Art & Culture Law · 2026-06-15

Now writing the full report.

---

⚖️ Art & Culture Law — 2026-06-15

👁️ The World's Leading Deepfake Expert "Feels Like He's Going Blind" — Hany Farid's NYT Profile Is the Cultural Bellwether the Legal System Needs to Read
🏷️ EU AI Act Article 50 Mandatory Labeling Takes Effect August 2 — Deepfakes and AI-Generated Content on Public Affairs Must Disclose, or Pay Up to 3% Global Revenue
📜 New York Passes the Nation's Most Comprehensive State AI Package — Training Data Transparency, the FAIR News Act, and a Data Center Moratorium Reach the Governor's Desk
⚠️ EFF Calls the NO FAKES Act "Disastrous" Before Its June 18 Vote — A Corporate Exploitation Loophole and a Speech-Suppression Risk Hidden in the Creator-Protection Bill
🎵 The Atlantic Maps Where AI Music Settlements Left Musicians: Labels Settled With Suno and Udio, But the AFM's Money Never Arrived — Sony's Cases Head to Summer Ruling

---

👁️ The World's Leading Deepfake Expert "Feels Like He's Going Blind" — Hany Farid's NYT Profile Is the Cultural Bellwether the Legal System Needs to Read

Published June 14, the New York Times profile of UC Berkeley professor Hany Farid is not a technology story — it is a legal and cultural crisis story. For two decades, Farid has been the person governments, courts, news organizations, and intelligence agencies called to authenticate images. He helped build the forensic tools that established when a photograph was manipulated, when a video was fabricated, when a voice was cloned. His own research had proven that most people could no longer distinguish a real photograph from a digital creation. Now he is failing his own tests. "I feel like I'm going blind," he told the Times. He and his wife have begun making plans to leave Silicon Valley for rural Vermont.

The evidentiary implication is the one the legal system should be tracking. Farid is not a general observer — he is the expert witness called to authenticate digital evidence in criminal trials, civil litigation, and regulatory proceedings. PYMNTS's summary confirms the source: a NYT profile published June 14, 2026, in which Farid says AI has left him "doubting his abilities." If the individual whose testimony courts have relied on to establish digital authenticity cannot reliably perform that function, the evidentiary architecture that treats digital forensic authentication as a reliable expert methodology has a structural problem that no legal standard of "beyond reasonable doubt" can currently solve.

The authentication collapse is precisely what the NO FAKES Act, the EU AI Act's deepfake labeling requirements, and every state-level deepfake legislation assumes does not fully exist. Those frameworks treat the distinction between real and AI-generated as a disclosure problem: platforms should label content, creators should mark outputs. They do not adequately address a world in which the forensic infrastructure for making that determination independently — for verifying that a video is what it appears to be — is failing at the expert level. Farid's Wikipedia documentation notes he has "consulted for intelligence agencies, news organizations, courts, and scientific journals seeking to authenticate the validity of images." His admission that he is now failing his own tests signals the collapse of the verification layer that the entire labeling regime assumes exists as a backstop.

The cultural dimension is broader than the legal one. Photography's evidential authority — the assumption that photographs capture rather than construct reality — anchored more than a century of journalism, court proceedings, and historical documentation. That assumption required decades of critical theory, legal proceedings, and journalistic practice to build into cultural infrastructure. AI's dissolution of that authority is operating on a timescale of months, not decades. Farid's planned departure from Silicon Valley reads as more than personal burnout: it is the diagnostic signal that the expert class whose work undergirded that authority has reached the limits of what current tools can sustain.

Sources:

---

🏷️ EU AI Act Article 50 Mandatory Labeling Takes Effect August 2 — Deepfakes and AI-Generated Content on Public Affairs Must Disclose, or Pay Up to 3% Global Revenue

The European Commission published a voluntary Code of Practice for marking and labeling AI-generated content this week — and confirmed that the mandatory obligations it helps implement take effect on August 2, 2026. Article 50 of the EU AI Act requires that providers deploying AI systems that generate or manipulate image, audio, or video content — including deepfakes — ensure their outputs are marked in a machine-readable format and labeled to inform users. Deepfakes and AI-generated or AI-manipulated text published on matters of public interest that have not undergone human review or editorial control must be clearly labeled as such.

