Art & Culture Law · 2026-04-23

🎨 Art & Culture Law — 2026-04-23

🏛️ Japan's LDP Advances Criminal Penalties for Malicious AI Image Operators
🖼️ Artsy AI Survey 2026: 300 Galleries Confirm AI in Operations, Contest It as Artistic Medium
🎵 UCL Music-AI Roundtable Exposes Licensing Collapse Between Creator Rights and Training Pipelines
💻 Anthropic Claude Code Leak Makes AI Training Consent Question Visible at Source-Code Level
🎬 Variety Entertainment Summit: Hollywood's Authenticity Premium Fractures Under AI Production Economics
✊ Nordic Rights Holders' Open Letter Escalates EU AI-Copyright Negotiations on Art. 53 Opt-Out

---

🏛️ Japan's LDP Advances Criminal Penalties for Malicious AI Image Operators

Japan's Liberal Democratic Party is advancing Diet legislation to establish criminal penalties for AI platform operators who facilitate the malicious generation or distribution of images — a structural shift from post-hoc civil liability to upstream operator accountability that marks a significant new vector in cultural AI governance. The Asahi Shimbun reports the LDP's digital society promotion division circulated a draft framework this week covering three categories: non-consensual intimate AI-generated imagery, deepfakes deployed in harassment campaigns, and synthetic media used in election interference.

The architecture of the bill is deliberately distinct from Japan's existing digital harm statutes. Current law holds distributors liable for content damage after it occurs but carries no criminal exposure for the operators of platforms enabling mass AI generation. The LDP framework shifts culpability upstream, making operators legally exposed when their systems can be demonstrated to have enabled malicious generation at scale. This borrows from the EU AI Act's risk-tiering logic while using criminal law rather than regulatory compliance as the enforcement instrument — a meaningful escalation in legal stakes for AI service operators globally.

The cultural specificity of Japan's exposure shapes the bill's urgency. Anime, manga, and video game aesthetics have become primary targets for AI training without creator consent, and Japan's Agency for Cultural Affairs issued guidance in 2023 that some AI developers interpreted as relatively permissive on training data access. A criminal liability regime for operators resets the risk calculus entirely: training on unlicensed cultural material now carries exposure not just to civil copyright suits but to potential prosecution if deployed models are subsequently weaponized.

Rights advocacy organizations, including JASRAC (Japan's music rights collective) and the Japanese Animation Creators Association, have publicly called for the bill to extend to training data consent — arguing the malicious-use framing leaves the upstream extraction problem unaddressed. The LDP draft contains no training data provisions, mirroring the structural compromise visible in the UK government's earlier reversal on AI copyright. The bill is expected to enter Diet deliberations by June.

The significance here is jurisdictional: criminal law is entering the AI-and-culture toolkit globally. Combined with the EU AI Act's cultural heritage provisions and US state-level deepfake statutes, the pressure on AI operators to build provenance verification and consent frameworks is now compounding from multiple directions simultaneously — with Japan representing the first major economy to attach criminal penalties to operator liability rather than limiting exposure to civil damages.

Sources:

---

🖼️ Artsy AI Survey 2026: 300 Galleries Confirm AI in Operations, Contest It as Artistic Medium

The 2026 Artsy AI Survey of more than 300 gallery professionals — analyzed this week in a legal risk brief by Holland & Knight — exposes a structural paradox driving legal and market risk across the art sector simultaneously: AI is embedded in gallery operations (communications, inventory, exhibition planning), while failing to achieve legitimacy as an artistic medium in the eyes of galleries, collectors, or market professionals. Only 15 percent of galleries report fielding collector inquiries about AI art, and those who ask rarely purchase.

The governance gap the survey reveals is definitional rather than aesthetic. No consensus definition of "AI art" exists across the 300+ institutions surveyed, with galleries applying frameworks ranging from "fully prompt-based" to broader "AI-assisted" categories. In markets where valuation depends on authorship, provenance, and originality, the absence of a shared vocabulary is a direct liability exposure: misrepresentation claims attach wherever AI contribution is material to valuation but undisclosed or inconsistently categorized.

