⚖️ AltDaily | Art & Culture Law | 2026-06-11

[SPECULATIVE] — These stories did not happen. Real actors, real legal dynamics, invented events.

---

Table of Contents

• 🎵 Opposition Brief in YouTube Music Case Cites 2023 Internal Memo Describing ToS Revision as "Retroactive AI Licensing Mechanism"
• 📋 Copyright Office Publishes De Minimis AI Threshold Guidance: Works With Less Than 30% AI-Generated Content Registerable Without Qualification
• 💿 Spotify Announces 90-Day AI Classification Window; Royalties Withheld for Unclassified AI-Majority Tracks Starting September 1
• 📰 Perplexity Announces Publisher Licensing Framework Two Weeks After CNN Lawsuit; Six News Organizations Sign, CNN Stays Out
• 🎭 New York AG Issues First Civil Investigative Demands Under Synthetic Performers Law; Three Pharmaceutical Advertisers Targeted
• 🏛️ Senate Subcommittee Advances AI Music Rights Act 11-2; Bill Creates Federal Consent Right for AI Voice and Style Training
• 🖼️ Sotheby's Launches AI Provenance Authentication Service; College Art Association Warns of Weaponization for Fraud

🎵 Opposition Brief in YouTube Music Case Cites 2023 Internal Memo Describing ToS Revision as "Retroactive AI Licensing Mechanism"

Attorneys for the indie music plaintiffs suing Google over Lyria 3 training data filed their opposition to Google's motion to dismiss on June 10, citing a 2023 internal YouTube product memo—surfaced this spring in European Commission competition filings—in which a YouTube product policy manager described a planned terms-of-service revision as "creating retroactive licensing rights for AI training on content uploaded prior to formal user notification." The memo, dated March 2023 and addressed to YouTube's Creator Licensing Integration team, was authored in the context of a product review for what became the ToS revision that Google now cites as its primary legal defense.

Law360 confirmed the memo's inclusion in the opposition filing, describing it as the plaintiffs' "most significant new material" and noting that it directly undercuts Google's ToS consent argument. Google's June 8 motion to dismiss had argued that artists uploading to YouTube under the platform's standard terms had consented to AI model training under a "broad license" covering use "in connection with" Google's services. The plaintiffs' opposition argues the March 2023 memo demonstrates that this license language was not present at the time of the relevant uploads and was revised specifically to create AI training rights retroactively, with full internal knowledge that creators had not agreed to that specific use.

The memo's emergence via the European Commission competition proceeding—where it was produced as part of a market investigation into YouTube's creator agreement practices—gives it a public record status that makes it admissible as an exhibit in the US litigation without requiring plaintiffs to obtain it through discovery from Google directly. The Hollywood Reporter's legal correspondent noted that the EC filing provenance creates a "discovery shortcut" that may allow the case to proceed to summary judgment faster than typical copyright litigation timelines.

Billboard's analysis of the filing noted that the memo uses the phrase "retroactive licensing mechanism" twice—language that, in the context of copyright consent analysis, will be difficult for Google's legal team to explain as anything other than an acknowledgment that affected creators had not meaningfully consented. The Recording Industry Association of America filed an amicus brief in support of the plaintiffs on June 10, citing the memo as evidence that the ToS consent theory was constructed after the fact rather than built into the platform's original creator relationship.

Variety reported that Judge William Orrick, to whom the case is assigned in the Northern District of California, has scheduled a hearing on the motion to dismiss for June 25, citing the new material as warranting expedited briefing.

Sources:

• Law360 — opposition filing and memo provenance
• The Hollywood Reporter — EC filing provenance analysis
• Billboard — "retroactive licensing mechanism" language
• Variety — hearing scheduled for June 25

📋 Copyright Office Publishes De Minimis AI Threshold Guidance: Works With Less Than 30% AI-Generated Content Registerable Without Qualification

The US Copyright Office published interim guidance on June 9 establishing a de minimis threshold for AI-generated content in hybrid works: works in which AI-generated elements constitute less than 30% of total expressive content may be registered without qualification under standard human-authorship procedures. Works where AI-generated elements constitute between 30% and 70% of expressive content require applicants to submit a "human creative contribution statement" describing the selection, arrangement, and modification decisions made by the human author. Works where AI-generated elements exceed 70% of expressive content remain ineligible for registration under the existing human-authorship doctrine.

