages art through a small number of recurrent frames when it does appear.

What Art Becomes Legible

The verified corpus is not neutral across art forms. Literature and fiction dominate, followed by general art discourse and visual art. Digital / generative art is present but still marginal relative to how central it has become to public AI debate.

That distribution is substantively revealing. A writing-native discourse environment finds literature easiest to discuss. It is far less comfortable with forms that require institutional mediation, embodied spectatorship, or material encounter. The result is not a broad cultural field but a selective legibility surface in which forms already native to textual online communities are easier to value and easier to analyze.

The modest share for digital / generative art is especially important. The domain most obviously transformed by new AI systems is not absent, but it is not yet the dominant site through which these communities think about art either. Even in a corpus adjacent to AI discourse, generative art remains one small cluster inside a broader pattern of selective attention.

When Art Appears, It Is Usually Not Dismissed

The strongest conceptual result from the verified corpus is the framing distribution. This is where the project stops being a simple frequency note. The question becomes: what mode of value attribution is actually present when art enters alignment-adjacent discourse?

Terminal framing accounts for 17 of 34 posts, or 50% of the verified corpus. Instrumental framing accounts for 6 posts (18%). Labor / economic, threat / displacement, and governance / legitimation each account for 3 posts (9%), while outright dismissal and metaphorical use appear only once each. The stereotype that rationalist or EA-adjacent discourse simply treats art as frivolous is therefore not supported by the current hand-audited sample.

The sharper point is that the community does not need to reject art in order to reveal a limited discourse of value. Scarcity plus selective framing is enough. Terminal art engagement exists, but it is peripheral. Governance / legitimation appears at all, which is important for the bridge to WAE, but only as a thin slice. The paper’s claim is therefore not that alignment-adjacent communities hate art. It is that art is weakly and selectively legible inside the value model those communities publicly articulate.

Representative high-scoring examples still matter qualitatively, but the ranked retrieval table now appears in the appendix rather than the body. The main text stays focused on verified scarcity, selective framing, and only those comparative results that survive the current reliability gates.

Comparative Findings

The major upgrade in this draft is the comparative valuation grammar module, DRDEA_CVG_V0. Its purpose is not to ask whether alignment-adjacent and arts-critical writing have different emotional tone. Its purpose is to compare the grammar of valuation itself: what is valued, why, by whom, and in relation to AI. The body of the paper promotes only the comparative claims that survive the current reliability gates. Detailed proxy-pair, scaling, and low-margin review diagnostics appear in the appendix note below.

The comparative module now supports a narrower claim set than the original ontology envisioned. The historical arts corpus, passage shell, and balanced pilot are enough to show that cross-regime comparison is real and no longer merely aspirational. But at present only two axes survive strongly enough to matter in the body: AI relation and, more cautiously, object of valuation. Those axes are useful because they bear directly on the sociohistorical mechanism above. They show where each regime locates value and what each regime treats AI as doing inside culture.

AI Relation Is the Strongest Comparative Axis

The clearest surviving comparative difference concerns how AI enters the scene. In both regimes, most passages are still AI-absent. But when AI is present, alignment-adjacent writing is more likely to treat AI as constitutive medium or practice, while arts-critical writing more often frames AI as tool, opportunity, or governance problem. In the seeded full-shell map, alignment-adjacent passages over-index on constitutive-medium usage (6.0% vs 0.4%), while arts-critical passages over-index on tool / opportunity framing (10.1% vs 4.0%) and governance-problem framing (5.2% vs 2.0%). That pattern survives threshold sensitivity, which is why AI relation is the only comparative axis promoted here as broadly scalable.

Object of Valuation Differs, But Only Directionally

The second surviving comparative axis concerns what the discourse takes to be the thing under evaluation. Here the paper can make a substantive directional claim, but not a flat numerical one. Alignment-adjacent passages repeatedly compress ambiguous cases toward an artwork / technology boundary. Arts-critical passages distribute ambiguity more broadly across artwork, institution, and artist. That is enough to say that the two regimes do not simply talk about art at different rates; they also center different objects when value is assigned.