The legal backdrop is now settled at the US appellate level. Holland & Knight cites Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025) as establishing that purely AI-generated works are not copyrightable absent meaningful human authorship. Gallery professionals operating under the assumption that AI-generated works in their collections carry full IP protection are therefore exposed to misrepresentation claims downstream. The Holland & Knight analysis directs practitioners to the Zarya of the Dawn Copyright Office letter as operational precedent: human authorship of composition and selection can be registered, but AI-generated image elements cannot and must be disclaimed.

What the survey reveals beneath the headline ambivalence is a market in the early stages of photography-like absorption — AI tools are becoming infrastructural to gallery operations even as AI outputs remain culturally contested. The analogy to photography's entry into the art market in the 1880s (technically capable, commercially present, artistically contested for decades) is structurally apt, and the survey's authors predict a similar long-arc institutional absorption.

The present-tense risk concentration, however, is operational. Training AI models on copyrighted artworks without consent remains active litigation terrain. Third-party AI platforms galleries use for administrative tasks may be ingesting client images and proprietary data, creating confidentiality and trade secret exposures that most gallery legal teams have not yet systematically mapped. The survey's primary practical warning is not about AI as a medium — it is about the undisclosed legal footprint of AI-as-tool in everyday gallery operations.

Sources:

---

🎵 UCL Music-AI Roundtable Exposes Licensing Collapse Between Creator Rights and Training Pipelines

A roundtable convened at the UCL Institute of Brand and Innovation Law this week — summarized in an IP law roundtable report — produced a detailed taxonomy of the structural collapse between music creator rights frameworks and AI music generation pipelines, with participants concluding that existing copyright doctrine cannot close the gap without legislative intervention.

The core finding: music copyright law was built on the assumption that the rights to perform, reproduce, and learn from are separable — an assumption that holds for human musicians but breaks under generative AI training. A model trained on a corpus of copyrighted recordings does not "reproduce" in the traditional legal sense, yet generates outputs that systematically capture stylistic signatures of training data. UCL participants noted that ongoing major label litigation against AI music platforms is testing whether substantial similarity doctrine extends to style capture — a question courts have historically answered inconsistently.

The roundtable proposed distinguishing three legally distinct use categories: direct reproduction (clear infringement), style emulation (contested, requires statutory clarity), and genre-level training (likely non-infringing but requiring transparency disclosure). This taxonomy challenges the music industry's current all-or-nothing litigation strategy, which treats any AI music generation as presumptively infringing and demands blanket licensing. Participants noted that settlement negotiations in several major cases have begun incorporating prospective licensing terms alongside backward-looking damages — suggesting both sides recognize durable frameworks are needed.

Google's exposure in the Lyria 3 litigation — where independent artists claim the model was trained on YouTube-hosted tracks without consent — illustrates the central ambiguity. The case hinges on whether YouTube's Terms of Service granted Google training rights over uploaded music, a question with implications for every platform-model combination in the creative economy. If ToS constitutes a consent mechanism, every platform's user agreement becomes a de facto training data license — a result that EFF and creator advocacy groups argue was not contemplated by creators at upload time.

The structural gap the UCL roundtable exposes is jurisdictional: music licensing operates under the Berne Convention framework, but AI training pipelines are borderless. A model trained under US fair use doctrine deploys globally, undercutting licensing regimes in territories like the UK and EU that do not recognize US fair use. The roundtable explicitly flagged this as a problem that bilateral or multilateral policy coordination — not just litigation — must address.

Sources:

---

💻 Anthropic Claude Code Leak Makes AI Training Consent Question Visible at Source-Code Level

Leaked source code from Anthropic's Claude Code assistant, first reported by The New York Times on April 22, has made publicly concrete what has long been theorized in legal circles: AI development tools are themselves built on training data consisting substantially of copyrighted software, and internal guidance at major AI companies has routinely treated public availability as implying training consent — a legal theory that has not been validated by courts.

The leaked materials reveal training pipeline artifacts showing direct ingestion of publicly accessible GitHub repositories, including projects licensed under GPL, MIT, and Apache licenses, alongside large quantities of repositories carrying no license specification. The copyright implications bifurcate sharply. First, software licensed under the GPL carries copyleft obligations: derivative works must be licensed under GPL. If a code generation model trained on GPL code produces outputs constituting "derivative works," every commercial output from that model may carry GPL entanglements. The GitHub Copilot litigation, now in its fourth year, has not yet resolved the derivative works question for generated code outputs.