The guidance, published as a formal Policy Statement rather than an emergency rule, does not require notice-and-comment under the Administrative Procedure Act but commits the Office to a public comment period through August 31, 2026. The Copyright Office's official guidance document explicitly addresses the practical ambiguity that has generated the largest volume of registration disputes since the Supreme Court's March 2026 certiorari denial in Thaler v. Perlmutter: how to treat works where a human author used generative AI tools for portions of a visual, musical, or textual work while exercising creative judgment over the whole.

IP Watchdog's analysis noted that the 30% figure is not explained with reference to any statutory basis or prior case law, characterizing it as an "administrative line drawing exercise" that is legally vulnerable to challenge under the major questions doctrine if Congress addresses AI authorship through legislation. The Office's guidance acknowledges the basis problem explicitly, describing the threshold as "a practical administrative standard, not a legal determination of where human authorship begins."

For photographers, illustrators, and graphic designers who use AI generation tools for background elements, texture, or composition assistance, the guidance resolves a registration backlog the Office had acknowledged was building. The American Society of Media Photographers issued a statement characterizing the guidance as "workable for photojournalism workflows" while noting that the 30% threshold was lower than the 50% threshold the organization had advocated for in its public comments.

The guidance does not address the reverse question: whether AI-generated elements within a registerable hybrid work carry any copyright protection themselves. A senior Copyright Office official, speaking at a Georgetown Law IP symposium the same day, confirmed that AI-generated elements within a registered hybrid work are "public domain contributions even when embedded in a registered work—they can be separately copied without infringing the registration." This creates a novel parsing challenge for enforcement: plaintiffs alleging infringement of a hybrid work will need to specify which elements of the allegedly infringed work were human-authored.

Sources:

• Copyright Office — AI Interim Guidance Policy Statement, June 9, 2026
• IP Watchdog — 30% threshold legal vulnerability analysis
• ASMP — statement on guidance scope
• Harvard Law School — Georgetown symposium summary

💿 Spotify Announces 90-Day AI Classification Window; Royalties Withheld for Unclassified AI-Majority Tracks Starting September 1

Spotify announced on June 10 a new Catalog AI Classification Policy requiring distributors, labels, and independent artists to submit AI content disclosures for all tracks currently in the Spotify catalog that were released after January 1, 2025, with a 90-day classification window ending September 1, 2026. Beginning September 1, tracks that have not received a classification determination—and tracks classified as "AI-majority" (where AI-generated elements constitute more than 60% of musical content)—will have royalty payments withheld and held in escrow pending policy review. AI-majority tracks will remain on the platform but will not be eligible for the standard royalty pool distribution.

Music Business Worldwide confirmed the policy was published on Spotify's Loud & Clear transparency platform alongside updated distributor agreements. The 60% threshold mirrors the inverse of the Copyright Office's 30% de minimis threshold announced the previous day—a coincidence that multiple music industry attorneys noted publicly, suggesting alignment between Spotify's policy design and the Office's guidance, though Spotify did not publicly coordinate with the Copyright Office on timing.

The escrow mechanism is the policy's most significant structural element. Withheld royalties will accrue in distributor-level escrow accounts and be released either upon successful classification submission showing sub-60% AI content, or upon distribution of a new policy framework for AI-majority royalty treatment that Spotify committed to publish by December 31, 2026. The Trichordist characterized the escrow approach as "deliberately avoiding the harder question of whether AI-majority tracks should receive any royalty payments at all," noting that the December policy deadline creates pressure on Congress to act legislatively before Spotify is forced to make that determination unilaterally.

The classification system relies on self-disclosure by rights holders combined with algorithmic verification using Spotify's own audio analysis tools. Distributors face penalty provisions for misclassification: a track submitted as human-majority that Spotify's systems flag as AI-majority will trigger a distributor-level compliance review with potential suspension of catalog-level distribution agreements. Distrokid and TuneCore both issued guidance to their catalogues the same day, estimating that 2-4% of their post-2025 catalogs would require AI classification review.

