Art & Culture Law Watcher - March 25, 2026 (Draft Iteration 2)

---

Table of Contents

1. Baltimore Sues xAI Over Grok's Nonconsensual Deepfake Generation 2. Trump Administration Declares AI Training Inherently Legal 3. India Enforces World's Fastest Deepfake Takedown Rules 4. Korean Music Industry Builds Blockchain Defense Against AI Exploitation 5. Art Galleries Adopt AI Without Governance Policies 6. Publishers Launch Mass Copyright Litigation Against Eight Tech Giants 7. Washington State Criminalizes Unauthorized Digital Replicas

---

Research Papers

A Unified Framework to Quantify Cultural Intelligence of AI (arXiv:2603.01211)

Sunipa Dev et al., March 20, 2026 | arXiv

Multi-institution research team proposes systematic framework for measuring how AI systems handle "offline" cultural knowledge—rituals, dialects, social norms—underrepresented in web-scale training data. Direct community engagement surfaces long-tail facets that reveal where models default to Western-centric stereotypes or shallow pattern matching. Framework addresses gap between platform claims of cultural competence and actual performance on hyper-local customs.

Do Deepfakes, Digital Replicas and Human Digital Twins Justify Personality Rights? (Journal of World Intellectual Property)

Bosher, March 2026 | Wiley

Legal analysis argues criminalizing intimate deepfake creation would impose stronger obligations on hosting platforms and payment providers. If creation itself were unlawful, Google would struggle justifying deepfake porn sites at top of search results, social platforms couldn't advertise nudify apps, payment processors couldn't fund deepfake ecosystems. Proposes high-harm capability licensing for tools enabling realistic sexual deepfakes or intimate transformations of real persons.

---

Implications

The week's developments expose structural divergence between jurisdictions attempting regulatory intervention (India, Korea, Washington State) and frameworks designed to protect AI industry velocity (Trump administration, failed UK opt-out regime). India's three-hour takedown rule and Korea's blockchain tracking system represent operational enforcement at technical infrastructure level—not aspirational guidelines. Meanwhile, U.S. federal policy explicitly declares training on copyrighted works non-infringing, preempting state-level experimentation.

The economic model is fracturing along cultural boundaries. Mistral's proposed 1-1.5% revenue levy acknowledges AI developers extract value from European cultural production without compensation, while Trump's framework treats such compensation as market-distorting regulation. Art galleries (84% using AI tools, 8% with formal policies) demonstrate the governance vacuum at cultural production sites—tools deployed faster than ethical frameworks can form.

Most revealing: Korean music industry's "golden time" framing—next 24 months determine survival. Not abstract policy debate, but existential timeline for creators whose work fuels training datasets while their economic models collapse. Blockchain provenance tracking and mandatory human-contribution certification aren't technological solutions seeking problems; they're last-resort defenses when legal frameworks can't keep pace with extraction speed.

Eric Schmidt's "build first, hire lawyers later" doctrine isn't scandal—it's operational reality. Chicken Soup for the Soul suing eight companies simultaneously (Apple, Meta, xAI, Google, Anthropic, OpenAI, Perplexity, NVIDIA) for using pirated book datasets reveals how normalized mass copyright infringement has become in AI development cycles.

What's absent: meaningful discussion of what cultural production looks like when training data extraction continues at current velocity while new human-created work becomes economically unviable. The gap between jurisdictions enforcing rapid content removal (India: 3 hours) and jurisdictions declaring training inherently legal (U.S. federal framework) creates regulatory arbitrage that platforms will exploit through jurisdictional shopping and infrastructure fragmentation.

Baltimore's lawsuit against xAI marks the first municipal challenge to deepfake generation capabilities themselves—not just distribution. The shift from targeting content dissemination to attacking creation tools indicates regulatory strategy evolution as intimate deepfakes proliferate beyond platform moderation capacity.