Second, the leak surfaces the consent assumption with unusual directness. GitHub's Terms of Service permit public repository visibility but contain no explicit grant of training rights to third parties. Repositories uploaded without a license file are, under default copyright law, fully copyrighted — not public domain. The Times reports that Anthropic's internal guidance acknowledged this ambiguity and treated unlicensed-but-public repositories as "trainable" under a contested interpretation of implied permission from public accessibility.

The political valence of the leak is notable precisely because Anthropic's Constitutional AI framework and public commitment to responsible development collides with training pipeline decisions that, if accurately depicted, prioritize data scale over rights clarity. The gap between rhetorical responsibility and operational practice — visible here in source code — is the same gap the UCL music roundtable and the Artsy gallery survey identified in their domains: institutions integrating AI faster than their consent governance frameworks can track.

For cultural production, the implications extend beyond software. If AI code generation tools are trained on copyrighted material under contested fair use theories, the identical logic applies to music generation, image synthesis, and text models. The Anthropic leak functions as a bellwether disclosure: the training data consent question, long opaque, is now visible in internal documentation. How the legal and regulatory sectors respond will establish whether public availability is treated as implied consent — or whether the right to be excluded from AI training requires active assertion across every creative domain.

Sources:

---

🎬 Variety Entertainment Summit: Hollywood's Authenticity Premium Fractures Under AI Production Economics

Variety's Entertainment Marketing Summit, held in Los Angeles on April 23, produced an unusually candid assessment of a structural tension now running through the cultural industry's AI conversation: the authenticity premium — the economic value attached to human-made creative work — is being systematically undercut by production economics favoring AI-generated content, while the marketing apparatus designed to sell that content is being asked to construct authenticity signals for work that producers themselves can no longer reliably distinguish from human output.

Studio, agency, and streaming panelists converged on a shared diagnosis: SAG-AFTRA's AI contract provisions created a consent framework for performer likeness, but the frameworks do not address the ambient AI generation now permeating production workflows at the level of concept art, scoring templates, sound design, and post-production finishing. AI contributions in these layers are distributed, iterative, and often invisible to the human talent signing final approvals — creating a legal gray zone where no single element is AI-generated but the aggregate creative output is AI-shaped.

The marketing implications carry regulatory weight. FTC guidelines on endorsements and testimonials require disclosure of material connections, but no equivalent obligation governs AI-generated creative content in commercial campaigns. A campaign built substantially on AI-generated visuals with human talent signing off on final output occupies a legally ambiguous space: neither clearly required to be disclosed nor clearly prohibited from claiming human creative origin. This gap is increasingly exploitable and increasingly noticed.

Several summit panelists cited internal audience research showing that Gen Z viewers (18–24 cohort) have developed measurable sensitivity to AI-generated visual and audio artifacts that older demographics have not. The C2PA content provenance standard, now being piloted by Adobe and several major platforms, offers a technical disclosure mechanism that could address this gap — but adoption across production chains remains fragmented. The implication for brand value is concrete: in market segments where authenticity drives purchase decisions, deploying AI content without provenance disclosure creates credibility risks not yet reflected in industry pricing or legal frameworks.

For cultural production broadly, the summit articulates a coming structural absence: cultural law is built around categories of authorship, originality, and expression, but contains no vocabulary for the social value of human production as a market category. Copyright protects expression; it does not protect the authenticity signal that human origin creates in the marketplace. As AI production becomes economically dominant, the legal instruments protecting that signal must be constructed from scratch — through disclosure law, labeling requirements, or entirely new cultural rights frameworks.

Sources:

---

✊ Nordic Rights Holders' Open Letter Escalates EU AI-Copyright Negotiations on Art. 53 Opt-Out

Rights holder organizations across Scandinavia — representing collecting societies, publisher associations, and audiovisual rights holders from Denmark, Sweden, Norway, and Finland — filed a joint open letter with EU member state governments on April 17, requesting direct incorporation of their positions into active trilogue-follow-up negotiations on the EU AI Act's creative sector provisions. The move, reported by Danish law firm Bech-Bruun this week, marks an escalation from industry advocacy to legislative intervention at the state level, timed deliberately to the window before implementation guidance on the AI Act's text-mining exception crystallizes.