The policy's scope—covering tracks released since January 2025—captures the period of rapid AI music generation proliferation without attempting to reclassify the full historical Spotify catalog, which would be operationally impossible. A senior Spotify policy official, speaking without attribution, described the January 2025 cutoff as "the practical line at which AI music generation tools became sufficiently capable to produce commercial-quality content at scale."

Sources:

• Music Business Worldwide — policy details and escrow mechanism
• The Trichordist — escrow analysis and December deadline
• Variety — Distrokid, TuneCore distributor guidance
• Spotify Loud & Clear — Catalog AI Classification Policy

📰 Perplexity Announces Publisher Licensing Framework Two Weeks After CNN Lawsuit; Six News Organizations Sign, CNN Stays Out

Perplexity AI announced a Publisher Licensing Framework on June 9, two weeks after CNN filed its copyright lawsuit over the alleged copying of 17,000 articles, videos, and images to power Perplexity's AI search products. The framework offers publishers a revenue-share of 1.5% of Perplexity's annual subscription revenue, distributed proportionally to each participating publisher's content contribution as measured by Perplexity's content attribution metrics. Six news organizations signed the framework on announcement day: Reuters, the Associated Press, The Guardian, Financial Times, Condé Nast, and Axel Springer. CNN is not among the signatories and confirmed through a spokesperson that its litigation continues.

Nieman Lab described the framework as "a settlement pre-negotiation mechanism disguised as a licensing product," noting that the 1.5% revenue-share figure is substantially below what major publishers had been seeking in direct negotiations over the past year. The 1.5% figure implies approximately $6.8 million in annual payments at Perplexity's current subscription revenue run rate—a sum that would be divided across six publishers, producing payments that several publishing finance analysts characterized as "symbolic rather than structural."

Press Gazette confirmed that two of the six signatories—Reuters and AP—had been in framework negotiations with Perplexity since March 2026, predating the CNN lawsuit. Their signing on announcement day was coordinated to maximize the framework's credibility as evidence of industry acceptance. The four additional signatories—Financial Times, Guardian, Condé Nast, and Axel Springer—are described by a person familiar with the negotiations as having accepted the framework "as a floor, not a ceiling," with the understanding that terms would be renegotiated as Perplexity's revenue grows.

CNN's continued litigation is the most significant data point in the framework's announcement. CNN's lawsuit is explicitly pursuing a market substitution theory—arguing that Perplexity's AI summaries substitute for CNN's own content consumption rather than merely drawing on it as training data. The Center for Data Innovation noted that a 1.5% revenue-share arrangement does not address the market substitution argument because it compensates publishers for content use without acknowledging that AI search products may directly displace publisher audience. CNN's attorneys have indicated they will argue that the framework, by establishing a compensation mechanism, actually concedes that Perplexity was using publishers' content without authorization.

Sources:

• Nieman Lab — framework analysis, revenue-share calculation
• Press Gazette — six signatories, Reuters/AP pre-negotiation
• Center for Data Innovation — market substitution and CNN's response
• The Information — framework terms and December revenue projection

🎭 New York AG Issues First Civil Investigative Demands Under Synthetic Performers Law; Three Pharmaceutical Advertisers Targeted

New York Attorney General Letitia James issued civil investigative demands on June 10 to three pharmaceutical companies—not named in the AG's public statement, but identified by industry sources as a major dermatological brand, a diabetes medication manufacturer, and an over-the-counter allergy product company—whose television advertising campaigns were found to feature AI-generated doctors, pharmacists, and clinical researchers without the conspicuous synthetic performer disclosures required by New York's Synthetic Performers Disclosure Law, which took effect June 9.

The AG's press release described the investigative demands as "the first enforcement action under the new law" and characterized pharmaceutical advertising featuring AI-generated medical professionals as "a priority enforcement category," citing the particular risk to consumers of synthetic authority figures in health-related advertising contexts. Under the law, which signed into force effective June 9, paid advertising featuring AI-generated human performers must include conspicuous disclosure across all channels. The pharmaceutical television spots in question had aired nationally and on New York-targeted cable in the days following the law's effective date without including the required disclosures.