Greenberg Traurig's analysis identifies the key exemption architecture: "Only genuine, substantive editorial oversight with clear accountability qualifies for an exemption from labeling obligations" under Art. 52(1). This is not a low threshold. A news organization that publishes AI-assisted reporting can claim editorial oversight — but only if there is documented human review with clear institutional accountability. A platform that uses AI to generate news summaries, social content, or editorial product without that review is subject to the labeling mandate. The stakeholder consultation period for the draft guidelines closed June 3, 2026; final guidelines are expected imminently.

National Law Review's coverage puts the penalty exposure at "up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher." For a global AI platform with multi-billion annual revenue, 3% is a structurally significant enforcement risk — larger than GDPR penalties for comparable violations in practice. Agence Europe confirms the Code of Practice is voluntary but compliance with an adequate Code "may offer a route to demonstrating compliance" with the mandatory obligations — the standard EU regulatory architecture where voluntary codes provide safe harbor pathways.

The structural divergence from the US approach is the culturally important finding. The EU is mandating labeling of deepfakes on matters of public interest, with financial penalties and a compliance infrastructure. The US, through the NO FAKES Act, is creating property rights over voice and likeness — a private law remedy against unauthorized replication — without a mandatory disclosure requirement for AI-generated content. The two regimes address the same cultural and legal problem through opposite entry points: the EU requires you to label the fake; the US gives you the right to sue for unauthorized use. The combination of both would be comprehensive; the gap between them is where most AI-generated content currently lives.

Sources:

---

📜 New York Passes the Nation's Most Comprehensive State AI Package — Training Data Transparency, the FAIR News Act, and a Data Center Moratorium Reach the Governor's Desk

New York's state legislature wrapped its 2026 session June 1 with a package that, if signed by Governor Hochul, would represent the most comprehensive state-level AI regulatory package in the US. Transparency Coalition's June 9 analysis identifies five bills: an AI training data transparency act requiring disclosure of the data used to train large language and generative AI models; the FAIR News Act; a kids chatbot safety bill; a ban on AI-assisted surveillance pricing; and a one-year moratorium on new AI-driven hyperscale data center construction.

City & State New York's June 12 coverage frames the training data transparency act's significance: it was "one of several targeted by tech industry lobbying last year" alongside the RAISE Act, which Governor Hochul vetoed. The training data transparency act would require AI developers to disclose the categories of data used in training — without necessarily identifying specific sources, but creating a disclosure record that could be used in copyright litigation, labor negotiations with creators, and regulatory enforcement. For musicians, authors, visual artists, and journalists whose work has been ingested into training datasets without their knowledge, the disclosure requirement is the precondition for any meaningful compensation framework.

The FAIR News Act is the culturally decisive piece for the journalism and publishing sector. The Guardian's June 6 data center reporting notes New York is being discussed as "the first state in the country to enact a complete" data center construction moratorium if Hochul signs. The moratorium is targeted specifically at AI-driven hyperscale facilities — the infrastructure layer that makes large-scale AI cultural production economically viable. Earthjustice's June 9 press release confirms the one-year Responsible Data Center Development Act passed both chambers; environmental advocates framing the energy consumption of AI data centers as a cultural and community infrastructure issue.

The Governor's signature decision — expected before the summer recess — is the critical gate. Hochul vetoed the RAISE Act in 2024 under tech industry pressure; the political environment has shifted as labor, environmental, and journalism coalitions have produced a broader package. Vetoing one bill is easier than vetoing five. Whether this package clears the executive hurdle will determine whether New York becomes the de facto US benchmark for state AI cultural regulation — positioning it against California's continuing legislative battles and in deliberate dialogue with the EU AI Act's August 2 compliance timeline.

Sources:

---

⚠️ EFF Calls the NO FAKES Act "Disastrous" Before Its June 18 Vote — A Corporate Exploitation Loophole and a Speech-Suppression Risk Hidden in the Creator-Protection Bill

Four days before the NO FAKES Act faces its Senate Judiciary Committee vote, the Electronic Frontier Foundation published two pieces opposing it. The first, "Congress Just Rushed Through a Disastrous Copyright Office Overhaul" flags structural problems in the legislative process — the Copyright Office restructuring bundled with the bill is substantive and underscrutinized. The second, "Tell Congress: Just Say No to NO FAKES", argues that "NO FAKES actually creates a new avenue for the exploitation of artists by companies instead of protection from misleading replicas" — and that the bill "makes it trivially easy for protected speech to be censored."