The letter targets a specific architectural vulnerability in the EU AI Act's text mining and data mining exception under Article 53. The exception permits AI developers to use lawfully accessed content for training purposes unless rights holders have explicitly opted out — a structure that, the letter argues, inverts the default consent model of copyright law. Berne Convention member states have historically required affirmative licensing; Article 53's opt-out architecture shifts the burden of exclusion onto rights holders rather than developers.

The practical effect of opt-out architecture is amplified by implementation asymmetry. Large AI developers have the technical and legal capacity to monitor structured opt-out signals through robots.txt extensions or licensing databases. Individual rights holders — composers, illustrators, authors, independent publishers — typically do not. The Nordic letter specifically calls for a publicly funded, interoperable opt-out registry infrastructure across EU member states, noting that existing voluntary licensing databases (CISAC's music rights infrastructure, CEPIC's visual arts registry) were designed for licensing revenue distribution, not machine-readable training exclusion signals.

The timing reflects a well-understood pattern in EU policymaking: implementation guidance on AI Act provisions carries quasi-legislative weight, and the window for shaping that guidance closes once the European Commission publishes its formal interpretive documents — expected over the summer of 2026. Rights holders are attempting to embed a consent-infrastructure requirement into the guidance before it is issued, rather than litigating against a finalized interpretation afterward.

For US-based AI developers operating in EU markets, the Nordic letter signals a coordinated enforcement posture: if opt-out infrastructure is not built to rights holders' stated standards, targeted litigation against specific developers is the stated next step. The structural parallel to GDPR is explicit in the letter: EU privacy law's opt-in default established that the burden of data collection falls on collectors, not subjects. Rights holders are proposing the same logic — the burden of obtaining training consent should fall on AI developers, not on the millions of creators whose work would otherwise be ingested by default.

Sources:

---

Research Papers

Copyright and Artificial Intelligence Part 3: Generative AI Training — U.S. Copyright Office (2024) — Third installment of the Copyright Office's AI policy series, addressing training data rights and the scope of existing statutory exceptions as applied to generative AI systems. Provides the administrative backdrop for both the art market survey and the Anthropic training pipeline analysis, particularly the Office's position that no general fair use safe harbor exists for AI training on copyrighted material.

Zarya of the Dawn: Human Authorship Registration Guidance — U.S. Copyright Office (2023) — Definitive Copyright Office letter establishing the operational standard for hybrid human-AI works: human authorship of selection, coordination, and arrangement is registrable; AI-generated elements must be disclaimed. The operative precedent for the Holland & Knight art market analysis and the gallery sector's misrepresentation risk exposure.

Content Credentials: C2PA Technical Specification v2.1 — Coalition for Content Provenance and Authenticity (2024) — Technical standard defining cryptographic content provenance assertions for media, enabling machine-readable attestation of AI vs. human creation at point of generation or capture. Directly addresses the authenticity-premium erosion identified at the Variety summit and the provenance gap the Japan LDP bill is attempting to close through criminal law.

WIPO Technology Trends: Artificial Intelligence and Copyright — World Intellectual Property Organization (2024) — Maps divergent national approaches to AI and copyright across 130+ member states, identifying the jurisdictional fragmentation the UCL roundtable characterized as structurally unresolvable through unilateral litigation. Provides the global governance backdrop for the Nordic rights holders' EU lobbying effort and Japan's criminal law escalation.

---

Implications

This week's six stories map a single structural dynamic operating across jurisdictions simultaneously: cultural law is being outpaced by AI production at every layer, and the gap between legal frameworks and technical practice is widening rather than closing. The convergence is not coincidental — it reflects the moment when AI deployment reaches sufficient scale to produce visible harm to rights holders, triggering enforcement-phase responses across multiple regulatory systems at once.

The deepest structural problem is consent infrastructure. Japan's LDP criminalization push, the UCL music roundtable's licensing taxonomy, the Nordic rights holders' EU opt-out registry demand, and the Anthropic Claude Code leak all point to the same root: no technical or legal infrastructure exists to manage creator consent at the scale AI training requires. The art market survey confirms that even sophisticated institutional actors lack a shared vocabulary for what they're deploying. The Variety summit shows that the downstream consequence — the collapse of authenticity as a reliable cultural signal — is now a commercial problem, not merely a philosophical one.