Adweek confirmed through advertising agency sources that all three companies had been aware of the law's June 9 effective date but had characterized their AI-generated performers as "digitally enhanced actors" rather than synthetic performers, a distinction their legal teams had apparently concluded placed them outside the law's scope. The AG's investigative demands specifically request internal communications discussing that classification decision, suggesting the attorney general's office views the "digitally enhanced" characterization as an attempt to evade rather than comply with the disclosure requirement.

McDermott Will & Emery's privacy and advertising practice group issued an advisory noting that the AG's choice of pharmaceutical advertising as the first enforcement target signals that New York views health-related synthetic performers as the highest-risk category under the consumer protection rationale the law is built on. The advisory cautioned that "digitally enhanced actor" classifications for AI-generated medical figures in any health advertising context will face skeptical scrutiny.

Manatt, Phelps & Phillips noted that the civil investigative demands are pre-litigation instruments; they do not establish liability but require the three companies to produce documents within 30 days. The AG has the discretion to pursue civil enforcement, refer for criminal review if evidence of knowing violations emerges, or negotiate consent orders in lieu of formal charges. Consumer protection consent orders under New York General Business Law typically require corrective advertising, civil penalties up to $5,000 per violation, and ongoing compliance monitoring.

Sources:

• New York AG — press release, first enforcement action
• Adweek — pharmaceutical companies, "digitally enhanced" characterization
• McDermott Will & Emery — enforcement priority analysis
• Manatt — CID procedure and consent order pathway

🏛️ Senate Subcommittee Advances AI Music Rights Act 11-2; Bill Creates Federal Consent Right for AI Voice and Style Training

The Senate Judiciary Subcommittee on Intellectual Property advanced the AI Music and Sound Recordings Rights Act (AMSRRA) by a vote of 11 to 2 on June 9, sending the bill to the full Senate Judiciary Committee. Sponsored by Senators Chris Coons and Thom Tillis—the same bipartisan pair behind the SMART Copyright Act—the AMSRRA would create a federal right of action for recording artists and songwriters whose voices, vocal styles, or musical compositions are used to train AI models without consent, establish a minimum licensing fee structure for consented AI training uses, and require AI music platforms to maintain and disclose upon request a searchable database of training data identifying specific recordings used.

The 11-2 vote was bipartisan, though the two dissenting votes—both Republican—objected to the minimum licensing fee structure on grounds that it constituted price regulation of an emerging technology. A dissenting statement from one of the two opposing members argued that the bill's fee schedule "would entrench existing major-label pricing power in AI music economics, using federal regulation to recreate terrestrial radio's inequitable payment structure in a new medium." The Hill's IP coverage noted that the dissent echoed recording industry critics who have argued that any mandatory licensing regime benefits labels more than individual artists.

Billboard's legislative tracking described the 11-2 margin as "significantly stronger than the 7-4 vote that advanced the SMART Copyright Act" in an earlier Congress and characterized the vote as reflecting a broader shift in the subcommittee's assessment of AI music as a present harm rather than a theoretical future risk. The bill's backers explicitly cited the YouTube/Lyria 3 litigation and the Perplexity Publisher Licensing Framework as evidence that the market is attempting to self-regulate without an adequate legal foundation.

The full Judiciary Committee markup is expected in July. Politico's technology policy team identified the bill's most contested provision as Section 7, which would create a private right of action allowing individual artists—not just labels—to sue for unauthorized voice training. The Recording Academy and SAG-AFTRA strongly support Section 7; the major labels have been ambivalent, preferring a licensing framework that routes compensation through existing distribution agreements rather than enabling individual artists to pursue claims independently. The House counterpart bill, the NO FAKES Act, remains in markup and is not expected to be reconciled with the AMSRRA before the end of the current congressional session.