What the current draft does not do is promote object-of-valuation percentages as stable headline results. Repeated low-margin review shows that this axis remains review-sensitive, especially in alignment-adjacent passages where the main ambiguity zone is still `artwork ↔ technology / tool`. On the arts-critical side, the recurring ambiguity zones remain `artwork ↔ institution` and `artist ↔ artwork`. In the body of the paper, object of valuation therefore supports interpretation rather than precise regime totals.

Paired Readings: What Each Regime Notices

Quantitative differences become clearer when read back into concrete passages. The point of the paired readings below is not to elevate a few texts into definitive proof. It is to show how the comparative axes sound in practice when each regime confronts the same historical pressure: AI makes new forms of cultural production abundant, but leaves judgment unsettled.

Discussion

The paper’s strongest conclusion is now sociohistorical rather than merely descriptive. The problem is not simply that alignment-adjacent communities discuss art infrequently. The problem is that under conditions of cheapening intelligence, they render art legible through a thinner and differently situated evaluative surface than arts-critical discourse.

That matters because AI governance discourse does not merely optimize policy variables. It also carries assumptions about which domains deserve preservation, public reasoning, and institutional care. A regime that rarely encounters art, and that most easily recognizes the forms already native to text-heavy online communities, may remain internally coherent while still narrowing the cultural objects it can treat as publicly consequential.

The comparative valuation grammar module sharpens that concern in a narrower but more defensible way than the original ontology promised. The cleanest current comparative difference is how AI enters the frame. The second is what the discourse takes to be the thing being valued. Those are not trivial additions to a scarcity paper. They move DRDEA from “art is rare here” to a stronger claim: alignment-adjacent and arts-critical discourse are rival elite regimes of judgment under relocated scarcity.

Counterarguments

Three objections deserve to be answered directly.

First: perhaps these differences are trivial because the audiences are different. That objection is partly right and does not damage the paper. The claim is not that one regime has failed to approximate the other. The claim is that these are historically distinct institutions of judgment, and that their differences matter because both now participate in shaping how AI and culture are publicly understood.

Second: perhaps scarcity alone only reflects topic mix, not a theory of value. That would be a valid criticism of the paper’s first version. It is why the current draft does not stop at corpus frequency and why the comparative module works at the level of art-relevant passages rather than whole corpora. The surviving comparative axes and paired readings matter precisely because they show that when art does appear, it is not simply slotted into the same evaluative machinery on both sides.

Third: perhaps the arts corpus is too partial to bear a regime-level argument. This is the strongest empirical objection, and the paper accepts its force. The arts side is a bounded five-family slice, not the whole arts-critical public sphere. That is why the paper promotes only `ai_relation` as a scalable comparative axis, keeps `object_of_valuation` directional, and leaves the deeper ontology in the appendix.

Fourth: perhaps the paired readings are cherry-picked. Of course they are curated; that is what close reading is. But they are not being used as prevalence estimates or as substitutes for the bounded comparative map. Their job is narrower: to make the surviving axes audible in prose. If `ai_relation` and the directional object-of-valuation contrast disappeared, these pairs would lose evidentiary force as well.

The important negative result is equally useful. The paper does not yet know enough to promote stable claims about evaluative mode or source of authority. Those remain the deeper theoretical prize, but they should stay in the appendix and methods frame until reliability improves. That is a strength of the current draft, not a weakness. It means the paper is organized around admissible claims rather than around the largest ontology it could imagine.

The broader implication is not that artistic or intellectual labor disappears. It is that as the marginal price of many outputs falls, value shifts toward ownership, filtering, infrastructure, rights, and public judgment. Art is diagnostic because it concentrates those pressures in an unusually visible form. What each regime can recognize as worth preserving becomes an indicator of the kind of human future it is prepared to optimize for.

Limitations

Four limitations should remain explicit.