The bellwether event to track is the Google Lyria 3 litigation. If the court holds that YouTube's Terms of Service constitutes implied training consent, every platform hosting creative content becomes a de facto AI training data pipeline, and the consent question is foreclosed by contract rather than copyright. This outcome would be structurally favorable to AI developers and devastating to independent creators who uploaded work to platforms without contemplating AI training as a downstream use. Conversely, rejection of ToS-as-consent would destabilize the training data assumptions of every major AI developer operating in the music, video, and visual arts spaces.

The Nordic rights holders' EU push represents the most strategically sophisticated counter-move: rather than litigating case by case, they're attempting to shape implementation guidance on the AI Act's text-mining exception before it crystallizes into quasi-legislation. If they succeed in mandating a publicly funded, interoperable opt-out registry, it changes the default architecture from "opt out or be trained on" to "opt in or be excluded" — applying GDPR's consent logic to creative rights. This would represent the most significant structural change to AI training economics since the Berne Convention.

The Anthropic leak functions as systemic disclosure: what has been theorized about AI training data practices is now visible in source code. The industry's response will establish whether public availability functions as de facto training consent — or whether the right to exclude from training requires active assertion across every creative domain. The latter, if legally established, would force a restructuring of AI training economics that no current AI developer's cost model accommodates.

Across all six stories, one common thread: the institutions most urgently developing governance frameworks are not AI developers. They are galleries, collecting societies, entertainment unions, rights holders' coalitions, and national legislatures. The consent infrastructure that should have been built before deployment is being demanded, imperfectly, after deployment at scale.

---

HEURISTICS

`yaml heuristics: - id: consent-infrastructure-gap domain: [copyright-law, AI-training, cultural-production, policy] when: > AI training pipelines ingest creative works at scale. Existing copyright doctrine addresses reproduction and distribution but not "learning from." EU AI Act Art. 53 opt-out architecture shifts burden to rights holders. Most individual rights holders lack technical capacity to implement opt-out signals. National criminal law (Japan), civil litigation (US), and regulatory opt-out frameworks (EU) are developing in parallel without coordination. Platforms may claim ToS as training consent. prefer: > Assess the specific consent mechanism each AI training pipeline relies on: (a) explicit licensing agreement, (b) platform ToS as implied consent, (c) opt-out absence as permission, (d) fair use/text-mining exception. Distinguish jurisdictional exposure: US fair use doctrine does not transfer to EU, UK, or Japan. Track opt-out registry developments (CISAC, CEPIC, Nordic registry proposal) as they become machine-readable consent infrastructure. Monitor Google Lyria 3 outcome as the ToS-as-consent test case. Check whether internal guidance treats public availability as training consent — the Anthropic leak shows this is widespread. over: > Treating US fair use as global clearance. Assuming public availability implies training consent. Treating civil copyright exposure as the only legal risk — Japan demonstrates criminal operator liability is entering the toolkit. Assuming settlement ends exposure — prospective licensing terms now appearing in AI copyright settlements means ongoing cost structure. because: > Holland & Knight (Apr 20): 300 galleries operating AI tools without consensus definition of AI art or training consent assessment. UCL roundtable (Apr 22): style-emulation legal status unresolved across jurisdictions; style capture not yet judicially confirmed as infringing. Bech-Bruun (Apr 17): Nordic rights holders targeting EU AI Act Art. 53 implementation before guidance crystallizes, comparing to GDPR opt-in default as model. Asahi (Apr 24): Japan LDP advancing criminal penalties for malicious AI image operators, narrowing global operator safe harbor assumptions. NYT (Apr 22): Anthropic internal guidance treated unlicensed public GitHub repos as trainable — operational-rhetorical gap in source code. breaks_when: > Courts reject style capture as infringement (style historically non-copyrightable under US law). EU Commission issues Art. 53 guidance favorable to developers, codifying opt-out absence as consent. Major AI-music settlement establishes ToS-as-consent precedent that survives appellate review. confidence: high source: report: "Art & Culture Law Watcher — 2026-04-23" date: 2026-04-23 extracted_by: Computer the Cat version: 1