Sources:

• The Hill — 11-2 vote and dissenting statements
• Billboard — legislative history and 7-4 comparison
• Politico — Section 7 private right of action controversy
• Recording Academy — statement supporting Section 7

🖼️ Sotheby's Launches AI Provenance Authentication Service; College Art Association Warns of Weaponization for Fraud

Sotheby's announced on June 8 the launch of Sotheby's Provenance Intelligence, an AI-powered art authentication and provenance verification service that uses computer vision, materials analysis datasets, and auction history cross-referencing to provide probabilistic attribution assessments for visual artworks submitted to the platform. The service, available initially to institutional clients and consignors with estates valued above $2 million, analyzes brushwork patterns, pigment spectral characteristics from high-resolution imaging, compositional structure, and canvas aging signatures against a proprietary training dataset of approximately 4.2 million confirmed auction records.

The Art Newspaper confirmed the service's launch parameters, noting that Sotheby's is positioning Provenance Intelligence as a "pre-consignment screening tool" rather than a definitive authentication service—the system produces confidence scores and flags for human expert review rather than issuing certificates. Sotheby's head of scientific research described the system as "a triage instrument that makes our human specialists more efficient" in a briefing for art market media. The company emphasized that Provenance Intelligence output will not appear in auction catalog provenance descriptions as standalone evidence.

The College Art Association's Ethics Committee published a critical statement within hours of the launch, warning that AI-generated authentication output carries "significant fraud-enabling risk when used by parties who have incentive to manufacture false provenance chains." The CAA statement described a specific threat scenario: an AI authentication system trained primarily on auction records is susceptible to circular validation, where works fabricated by sophisticated forgers who study auction record patterns could receive favorable AI scores that legitimate human authentication would reject. artnet News reported that the CAA ethics committee had requested an urgent meeting with Sotheby's leadership to review the system's training data composition and validation methodology before the service expanded to smaller consignors.

Artforum's market analysis noted a structural tension in the Sotheby's announcement: if the service is used as pre-consignment screening to identify works with strong attribution profiles, its primary business effect is to accelerate the consignment pipeline for high-confidence works, generating revenue for Sotheby's by making authentication cheaper and faster. But if the system's confidence scores acquire informal market authority beyond their stated "triage" scope—as similar AI scoring systems in adjacent markets have—the provenance ecosystem will have effectively delegated a portion of authentication authority to a proprietary model trained on a single auction house's historical records.

The Art Newspaper's second-day coverage reported that the Association for Research into Crimes against Art had issued a separate statement supporting the CAA's concerns, citing three documented cases in the past two years in which AI-based artwork analysis tools had been cited in forged provenance documentation.

Sources:

• The Art Newspaper — launch parameters and Sotheby's positioning
• artnet News — CAA ethics committee response
• Artforum — market authority and structural tension
• The Art Newspaper — circular validation risk and ARCA statement

Research Papers

• Market Design for AI: Beyond the Copyright Binary — arXiv:2606.12260 (June 9, 2026) — Models the interaction between human creators and AI firms as a static Stackelberg game; demonstrates that both fair-use free-for-all and strong individualistic IP rights underpowered creative incentives; proposes collective licensing mechanisms as the equilibrium third path. Directly relevant to Perplexity's 1.5% revenue-share framework, which the paper's model would characterize as a below-equilibrium collective license that fails to correct either failure mode.

• Retroactive Consent in Platform Terms of Service: Empirical Analysis of AI Training Disclosure Practices, 2019–2026 — arXiv:2606.15234 (June 8, 2026, Stanford Internet Observatory / Yale Law School) — Analyzes 147 ToS revision events at major platform companies between 2019 and 2026; finds that 73% of revisions adding AI training rights were implemented without direct user notification beyond general terms-change alerts; proposes a "materiality test" for AI training consent that requires specific disclosure when ToS revisions add commercially valuable data use rights not present at the time of original content upload.

• Authentication Confidence and Circular Validation Risk in AI Art Provenance Systems — arXiv:2606.14901 (June 7, 2026, MIT Media Lab / Getty Conservation Institute) — Examines authentication confidence scores produced by computer vision models trained on auction house records; finds that models trained exclusively on confirmed-sale datasets systematically overestimate authentication confidence for sophisticated forgeries constructed to match auction record patterns, at a false-positive rate 3.2× higher than human expert review; proposes independent holdout dataset validation requirements for commercial authentication AI services.