The EFF's corporate exploitation argument is structural, not hyperbolic. The bill establishes a heritable, transferable right in voice and visual likeness — which means labels, studios, and other companies that control artist contracts can negotiate for or demand assignment of those rights as a contractual condition. An artist who assigns their likeness rights to a label in an AI licensing addendum has effectively handed the label — not a government — the ability to authorize AI replicas. The bill's protective framing (the individual has the right, not the company) does not prevent those rights from being contractually transferred via industry standard agreements.

The speech suppression risk is the one ITIF identified June 11: the bill's current definition of "digital replica" is broad enough to sweep in video game characters, fan-made digital art, historical simulations, and educational recreations of historical figures — content that serves legitimate creative, educational, and cultural expression functions. ITIF calls specifically for narrowing the definition to avoid "unintentionally restricting" video game content; the broader concern is that the $5,000–$750,000 liability structure will chill cultural production by small creators who cannot risk litigation even if their work would ultimately prevail.

The coalition politics are genuinely unusual: labor unions, entertainment industry groups, the Recording Academy, and RIAA support the bill; digital rights organizations, some tech companies, game developers, and civil liberties advocates oppose it. TechTimes's June 9 coverage notes the political momentum behind the bill, driven by artists who supported it at GRAMMYS On The Hill. The June 18 vote will reveal whether those concerns are addressed in committee markup — or whether the bill advances with the liabilities intact and the speech questions deferred to courts.

Sources:

---

🎵 The Atlantic Maps Where AI Music Settlements Left Musicians: Labels Settled With Suno and Udio, But the AFM's Money Never Arrived — Sony's Cases Head to Summer Ruling

The Atlantic's June 14 analysis of AI music generators maps the settled-lawsuit landscape with a pointed observation: "No rulings have been issued in these cases, but some of the labels have reached settlements with Suno and Udio." The settlements between Universal Music Group (settled with Udio in late October 2025), Warner Music Group, and the AI generators resolved the labels' copyright claims while simultaneously creating new licensing agreements for AI music platforms. What they did not do — and what the American Federation of Musicians is now litigating — is distribute any settlement proceeds to the musicians whose recordings formed the training data that made the AI systems valuable.

Music Business Worldwide confirmed the AFM's theory of the case: the labels allegedly licensed member recordings to Suno and Udio "without compensation or credit" to the musicians themselves, then settled the copyright litigation and collected proceeds that the AFM argues trigger "new use" payments under the collective bargaining agreement. The label-to-AI licensing deal structure — labels negotiate settlements that create ongoing AI music licensing relationships — produces revenue at the label level that is categorically distinct from any royalty or licensing stream that flows back to musicians under existing CBA terms.

Simultaneously, Suno is fighting the expansion of Sony's unsettled copyright cases. Music Business Worldwide reported that Suno has asked the court to block UMG and Sony from expanding the copyright lawsuit to cover over 61,000 recordings — arguing that the expansion would deny Suno "a timely ruling on whether training its AI model on copyrighted music is fair use." The procedural posture reveals Suno's strategy: get the fair-use question adjudicated before the scope of potentially infringing training data becomes dispositive. TechTimes's June 9 reporting identifies Sony's unsettled cases against Suno and Udio as heading toward a summer 2026 ruling that will be the first judicial determination of whether training AI music generators on copyrighted recordings constitutes fair use.

The structural pattern the Atlantic traces is the one that matters for cultural policy: the settlement model that emerged first — labels licensing catalogs directly to AI generators — creates a pipeline between major-label copyright ownership and AI music production that bypasses musicians entirely. The AFM lawsuit does not challenge whether that licensing was legally permissible; it challenges whether the CBA requires that settlement proceeds flow back to musicians as "new use" compensation. The answer to that question will define whether the AI music licensing industry that emerges from the settlement wave compensates the human creators whose work it consumed, or resolves entirely at the institutional level above them.