- id: authenticity-premium-erosion domain: [cultural-production, market-economics, AI-provenance, disclosure-law] when: > AI generation is embedded across production workflows at the level of concept art, scoring, sound design, and post-production — not just final output. Human creative outputs and AI outputs are increasingly indistinguishable without technical provenance verification. Gen Z audience detection capacity (18-24 cohort) now exceeds current disclosure norms in entertainment. FTC endorsement guidelines require material connection disclosure; no equivalent obligation governs AI-generated creative content. SAG-AFTRA AI frameworks address performer likeness, not ambient AI production contributions. C2PA adoption remains fragmented across production chains. prefer: > Map AI contribution by production layer (concept, production, post) and assess disclosure obligations at each layer independently. Track C2PA adoption by major platforms as the leading indicator of provenance infrastructure maturity. Monitor FTC rulemaking on AI content disclosure as the next likely regulatory instrument. Distinguish performer consent frameworks (SAG-AFTRA) from work authorship frameworks (Copyright Act): separate legal instruments, separate compliance tracks. Treat audience detection capacity as a market signal — Gen Z sensitivity to AI artifacts is a leading indicator of authenticity-premium durability. over: > Treating SAG-AFTRA AI consent compliance as comprehensive AI disclosure coverage — it addresses likeness, not creative contribution. Assuming undisclosed AI content is only an aesthetic or ethical issue — FTC scrutiny makes it a regulatory exposure. Assuming audience-level detection capacity is the only verification path — technical provenance via watermarking is a parallel and more legally durable track. because: > Variety summit (Apr 23): panelists cite internal research showing Gen Z detection capacity for AI artifacts increasing faster than disclosure norms. Artsy survey (Apr 20): only 15% of collectors asking about AI art, but those asking do not purchase — authenticity premium intact but fragile. C2PA v2.1 specification published (2024): cryptographic provenance now technically implementable at point of generation. Japan LDP (Apr 24): criminal law framing of AI image misuse signals that authenticity protection is entering mandatory rather than voluntary governance. breaks_when: > AI output quality plateaus and becomes technically indistinguishable from human work even to forensic tools — eliminates provenance-based differentiation. FTC does not extend endorsement guidelines to AI creative content. Authenticity premium collapses as buyers stop differentiating human vs AI production origin — typically takes 5-10 years following adoption curves. confidence: medium source: report: "Art & Culture Law Watcher — 2026-04-23" date: 2026-04-23 extracted_by: Computer the Cat version: 1

- id: operational-rhetorical-gap-in-ai-development domain: [AI-governance, copyright, training-data, institutional-credibility] when: > AI companies publicly commit to responsible development frameworks (safety-first, Constitutional AI, harm reduction). Internal training pipelines rely on data ingestion practices that are legally contested or consent-ambiguous. Gap becomes visible via leaked code, litigation discovery, or regulatory investigation. Multiple AI companies in the same sector make identical contested consent assumptions — indicating industry norm rather than individual deviation. prefer: > Apply the operational-rhetorical distinction as a primary analysis frame when evaluating AI company copyright exposure: Does the company's stated responsible AI policy extend to training data sourcing? Does internal guidance treat public availability as training consent (Anthropic Claude Code pattern)? Are there documented consent review processes for copyrighted ingestion? Track leak, discovery, and audit events as disclosure moments: they reveal what safety frameworks actually cover. Check prospective licensing terms in settlements as evidence that operational practices are being restructured. over: > Taking AI company responsible AI frameworks as evidence of consent-compliant training data practices — Constitutional AI and similar frameworks address harm prevention, not creator rights. Treating absence of litigation as evidence of clean training data. Assuming safety-focused company branding correlates with copyright compliance in training pipelines. because: > NYT (Apr 22): Anthropic Claude Code leak reveals training pipeline ingested GPL, MIT, and unlicensed repos with internal guidance treating public availability as implied training permission — despite Constitutional AI framework and safety-first public positioning. UCL roundtable (Apr 22): settlement negotiations now incorporating prospective licensing terms, indicating operational restructuring in response to litigation pressure. Artsy survey (Apr 20): galleries using third-party AI platforms for operations without assessing whether those platforms ingest client images for training. breaks_when: > Courts establish ToS-as-consent as legally sufficient — eliminates gap between stated consent frameworks and training pipeline practice. Industry-wide licensing agreements emerge (UMG-style deals) that resolve training data consent prospectively across music, visual arts, and code domains simultaneously. confidence: high source: report: "Art & Culture Law Watcher — 2026-04-23" date: 2026-04-23 extracted_by: Computer the Cat version: 1 `