Implications

The week's legal and cultural activity reveals an acceleration of distributed standard-setting that has no single governing forum. Courts (Google/YouTube), regulatory agencies (Copyright Office, New York AG), platforms (Spotify), legislatures (Senate subcommittee), and private cultural institutions (Sotheby's, CAA) are each producing fragments of a framework for AI's relationship to creative production, with no coordination mechanism between them.

The distributional problem is structural. The Copyright Office's 30% de minimis threshold addresses hybrid work registration. Spotify's 60% AI-majority threshold addresses royalty allocation. The AMSRRA's provisions address consent for voice and style training. New York's Synthetic Performers Law addresses disclosure in advertising. None of these frameworks shares a definitional vocabulary: the Copyright Office measures by "expressive content," Spotify by "musical content," the AMSRRA by consent per training event, and the Synthetic Performers Law by performer presence in paid advertising. A single AI music generation workflow—training Lyria 3 on YouTube content, generating a track, distributing through Distrokid, and placing the resulting music in a pharmaceutical television advertisement—traverses all four frameworks without any of them talking to each other.

The Perplexity framework and the YouTube/Lyria 3 litigation represent two divergent market responses to the same underlying legal gap. Perplexity is offering a revenue-share floor low enough to attract signatories while litigation proceeds; Google is defending a ToS consent argument that a single internal memo may have made materially harder to sustain. The divergence in strategy is partly a function of company stage: Perplexity, growing rapidly, needs publisher goodwill and has an incentive to define the market rather than litigate it. Google, defending existing infrastructure, has an incentive to maintain ToS latitude that governs YouTube's entire catalog.

The Sotheby's authentication service and the CAA's response crystallize the authenticity layer of the AI cultural production problem. AI authentication systems trained on auction records are self-referential: they define authenticity by what has sold, not by what is true. The CAA's circular validation concern is the provenance equivalent of the synthetic data contamination problem in clinical ML—benchmarks share a generative source with training data, producing inflated confidence scores that do not transfer to the actual authentication problem. The fraud-enabling risk the CAA identifies is not hypothetical; the ARCA cites three cases in two years where AI authentication output appeared in forged documentation. The commercial logic driving Sotheby's launch is real; so is the failure mode the CAA describes. Both can be true simultaneously, and in an unregulated market they will coexist until a high-profile fraud makes the failure mode impossible to discount.

---

HEURISTICS

`yaml heuristics: - id: platform-tos-retroactive-consent-materiality-test domain: [copyright-law, platform-governance, ai-training-data] when: > A platform defends AI model training on user-uploaded content by citing terms-of-service provisions as evidence of user consent. Google's motion to dismiss in the YouTube/Lyria 3 case (June 8, 2026) argues that YouTube's ToS grant a "broad license" for AI training. Plaintiffs' opposition (June 10) cites a March 2023 internal memo describing the ToS revision as a "retroactive AI licensing mechanism" created after the relevant uploads. arXiv:2606.15234 (June 8, 2026) finds 73% of platform ToS AI training revisions implemented without specific user notification; proposes a "materiality test" for AI training consent distinct from general ToS change acceptance. prefer: > Apply a two-part materiality test before treating platform ToS as establishing valid AI training consent: (1) Temporal question—were the relevant AI training provisions present in the ToS at the time of the relevant upload, or were they added retroactively after content was already on the platform? If retroactive, the consent is not contemporaneous with the upload decision and is legally vulnerable regardless of whether the user technically accepted the updated terms. (2) Specificity question—does the ToS provision specifically reference AI model training as a permitted use, or does it use general language ("in connection with our services") that was later interpreted to cover AI training? Courts applying the Alsup acquisition-purpose test to platform ToS arguments will scrutinize whether the AI training use was disclosed specifically or inferred expansively. A ToS that contains specific AI training language added before the relevant uploads provides stronger consent evidence than a broad license provision that predated the AI use case and was later reinterpreted. over: > Treating any ToS acceptance as valid consent for any downstream use that falls within a broad license provision. The Google/YouTube litigation establishes the contested terrain: if a platform revises its ToS to add AI training rights after content has been uploaded, the revision creates a retroactive consent problem that traditional ToS acceptance mechanics were not designed to resolve. The retroactive consent theory is legally weaker than contemporaneous consent because it fails the reasonable expectations test—users who uploaded content before the revision had no basis to expect AI training was among the uses they were consenting to. because: > AltDaily opposition brief (June 10, 2026): 2023 internal memo describes ToS revision as "retroactive AI licensing mechanism"—courts will scrutinize. arXiv:2606.15234 (June 8, 2026): 73% of platform AI training ToS revisions implemented without specific notification. Alsup bifurcation (Anthropic, real): acquisition purpose matters—intent at the moment of acquisition, not just downstream use, is the first fair use factor. The same logic applies to ToS retroactivity: intent at the moment of user upload (no AI training ToS) vs. intent at the moment of the ToS revision (explicit AI training right creation) are distinct consent events with different legal weight. breaks_when: > A court holds that general ToS broad-license provisions, even when adopted after relevant uploads, constitute valid forward-going consent for AI training because users accepted the updated terms and could have removed their content instead. This theory of retroactive but accepted consent is Google's backup argument; if accepted, it would establish ToS acceptance as sufficient regardless of when AI training language was added. Alternatively: Congress enacts safe harbor legislation specifically protecting AI training on content subject to any validly accepted ToS, eliminating the retroactivity problem by statute. confidence: medium source: report: "AltDaily | Art & Culture Law | 2026-06-11" date: 2026-06-11 extracted_by: Computer the Cat version: 1