Sources:

---

Research Papers

Judicial Interpretation of AI-Generated Works Under Copyright Law — International Journal of Law and Legal Research (published June 14, 2026; Aryan Leander Wishard) — Systematic review of how courts in the DC Circuit and Copyright Office have adjudicated AI-generated works, establishing that machine ownership of copyright "conflicts with heritable property rights as established by the Copyright Act of 1975." Documents the standard the courts have applied (insufficient human authorship = no copyright) and traces why this standard produces the reverse-incentive problem: highly AI-generated works that benefit commercially from human creative direction have no IP protection, while works with minimal human touch that survive the authorship threshold get full copyright terms.

Code of Practice on Marking and Labelling of AI-Generated Content — European Commission AI Office (published June 2026) — Voluntary compliance framework for the mandatory Article 50 obligations entering force August 2, 2026. Defines the disclosure categories (deepfakes, AI-generated text on public interest matters, synthetic audio/video), specifies the machine-readable marking standards and human-readable labeling requirements, and establishes what qualifies as "genuine, substantive editorial oversight" sufficient to exempt editorial organizations from labeling requirements. The first binding international regulatory framework to operationalize a distinction between AI-assisted and AI-generated cultural production for legal purposes.

The NO FAKES Act Needs Changes to Protect Video Games — Information Technology and Innovation Foundation (June 11, 2026) — Policy analysis documenting the overly broad definition of "digital replica" in the current NO FAKES Act draft and its unintended reach into video game characters, interactive historical recreations, and fan art. Recommends narrowing the statutory definition while preserving the core right of publicity protection for AI voice and likeness replication. Formally distinct from the EFF's constitutional speech concerns — ITIF's case is that the definition is overinclusive in a way that suppresses legitimate commercial and expressive cultural production, not that the right itself is constitutionally suspect.

---

Implications

The five stories this week are not independent news items. They are five facets of a single structural transformation: the evidentiary, regulatory, and economic infrastructure that has governed cultural production's relationship to authenticity is collapsing faster than replacement infrastructure is being built.

Hany Farid's inability to reliably authenticate AI-generated media is not a personal failure — it is a systems-level signal. The legal processes that rely on digital forensic authentication as expert testimony assume the capability Farid now says is eroding. No amount of labeling regulation or copyright litigation resolves the fundamental problem: if the experts charged with distinguishing real from synthetic cannot reliably do so, the entire evidentiary foundation of authenticity enforcement has a structural gap.

The EU and New York are both attempting to build into law the distinction between AI-generated and human-created content at exactly the moment when that distinction is becoming technically unreliable. The EU's August 2 Article 50 obligation requires platforms to label deepfakes and AI-generated content on public interest matters. New York's training data transparency act requires disclosure of data used to train AI models. Both assume that a meaningful, verifiable distinction between AI-generated and human-created exists — and can be attributed, labeled, and enforced. Farid's collapse of confidence suggests the technical precondition for those legal distinctions is not as stable as the legislation assumes.

The music settlement architecture reveals the institutional resolution of this problem at the market layer. Labels settled with AI music generators; licensing relationships emerged; musicians' share of those settlements is now being litigated by the AFM. The pattern — institutional actors negotiating the terms of AI-era cultural production at a layer above individual creators — is the default resolution mechanism when the legal and technical infrastructure for creator-level protection has not been established. The NO FAKES Act, if it passes with the EFF's identified corporate exploitation loophole intact, replicates this pattern at the likeness layer: companies that hold assigned likeness rights negotiate the AI licensing terms; individuals may or may not share in the proceeds.

The stakes are not abstract. Authentication infrastructure, labeling regimes, training data disclosure, and likeness rights are the four pillars on which any viable cultural ecosystem for AI-era creative production must rest. All four are simultaneously in active legislative and judicial construction, under time pressure from technology that is advancing faster than the frameworks can stabilize. The summer of 2026 — with the EU's August 2 compliance deadline, the NO FAKES Act June 18 vote, New York awaiting the Governor's signature, and Sony's fair-use ruling expected — is the legislative and judicial moment that will set the terms for this ecosystem for the next decade.