---

Stories

Baltimore Sues xAI Over Grok's Nonconsensual Deepfake Generation

Baltimore filed the first U.S. municipal lawsuit against Elon Musk's xAI on March 24, alleging its Grok chatbot transforms ordinary photos into nonconsensual sexualized deepfakes. The complaint argues xAI's image generation capabilities create legal liability by enabling users to produce intimate deepfakes without subject consent—a product design choice, not mere misuse by bad actors.

The lawsuit follows Malaysia and Indonesia blocking Grok access after users demonstrated the platform could generate explicit imagery from regular photographs. xAI restricted image generation features in jurisdictions where such content is illegal, but many countries lack comprehensive laws against AI-enabled sexual abuse material. When technology outpaces regulation and lacks built-in safeguards, victims remain exposed while those enabling harm avoid accountability.

Baltimore's legal theory treats high-harm AI capabilities as product safety issues requiring licensing or registration—not protected speech. The suit cites research proposing "high-harm capability class" regulation for tools enabling realistic sexual deepfakes or intimate transformations of real persons. If deepfake creation itself were criminalized, payment providers couldn't fund the deepfake ecosystem, Google couldn't return deepfake porn sites at top of search results, and app stores couldn't advertise nudify tools.

The case represents regulatory strategy evolution. Prior efforts targeted content distribution; Baltimore attacks generation infrastructure. This parallels India's three-hour takedown rules and Washington State's personality rights law (signed March 21, effective June 10)—jurisdictions regulating AI capabilities directly rather than waiting for federal frameworks.

xAI faces mounting legal pressure despite the Trump administration's March 20 framework declaring AI training non-infringing. Municipal and state lawsuits proceed independently of federal copyright policy, creating fragmented compliance landscapes where companies face liability in some jurisdictions while operating freely in others.

CNBC Coverage | Legal News Feed Analysis | Wiley JWIP Research

---

Trump Administration Declares AI Training Inherently Legal

The White House released its national AI legislative framework on March 20, explicitly stating "the Administration believes that training of AI models on copyrighted material does not violate copyright laws"—directly contradicting ongoing litigation from authors, publishers, and news organizations. The framework recommends Congress allow courts to resolve copyright disputes rather than establish legislative protections, effectively siding with AI companies in 19 active infringement cases.

The policy creates federal preemption architecture blocking state AI regulations. While states retain power over data center zoning and procurement, they cannot regulate AI development or penalize developers for third-party model use. This eliminates California-style regulatory experimentation that shaped tech governance over the past decade.

The framework's fair use position contradicts how tech companies protect their own intellectual property. OpenAI's terms forbid using ChatGPT outputs to train competing models; Anthropic, Google, and xAI have identical clauses. Meta sends DMCA notices demanding deletion of its "open" model copies from platforms. The structural asymmetry: we can train on your work, you can't train on ours.

Former Google CEO Eric Schmidt articulated the operational doctrine in an April 2024 Stanford lecture (video removed after one day): download whatever you need to build an accurate test version of your AI product. If the product succeeds, hire lawyers to clean up the mess. If nobody uses it, the stolen content doesn't matter. The framework institutionalizes this approach at federal policy level.

The copyright stance arrives as Chicken Soup for the Soul simultaneously sues Apple, Meta, xAI, Google, Anthropic, OpenAI, Perplexity, and NVIDIA for training on pirated book datasets (Books3, derived from Bibliotik). Encyclopedia Britannica filed similar claims March 16. The administration's position that such training is inherently lawful attempts to resolve these cases through executive policy rather than judicial process.

The framework claims to "respect intellectual property rights and support creators" while declaring the training they challenge non-infringing. This contradiction extends to AI outputs: platforms forbid users from training on ChatGPT responses while arguing their own training on copyrighted works constitutes fair use.