1. Small verified corpus. The current publication-honest result still rests on 34 hand-verified art posts. Frequency, art-form, and framing percentages are therefore descriptive and directional rather than statistically settled.
2. Expanded retrieval is not audited prevalence. The 50,378-post substrate and 1,616 candidate documents materially improve scale, but they have not yet been promoted through a second manual prevalence audit. Tier B is a retrieval surface, not a new verified headline rate.
3. The arts corpus is historical but still thin in two families. The current arts-side build is backfilled for `2023-2025` across the planned five families, but museum / institutional and independent review are each represented by a single outlet.
4. The comparative module is not independently double-coded. The strongest current comparative artifact is a solo-adjudicated proxy pilot plus a seeded model-based full-shell map on two axes. That is enough for bounded working-paper notes and internal scaling, but not for publication-grade regime-difference claims.

Inside the current solo workflow, the practical next step is no longer to widen the ontology. It is to consolidate the paper around the axes that survived, keep object-of-valuation claims review-gated, and treat independent double-coding as the main future strengthening step if stronger public comparative claims are needed. Outlet depth within the museum / institutional and independent-review families also still needs to improve before any pooled cross-regime claim should be treated as field-level. The sociohistorical argument should therefore be read as a disciplined mechanism claim, not as a license to overstate the settledness of the comparative coding.

Appendix Note: Comparative Reliability and Scaling

This appendix note records the comparative workflow status that should inform interpretation but not carry the main argumentative burden. The comparative module is real, historical, and operational. But not every axis is equally usable, and the main text has been cut back accordingly.

The historical pilot contains 207 balanced passages. An initial heuristic dual pass retained evaluative mode, object of valuation, source of authority, and AI relation above the working threshold. A stricter reviewed-vs-skeptical proxy pair narrowed the stable set to object of valuation (κ=0.658) and AI relation (κ=1.000). Because the remaining disagreements were one-sided blanks rather than contradictory non-empty labels, a solo-resolved version of the pilot could be constructed for internal comparative work.

Using that solo-resolved pilot as a seed set, I scaled object of valuation and AI relation across the full 36,843-row historical shell. Leave-one-out validation on the seed set is strong for object of valuation (accuracy 0.915; macro-F1 0.889) and strong enough for internal mapping on AI relation (accuracy 0.957; macro-F1 0.777), with the remaining weakness concentrated in a rare analogy-only class.

Sensitivity analysis then separated the two axes. AI relation is comparatively stable under confidence filtering. Object of valuation is not: at `margin >= 0.05`, only about half of alignment-adjacent rows survive, and the object mix shifts materially. Two bounded solo reviews of the lowest-margin object-of-valuation rows reinforce the same conclusion. Across 128 reviewed low-margin rows, 77 changed on review (60.2%). The recurring alignment-side ambiguity is `artwork ↔ technology / tool`; the recurring arts-side ambiguities are `artwork ↔ institution` and `artist ↔ artwork`.

Appendix Table A2: Highest-Scoring Art-Relevant Posts

The ranked retrieval table is retained here as a qualitative check on what the contrastive embedder surfaced most strongly. It is informative for retrieval sanity, but it is not one of the paper’s headline results.

Connection to Which Art Endures?

WAE and DRDEA now form a cleaner pair. WAE measures legitimation in allocation: which organizations receive money, continuity, and network position. DRDEA measures legitimation in discourse: which artistic domains become visible enough to be reasoned about at all, and through which value grammars.

There is a suggestive empirical overlap between the two. In the verified preliminary corpus, discourse attention has a moderate positive rank correlation with WAE funding shares across the five admissible domains, though the comparison is small-n and excludes digital / generative art because its institutional funding share is effectively zero.³ The main asymmetry is that literature is overrepresented in discourse relative to funding, while heavily funded performing and visual arts are relatively underdiscussed.

The stronger bridge is conceptual. Both projects study scarcity under conditions of relative abundance. WAE asks how scarce endorsement and relationship capital structure the flow of money once artistic production is abundant. DRDEA asks how scarce attention and evaluative legibility structure the flow of discourse once cultural output becomes more abundant and cognitive production becomes cheaper. Taken together, they describe two linked layers of the same legitimation economy: allocation legitimation and discourse legitimation.

---