Gather — 18 July 2026

Publisher Litigation Reaches Google #

• Google faces another AI training lawsuit from major publishers (TechCrunch) — Hachette, Cengage, Elsevier, author Scott Turow, and S.C.R.I.B.E. filed a class action against Google in the SDNY on July 14–15, 2026, alleging Gemini was trained on copyrighted books obtained via Google Books/Google Play (programs licensed for narrower purposes) plus pirate-site and paywalled material. The complaint cites an internal Google document allegedly warning the practice was “highly problematic” and could trigger “$10Bs-$100Bs in potential fines” — a pre-litigation admission surfacing directly in a filing rather than discovery. This is largely the same publisher coalition that sued Meta on May 5 (Elsevier, Cengage, Hachette, plus Turow), now running a near-identical legal theory against a second AI company, filed in SDNY rather than California — escaping the persuasive weight of the pro-fair-use Kadrey v. Meta ruling and landing in the same district as NYT v. OpenAI/Microsoft.

AI Music Training: Munich Verdict Due July 31, Boston Scope Fight Continues #

• GEMA, Suno copyright ruling postponed by Munich court to July 31 (MLex) — The Munich Regional Court’s 42nd Civil Chamber pushed its verdict in GEMA v. Suno from June 12 to July 31, 2026 for internal administrative reasons. At the March 9 hearing, Judge Elke Schwager had both original recordings and Suno-generated outputs played aloud in court for six contested works (“Rasputin,” “Daddy Cool,” “Mambo No. 5,” “Forever Young,” “Atemlos,” “Big in Japan”). This follows the November 2025 GEMA v. OpenAI Munich ruling already tracked in this journal — a second Munich decision on AI-music training within a year would consolidate Germany as the most active European jurisdiction on AI copyright, arriving five weeks before the CJEU’s own September 3 Advocate General opinion.
• Why a fight over 61,000 recordings could shape the future of AI music licensing (Music Business Worldwide) — In the Massachusetts Sony/UMG v. Suno case, Judge F. Dennis Saylor IV is weighing whether to let the labels expand their claim from 560 tracks to 61,026 recordings identified via Audible Magic audio-fingerprinting of Suno’s training data. The exposure swings sharply on the ruling: at the $150,000-per-work statutory maximum for willful infringement, 560 works caps exposure near $84M; 61,026 works would exceed $9B. Fact discovery closes September 30, 2026, and dispositive (summary judgment) motions on the fair-use merits aren’t due until April 2027 — later than several secondary sources suggested this cycle — meaning Munich will rule on the underlying training question well before Boston reaches it.

Australia Rejects the Opt-Out/Carve-Out Model #

• Australia vs. AI: Prime Minister Promises Local Creatives ‘Strongest Possible Protection’ Against Copyright Theft (TheWrap) — At the University of Sydney on July 15, 2026, PM Anthony Albanese declared “no company should use Australian books, music, art or news to build or train AI” without creator control, calling unlicensed training “theft,” and announced a new Office of AI (within the Department of Prime Minister and Cabinet) tasked with drafting legislation for introduction to Parliament in early 2027, with a framework going to National Cabinet in August 2026. This follows Australia’s Attorney-General ruling out a text-and-data-mining exception in October 2025 — Australia becomes the first jurisdiction outside the US/EU/UK triad this journal has tracked to take an explicit consent-first policy position, rejecting the UK’s now-abandoned opt-out model outright rather than attempting and retreating from it.

Data Provenance as Deal and Deployment Risk #

• Synthetic Data as a Deal Asset: Ownership, Provenance, and Diligence Considerations in AI Acquisitions (Mayer Brown) — Practitioner guidance on treating synthetic data as an M&A asset: because synthetic outputs can inherit infringement liability from the generating model’s own training data, diligence must trace the full upstream chain — source data, foundation-model terms, and whether the target tested for model collapse — not just the synthetic layer itself. Synthetic data’s copyright status remains unresolved (protection requires “sufficient human expressive contribution” per Copyright Office guidance), so most protection today rests on trade-secret status requiring documented access controls. This formalizes what was informal practice: synthetic data doesn’t launder away training-data provenance risk, it adds a layer buyers must diligence separately.
• Legal Implications in AI Development and Deployment: Training Data and Grounding Data (Sidley Austin) — Distinguishes training-data risk (one-time, at model-build stage) from grounding-data risk (recurring, at query time via retrieval-augmented generation) — a distinction largely absent from this journal’s prior coverage, which has focused almost exclusively on training-time liability. Grounding data creates infringement risk “at the moment of each query, not just during development,” plus continuous-access dependency (losing a data feed can break a live system) and terms-of-service exposure from real-time scraping or API use. This maps directly onto the “substitutive summary” output-liability line tracked since the May 22 gather — RAG and AI-search products are exactly where grounding-data risk and output-liability risk meet.

Cross-links #

• AI Impact on Society Australia’s new Office of AI and consent-first training-data framework (legislation targeted for early 2027) is a new sovereign AI-policy data point alongside the EU/Colorado/GAAIA timeline already tracked there.
• Claude Integrations Sidley’s training-vs-grounding-data distinction bears directly on RAG-based integrations — query-time infringement risk applies to any product performing retrieval over external content at runtime, not just to models trained on copyrighted works.
• Open vs Closed AI Ecosystems The same publisher coalition (Hachette, Cengage, Elsevier, Scott Turow) has now sued both Meta (open-weight) and Google (closed) with near-identical legal theories — publisher litigation risk is being applied uniformly regardless of open/closed status, distinct from the sourcing-liability asymmetry flagged in earlier gathers.

Meta-observations #

• Emerging pattern: The same publisher coalition is now running a sequential multi-defendant litigation strategy — Meta in May, Google in July — rather than each suit being an independent event. Watch for the same plaintiffs naming further defendants (Amazon, xAI) in the coming weeks.
• Method note: Several secondary/aggregator sources this cycle characterized a July 2026 “Sony Music v. Suno summary judgment hearing,” but docket-level reporting (Music Business Worldwide) shows dispositive motions aren’t due until April 2027 — the July court activity is a scope-of-evidence dispute (560 vs. 61,026 works), not a fair-use ruling. Verify date claims against primary litigation trackers before citing them.
• Emerging theme: Grounding data (RAG/query-time retrieval) is surfacing as a legal risk category distinct from training data, with its own liability profile — continuous, per-query exposure rather than one-time training exposure. This journal has tracked “substitutive summary” output liability since May but not yet the retrieval-time mechanics Sidley’s piece formalizes; worth tracking as its own thread going forward.
• Source to watch: MLex (GEMA v. Suno postponement) and Music Business Worldwide (Sony/UMG v. Suno scope fight) both delivered sharper, date-accurate detail than generic “AI copyright tracker” aggregator sites this cycle — prioritize primary litigation-beat reporting over roundup sites.