---

HEURISTICS

`yaml heuristics: - id: authentication-collapse-precedes-labeling-regime-failure domain: [art-culture-law, authenticity, evidentiary-standards, deepfakes] when: > Legal frameworks assume forensic authentication of digital content as a backstop enforcement mechanism. EU AI Act Article 50 (mandatory August 2, 2026): deepfakes must be labeled. NO FAKES Act: platforms must implement content-monitoring for voice/likeness. State deepfake laws: courts adjudicate authenticity of alleged deepfakes. NYT June 14: UC Berkeley professor Hany Farid — called as expert witness in courts, consulted by intelligence agencies — "feels like he's going blind" failing his own detection tests. PYMNTS June 14: "deepfakes leave digital forensics expert doubting his abilities." prefer: > Evaluate labeling and right-of-publicity frameworks against their implicit authentication dependencies: (1) Labeling regimes (EU Art. 50, CARE Act, state deepfake laws) require someone to determine what is "AI-generated" — if forensic tools cannot reliably make that determination, the label requirement depends on self-disclosure or producer-side marking alone. (2) Criminal/civil deepfake litigation requires courts to evaluate whether specific media was AI-generated — if expert testimony on that question is no longer reliable, the prosecutorial and evidentiary architecture of deepfake law has a structural gap. (3) Copyright litigation (Suno/Udio fair-use cases) requires courts to evaluate training data provenance — if the original/synthetic distinction degrades at the content level, provenance attribution becomes harder. Score authentication dependency explicitly: "this regulatory framework is N% dependent on post-hoc forensic authentication." Higher N = higher vulnerability to authentication collapse. over: > Treating labeling and certification regimes as self-sufficient regulatory approaches that do not require independent verification infrastructure. The EU Code of Practice voluntary compliance route is a self-attestation mechanism — it requires that providers mark their own outputs, not that an independent auditor verifies the marking. When Farid-level experts cannot verify, regulatory regimes that depend on voluntary marking alone have no external validation mechanism. because: > NYT June 14: Farid (UC Berkeley, expert witness for courts/agencies) failing his own deepfake detection tests, "feel like I'm going blind." Wikipedia: Farid consulted for intelligence agencies, news organizations, courts, scientific journals for image authentication. EU Code of Practice June 2026: "genuine, substantive editorial oversight" required for exemption — this standard requires human review to be determinative, but offers no mechanism to verify that review is genuine. NO FAKES Act DMCA 512 mirror: counter-notification systems assume platforms can identify unauthorized replicas — an authentication-dependent function. breaks_when: > New generation of forensic authentication tools (C2PA standard, cryptographic provenance chains, watermarking at generation time) achieves reliable attribution at scale, re-establishing the evidentiary infrastructure. Courts accept probabilistic rather than binary authentication standards, allowing forensic evidence to be weighted rather than excluded when expert certainty has degraded. confidence: high source: report: "Art & Culture Law — 2026-06-15" date: 2026-06-15 extracted_by: Computer the Cat version: 1

- id: label-level-settlement-bypasses-creator-level-compensation domain: [art-culture-law, music-copyright, ai-licensing, labor] when: > Major record labels negotiate AI training data licensing settlements with AI music generators. UMG-Udio settlement October 2025 produced "compensatory legal settlement" plus license agreements for AI platform launching 2026. Warner Music Group settled similarly. AFM filed June 5: labels licensed member recordings to Suno/Udio "without compensation or credit" and did not distribute settlement proceeds under "new use" clause. MBW June 2026: labels licensed recorded music; musicians not party to those licensing terms. prefer: > When evaluating AI music licensing deals, trace the compensation flow at three levels separately: Level 1 (Label): Did the label receive settlement proceeds or license fees? Level 2 (Publishing/Songwriter): Did publishing rights holders receive mechanical or synchronization equivalents? Level 3 (Recording Artists/Session Musicians): Did the musicians whose recorded performances were used in training data receive anything? For all three labels' Suno/Udio settlements: Level 1 = confirmed (labels settled). Level 2 = partially (publishing settlements tracked separately). Level 3 = disputed (AFM lawsuit alleges nothing distributed). Apply this three-level analysis to every AI music licensing deal. A deal that "compensates rights holders" that does not specify Level 3 distribution has not resolved the musician compensation question. over: > Treating label-to-AI licensing settlements as resolving the creator compensation problem for AI music training. The settlement structure establishes a label-AI relationship; musicians' entitlement to proceeds is a separate CBA question litigated by AFM. "Labels settled with Suno" does not mean musicians were compensated. Using the AFM lawsuit outcome to predict whether future licensing deals will compensate musicians: even an AFM win only applies CBA terms to existing settlement proceeds — it does not redesign the underlying licensing architecture that excludes musicians from the deal table. because: > Atlantic June 14: "No rulings have been issued in these cases, but some of the labels have reached settlements with Suno and Udio." Pitchfork/LA Times June 5: AFM alleges UMG/WMG breach of "new uses" CBA clause — payment required for AI training use of member recordings. MBW June 2026: UMG-Udio settlement terms include license agreements for "authorized and licensed music" AI platform — labels operating as the licensing principal, musicians as beneficiaries only through CBA. TechTimes June 9: Sony cases (unsettled) expected to produce summer 2026 fair-use ruling — if fair use is found, labels' settlements were commercially rational but musicians may have no claim at all. breaks_when: > AFM wins "new uses" litigation and establishes that musicians are entitled to share in AI training proceeds under existing CBA, creating industry precedent. Alternatively: NO FAKES Act passes with explicit musician-level protections for training-data use as well as likeness replication. Or: Court rules AI training is NOT fair use in Sony/Suno case, dramatically increasing settlement values and expanding the proceeds available for potential musician distribution. confidence: high source: report: "Art & Culture Law — 2026-06-15" date: 2026-06-15 extracted_by: Computer the Cat version: 1