White House Framework | Axios Analysis | PBS Coverage | Mayer Brown Legal Brief

---

India Enforces World's Fastest Deepfake Takedown Rules

India's Information Technology Amendment Rules, notified February 10, 2026, compress content removal timelines from 36 hours to three hours for AI-generated deepfakes and "synthetically generated information" (SGI). The rules establish the first comprehensive regulatory definition of deepfake content and impose mandatory labeling, traceability, and audit obligations on intermediaries enabling SGI creation or distribution.

Platforms must now maintain round-the-clock processes to ingest government removal orders, classify content against prohibited SGI categories (child exploitation, non-consensual intimate imagery, impersonation fraud, electoral deception, forged documents), and execute takedowns within three hours. Failure strips intermediaries of safe harbor protection, exposing them to criminal liability under Indian law.

The rules introduce heightened due diligence for AI tools and video editing apps that enable synthetic content creation. These platforms must warn users against generating prohibited SGI categories, implement provenance labeling systems to distinguish synthetic from authentic media, and maintain detailed audit logs proving compliance with removal timelines. For in-house legal teams, this triggers operational overhauls: updating terms of service, building SOPs for three-hour response windows, training moderation staff on deepfake identification, running tabletop exercises to test compliance capacity.

The compressed timeline prevents contextual analysis of satire, public interest journalism, or artistic experimentation—favoring over-removal to avoid penalties. Grey areas remain: how will courts treat parody, anonymized training data, or satire when electoral sensitivities are involved? Who bears responsibility across the stack when front-end apps, model providers, and infrastructure intermediaries all touch prohibited content?

The regulation intersects with criminal laws on impersonation and obscenity, evolving evidence standards for AI-generated content, and India's data protection framework. Indian cricketer Gautam Gambhir's recent Delhi High Court filing seeking ₹2.5 crore damages for deepfake impersonation signals how personality rights litigation will test these rules' boundaries.

The three-hour standard positions India among the world's fastest enforcement regimes—faster than EU takedown requirements and opposite the Trump administration's March 20 framework declaring AI training inherently legal. This creates compliance fragmentation where platforms face criminal liability in India for content the U.S. government considers protected activity.

Mondaq Legal Analysis | MediaNama RTI Coverage | Times Now Report | Economic Times Gambhir Case

---

Korean Music Industry Builds Blockchain Defense Against AI Exploitation

Six major Korean music rights organizations launched the K-Music Rights Organization Mutual Growth Committee on February 26, warning that "the next two years will decide the life or death of Korea's music industry." The coalition—spanning KOMCA, recording industry associations, performers' federations, and producers—adopted an "AI-Era Music Rights Declaration" demanding a ban on AI training without creator consent, mandatory transparency in AI generation processes, and legal distinctions between human and AI-created works.

Korea already experienced AI's economic impact. When KOMCA discovered trot singer Hong Jin-young's "Love Is 24 Hours" was composed by an AI program (EvoM), the organization froze royalty payments in July 2022. Korea's Copyright Act defines creative works as "creations expressing human thoughts or emotions." If AI is the creator, no legal basis for royalty payments exists. EvoM had generated 300,000 compositions over six years, selling 30,000 tracks for 600 million won—revenue that suddenly had no legal standing.

The core vulnerability: AI systems train on millions of existing recordings without permission, unconsciously mimicking melodies and styles. Korean law compounds the problem by not defining a singer's voice as copyrightable work. Even when AI clones K-pop idols' voices illegally, existing performer rights can't effectively stop unauthorized content flooding platforms. A 2023 Security Hero report found Korean singers and actresses comprise 53% of individuals in deepfake pornography worldwide—eight of the top ten targets were Korean female singers.

The coalition's response is technical infrastructure, not lobbying. They're building a blockchain-based unified system connecting international standard identification codes (ISRCs, ISWCs) with platform content identification systems, creating auditable records of AI training pathways. As of March 24, 2025, KOMCA requires all new registrations include signed statements certifying "AI was not used and the work consists solely of human creative contributions." False statements trigger legal liability, royalty freezes, and database removal.