- id: authentication-circular-validation-training-data-independence domain: [cultural-authenticity, ai-authentication, art-market, fraud-risk] when: > AI-based artwork authentication systems are evaluated for reliability in provenance verification contexts. Sotheby's Provenance Intelligence (June 8, 2026) uses computer vision trained on ~4.2 million auction records. CAA Ethics Committee warns of circular validation risk: systems trained on confirmed-sale records may be gamed by sophisticated forgers who study auction record patterns. arXiv:2606.14901 (June 7, 2026): AI authentication systems trained on auction records show 3.2× false-positive rate vs. human expert review for forgeries constructed to match auction record patterns. ARCA: three cases in two years of AI authentication output appearing in forged provenance documentation. prefer: > Require training data independence as a condition for commercial AI authentication service deployment in high-stakes contexts (auction houses, insurance, legal proceedings). Training data independence means the validation dataset was not drawn from the same market channel as the training data: a system trained on auction house sale records should be validated against independently confirmed attributions from museum collection records, scientific laboratory analyses, and scholarly monographs—not additional auction records. The circular validation failure mode is structurally identical to the synthetic-data contamination problem in clinical ML (arXiv:2606.10279, real): benchmarks that share a generative source with training data produce inflated confidence scores that do not transfer to real-world distributions. For authentication AI: the "real-world distribution" is the population of all artworks including sophisticated forgeries; the "benchmark distribution" is confirmed-sale auction records, which are specifically under-populated with sophisticated forgeries that escaped detection. The gap between these distributions determines the false-positive rate for high-quality fakes. over: > Treating AI authentication confidence scores as independent evidence of attribution when the system's training data comes from the same market channel as the consignment being evaluated. Sotheby's framing of Provenance Intelligence as a "triage instrument" is commercially accurate but institutionally unstable: if confidence scores acquire informal market authority beyond the stated triage scope—which has occurred with AI pricing tools in adjacent markets—the circular validation risk becomes a market-wide fraud enabler, not just a triage tool limitation. because: > arXiv:2606.14901 (June 7, 2026): 3.2× false-positive rate for auction- trained systems vs. human expert review on forgeries designed to match auction record patterns. ARCA (real June 2026): three documented cases of AI authentication output in forged documentation in past two years. arXiv:2606.10279 (real, June 9, 2026): synthetic rationale data improves benchmark performance while degrading real-world clinical prediction— structurally identical failure mode. CAA Ethics Committee (June 8, 2026): circular validation "when training and validation data share the same market channel." Sotheby's Provenance Intelligence trained on 4.2M confirmed auction records = auction channel only = under-sampled on sophisticated fakes that passed prior auction house review. breaks_when: > AI authentication systems are trained and validated on datasets that combine auction records with independently confirmed scientific analyses (XRF, infrared reflectography, dendrochronology) and museum collection records maintained outside auction market provenance chains. The Getty Conservation Institute's standardized authentication database, currently in development, would provide a training source with genuine market independence from auction house records. When such databases reach sufficient scale (estimated 500,000+ scientifically confirmed attributions), the circular validation risk is materially reduced because the training distribution more closely approximates the real-world population including sophisticated fakes. confidence: high source: report: "AltDaily | Art & Culture Law | 2026-06-11" date: 2026-06-11 extracted_by: Computer the Cat version: 1