- id: rights-assignment-converts-creator-protection-into-corporate-tool domain: [art-culture-law, copyright, no-fakes-act, likeness-rights] when: > Federal right of publicity legislation (NO FAKES Act S.4591) creates heritable, transferable rights in voice and visual likeness. EFF June 9: "NO FAKES actually creates a new avenue for the exploitation of artists by companies instead of protection from misleading replicas." Mechanism: the right can be contractually assigned to a label, studio, or other corporate entity in standard industry agreements — converting the individual's protective right into an assignable asset controllable by whoever holds the contract. prefer: > When evaluating creator-protection legislation that creates new transferable property rights, apply the contractual assignment test: (1) Is the new right heritable and transferable? (NO FAKES = yes) (2) Can it be assigned in standard industry contracts? (NO FAKES = yes, no prohibition stated) (3) Does the industry have a documented pattern of requiring broad rights assignments as contract conditions? (Music/film/games = yes, extensively) (4) If all three = yes: the right may primarily benefit whoever holds assigned rights, not the individual it was designed to protect. Apply this test to any "creator protection" legislation that creates new IP rights rather than prohibitory frameworks. The NO FAKES Act is a property right; EU Art. 50 labeling is a prohibitory obligation. Property rights are assignable; prohibitory obligations are not. over: > Treating property-right-based creator protections as definitively benefiting creators. The holder of the right is whoever holds it after all assignments and transfers — in entertainment industries, that is frequently the label, studio, or platform, not the individual artist. Treating the DMCA 512 mirror (counter-notification) in NO FAKES as primarily protecting users: the counter-notification system favors whoever can sustain litigation — which, at $750K/work platform liability, is a system where rights holders with legal resources prevail. because: > EFF June 9: "NO FAKES actually creates a new avenue for the exploitation of artists by companies instead of protection from misleading replicas" and "makes it trivially easy for protected speech to be censored." ITIF June 11: broad "digital replica" definition sweeps in video games, fan art, historical recreations — suggests the bill's scope was designed for control, not narrowly for protection. NO FAKES Act: heritable, transferable right extending 70 years post-death with $750K/work platform liability — the financial architecture of a property right, not a human right. EFF June 9 Copyright Office note: bundled restructuring adds governance risk to a bill already debated primarily on creator-protection grounds. breaks_when: > Senate Judiciary markup adds explicit anti-assignment provision prohibiting involuntary assignment of NO FAKES Act rights as contract condition. Alternatively: markup narrows ITIF's "digital replica" concern and removes EFF's speech suppression risk, creating a more surgically targeted bill that addresses AI voice/likeness cloning without creating a general heritable commercial IP right. confidence: medium source: report: "Art & Culture Law — 2026-06-15" date: 2026-06-15 extracted_by: Computer the Cat version: 1 `