The policy doesn't ban AI use—works where AI served as assistive tool while human contribution remains clear may still qualify for protection. This aligns with WIPO's 2024 guidance that "AI-centric creations are difficult to protect under current copyright frameworks."

HYBE's response reveals industry strategy: acquire the threat. The K-pop giant bought AI voice startup Supertone for 45 billion won (56.1% stake), internalizing voice synthesis technology rather than waiting for regulation. The February 26 launch positions Korea among the fastest-moving nations on AI music governance, creating technical defenses (blockchain tracking) and legal frameworks (mandatory human-contribution certification) while global regulatory frameworks remain unresolved.

Billboard Coverage | KPOPPOST Analysis | STARNEWS Korea

---

Art Galleries Adopt AI Without Governance Policies

84% of art galleries use AI tools in daily operations, but only 8% have formal policies governing that use, according to First Thursday's "AI in Galleries" report and Artsy's 2026 AI Survey published last week. The gap reveals how quickly AI infrastructure has embedded in cultural institutions without corresponding ethical frameworks or disclosure protocols.

Gallery AI use spans operational and creative domains: 23% deploy AI for contract review, grant writing, and financial planning; 19% use it for installation renderings and virtual exhibition design. The absence of governance creates accountability vacuums. Who reviews AI-generated contracts for bias or legal risk? When AI designs exhibition layouts, how are aesthetic decisions attributed? Most galleries haven't formalized these questions.

The findings parallel broader museum sector developments. OpenAI's cultural heritage showcase in India demonstrated AI's potential for collection interpretation, but curator Marion Carré emphasized: "AI should amplify the work of museum professionals, never replace it. The people who build exhibitions, who research collections, who interpret artworks, who make curatorial choices—they hold the expertise. They understand context, nuance and cultural sensitivity."

At Art Central Hong Kong 2026, Kaitlyn Hau's "Recursive Feedback Ritual 0.01" used motion-capture performance and AI to explore tensions between human intuition and algorithmic reasoning. The New Museum's reopening exhibition "New Humans: Memories of the Future" (March 21) examines how humans and technology shape each other across 700+ objects. These institutional experiments treat AI as artistic medium requiring critical examination—not operational efficiency tool deployed without policy.

The governance gap has market implications. Premium 2D art buyers increasingly demand visible AI governance frameworks before production scales. Authentication, authorship control, and review accountability become buyer requirements when synthetic content can flood markets without provenance tracking.

The de Young Museum's "Monet and Venice" exhibition (opened March 21) featured an AI typewriter experience with Anthropic's Claude during the first week—demonstrating how major institutions experiment with AI engagement tools while commercial galleries deploy similar technologies without documented policies. Museum AI Summit 2026 (virtual conference) explores how cultural institutions can implement AI responsibly, but the 84%/8% gap suggests institutional guidance hasn't reached commercial gallery operations.

ArtNews Report | Artsy Survey | Guardian New Museum Coverage | KCRA de Young Report

---

Publishers Launch Mass Copyright Litigation Against Eight Tech Giants

Chicken Soup for the Soul sued Apple, Meta, xAI, Google, Anthropic, OpenAI, Perplexity, and NVIDIA on March 18, alleging hundreds of copyrighted books were used to train AI models via pirated datasets. The lawsuit joins 19 active copyright infringement cases against generative AI companies. Encyclopedia Britannica filed similar claims against OpenAI on March 16.

The suits center on Books3, a dataset derived from Bibliotik (pirate site) containing 196,640 books scraped without permission or compensation. Publishers characterize the dataset's use as "straightforward and deliberate theft" that violates fundamental copyright protections. Apple's inclusion is notable—the company maintains Apple Intelligence doesn't use Books3, but plaintiffs argue downstream model deployment benefits from infringing training regardless of direct licensing.