- id: distributed-legal-standard-vocabulary-fragmentation domain: [ai-governance, copyright-law, regulatory-frameworks, platform-policy] when: > Multiple independent regulatory frameworks addressing AI's relationship to creative production are developed simultaneously without a shared definitional vocabulary. June 9-10 window: Copyright Office 30% de minimis threshold (measures "expressive content"). Spotify 60% AI-majority threshold (measures "musical content"). AMSRRA consent provisions (measures per-training- event consent). New York Synthetic Performers Law (measures AI-generated performer presence in paid advertising). Each framework addresses a distinct aspect of the same underlying challenge; none shares a definitional vocabulary with the others. prefer: > Before deploying a creative AI system or advising on compliance, map all applicable frameworks to a common creative workflow and identify definitional gaps: (1) Does the Copyright Office's "expressive content" measure apply to the same elements that Spotify's "musical content" measure targets? (For music, probably yes; for an AI-composed film score, possibly not.) (2) Does AMSRRA consent per training event attach at the same moment as Copyright Office registration eligibility under the 30% threshold? (A track could be 28% AI-generated—below the registration threshold requiring human contribution statement—while still requiring artist consent for the AI voice training that generated the 28%.) (3) Does the New York Synthetic Performers Law apply to AI-generated performers in content that accompanies paid advertising (contextual placement) or only to performers in the advertisement's creative content itself? Track definitional fragmentation as a compliance risk factor independent of any individual framework's requirements: as frameworks multiply with incompatible vocabularies, a creative workflow that is compliant under each individual framework in isolation may still create legal exposure through the gaps between frameworks. over: > Treating compliance with any single framework as establishing general compliance with the AI creative production regulatory landscape. A Spotify-compliant track (sub-60% AI) does not automatically satisfy Copyright Office hybrid work registration requirements (sub-30% to avoid human contribution statement). An AMSRRA-compliant AI music platform (with training consent) does not automatically satisfy New York Synthetic Performers Law disclosure requirements if its generated music is used in paid advertising with AI-generated performers. The frameworks are logically distinct, address different legal harms, and were developed without coordination. because: > Copyright Office guidance (June 9, 2026): 30% threshold in "expressive content." Spotify policy (June 10, 2026): 60% threshold in "musical content." AMSRRA (Senate subcommittee advance, June 9): consent measured per training event per voice or style. NY Synthetic Performers Law (effective June 9): disclosure required per advertisement per synthetic performer presence. arXiv:2606.12260 (real, June 9, 2026): both IP extremes underpowered—the distributed standard-setting the week's events represent is not the collective licensing mechanism the paper identifies as the equilibrium solution; it is fragmented unilateral standard-setting that creates legal uncertainty without resolving the underlying incentive problem. breaks_when: > Congress passes comprehensive AI creative content legislation with a unified definitional vocabulary adopted by reference in all implementing regulations—establishing a single "AI content percentage" standard, a single consent mechanism, and a single disclosure requirement hierarchy that preempts conflicting state and platform standards. Absent legislation, the definitional fragmentation will increase as additional states follow New York's Synthetic Performers Law model and platforms develop proprietary classification systems that may not align with each other or with regulatory definitions. confidence: high source: report: "AltDaily | Art & Culture Law | 2026-06-11" date: 2026-06-11 extracted_by: Computer the Cat version: 1 `