The litigation wave exposes industry's operational doctrine. Former Google CEO Eric Schmidt told Stanford students in April 2024 (lecture video removed after one day): download whatever you need to build an accurate test version of your AI product. If the product succeeds, hire lawyers to clean up the mess. If nobody uses it, the stolen content doesn't matter.

Companies invoke fair use, arguing AI training produces transformative outputs not derived from source material. But research demonstrates AI models can produce near-exact copies: complete passages from Harry Potter and the Sorcerer's Stone, fuzzy reproductions of existing artwork. Anthropic CEO Dario Amodei's 2021 internal memo (unsealed in current litigation) acknowledged AI could become "an increasingly extractive concentrator of wealth," proposing creators receive "a fraction of the profits from the model produced." Today, Anthropic argues fair use entitles authors to nothing.

U.S. federal court rulings began addressing training legality in 2025, but Schmidt's doctrine persists because litigation timelines extend years while product deployment happens in months. Meanwhile, news publishers lost antitrust standing when Judge Amit Mehta dismissed Helena World Chronicle and Emmerich Newspapers' case against Google on March 20, ruling they failed to prove monopoly power in online news markets.

The Trump administration's March 20 framework declaring AI training non-infringing attempts to resolve these 19 lawsuits through executive policy rather than judicial process. This creates direct conflict between federal guidance and active litigation, with publishers pursuing copyright claims the administration characterizes as legally baseless.

Atlantic Deep Dive | Reuters Britannica Coverage | AppleInsider Report | MediaNama Google Dismissal | Business Insurance Coverage

---

Washington State Criminalizes Unauthorized Digital Replicas

Governor Bob Ferguson signed legislation on March 21 making unauthorized use of AI-generated digital replicas illegal, with enforcement beginning June 10, 2026. The law updates Washington's personality rights statute to prohibit deepfake technology that forges voice or likeness without explicit consent, establishing civil penalties and stronger court remedies for misuse of AI-generated representations.

The bill, sponsored by Rep. Matt Boehnke (R-Kennewick), arrives as 31 states have enacted deepfake legislation—mostly targeting electoral contexts. Washington's approach focuses on personality rights violations across commercial, intimate, and impersonation scenarios. The timing puts Washington on collision course with the Trump administration's federal preemption framework announced March 20, which seeks to block state AI development regulations.

The law's passage coincided with Baltimore becoming the first U.S. city to sue xAI over Grok's image generation capabilities on March 24. Baltimore's lawsuit alleges Grok transforms ordinary photos into nonconsensual sexualized deepfakes, creating legal liability for platforms enabling such content. This follows Malaysia and Indonesia blocking Grok access for violating laws against AI-enabled sexual abuse material.

Washington's statute establishes legal standing for individuals whose digital identity is weaponized without consent—the framework Indian cricketer Gautam Gambhir invoked in his Delhi High Court filing seeking ₹2.5 crore damages for AI deepfakes. Deepfake legislation analysis notes the 2024-2025 election cycle "accelerated deepfake election legislation dramatically," but personality rights protections lagged behind electoral safeguards.

The legal landscape remains fragmented. Some states' child pornography statutes don't include AI-generated images. The Take It Down Act (signed May 2025) addresses nonconsensual intimate imagery including deepfakes, but enforcement mechanisms vary by jurisdiction. Legal analysis argues criminalizing intimate deepfake creation itself—not just distribution—would impose stronger platform obligations: payment providers couldn't fund the deepfake ecosystem, Google couldn't return deepfake porn sites at top of searches, app stores couldn't advertise nudify tools.

The June 10 effective date positions Washington among states regulating AI capabilities directly through personality rights frameworks rather than waiting for federal copyright resolution. This creates compliance fragmentation where companies face state-level criminal liability for activities the Trump administration's March 20 framework characterizes as federally protected.

Seattle Red Coverage | Elkhorn Media | Legal Analysis | AWARE Singapore Policy Context

---

Generated: March 25, 2026