AI Copyright Training

State of play.

• European courts are delivering the first binding AI training infringement rulings. Munich Regional Court held that Suno violated German copyright law by training on protected songs without a license, ordering revenue disclosure with damages to follow in a separate proceeding — one of Europe's first judicial decisions on this question (→ German court finds Suno’s AI music training infringed copyright).
• US publisher litigation against Google and Meta is escalating the stakes. Hachette, Cengage, Elsevier, and Scott Turow have sued Google over Gemini training; the same publishers previously sued Meta over Llama — with courts applying a fact-specific market-harm test rather than categorical rules, and training on pirated content creating near-certain liability .
• The Third Circuit is positioned to issue the first US federal appellate ruling on AI training fair use. The panel pressed ROSS and Thomson Reuters hard on market harm and transformativeness at oral argument; a decision is expected before year-end .
• Anthropic is simultaneously litigating fair use and settling at scale. Its transformative-use motion remains pending in California while the $1.5 billion Bartz class settlement — with over 100,000 claimants — moves toward final approval .
• For counsel advising AI developers, publishers, or rights holders, the practical baseline is a multi-jurisdictional liability picture: US courts are converging on a market-harm test, a German court has found infringement, and the Third Circuit will shortly produce the first appellate precedent — meaning the window for a settled fair use defense is narrowing.

Where things stand.

• US courts are applying a fact-specific market-harm test, not a categorical rule. Bartz v. Anthropic found training on legally acquired books qualifies as fair use; courts have similarly favored Meta in a separate decision; but training on pirated or unlicensed content creates near-certain liability .
• The Third Circuit's Thomson Reuters v. ROSS Intelligence ruling will be the first federal appellate decision on AI training fair use. The panel's questioning at oral argument focused on market substitution, Thomson Reuters' exclusive training rights, and the existence of a licensing market for headnotes — all factors that cut against ROSS .
• Bartz v. Anthropic is the leading settlement benchmark. The $1.5 billion class settlement covers claims from over 100,000 authors and rights holders; Anthropic's parallel transformative fair use motion — analogizing AI ingestion to human learning — remains pending before the same court .
• Publisher registration failures are surfacing as a litigation complication. Works ingested without proper pre-suit registration create standing and damages gaps that complicate both plaintiff and defense strategy .
• The UK has abandoned its proposed AI copyright opt-out exception. The UK government's published report signals a policy reversal, leaving rights holders and AI developers without a clear statutory training-data framework in that jurisdiction .
• Australia is legislating a creator-control copyright framework as part of a broader AI data center regime. The Albanese government's framework — to be legislated by early 2027 — includes a copyright component giving Australian creators control over pricing when their work is used in AI training, alongside mandatory energy and water standards for hyperscale facilities .
• Indonesia is advancing a copyright rewrite that would require AI platforms to pay for training use. The draft bill mandates compensation through state-supervised collective management organizations, bans AI from imitating a creator's distinctive style, and threatens operating license revocation for non-compliance — with Google already pushing back .
• PAE acquisition of orphaned AI-sector patents is a structural exposure vector. Failed AI startups leave foundational patents that migrate to assertion entities, which then target the companies that successfully commercialized the underlying technology — a pattern intensifying as secondary patent markets mature .

Latest developments.

• Munich Regional Court ruled against Suno in GEMA's suit, finding training on protected songs without a license violated German copyright law; damages to be calculated separately; Suno has indicated it will appeal (→ German court finds Suno’s AI music training infringed copyright).
• Hachette, Cengage, Elsevier, and Scott Turow sued Google in New York federal court over Gemini training; the same publisher group had previously sued Meta over Llama .
• Third Circuit heard oral argument in Thomson Reuters v. ROSS Intelligence, pressing both parties on market harm, transformativeness, and the existence of a licensing market for headnotes; ruling expected late 2026 .
• A $2 million publishing deal for debut novelist Jerry Falade collapsed after agents withdrew over AI authorship suspicions, with no standardized manuscript provenance process in place; Falade has denied AI use .
• Australia announced a mandatory AI framework including a creator-control copyright component for AI training, to be legislated by early 2027 .
• Indonesia's House of Representatives is advancing a copyright rewrite requiring AI platforms to compensate publishers through collective management organizations, with operating license revocation as the enforcement mechanism .
• DOJ established an AI Litigation Task Force formalizing AI's role in federal legal operations; recent rulings involving Anthropic establish that training data must be obtained legally, barring use of pirated content (→ DOJ Establishes AI Litigation Task Force as Courts Adapt AI Discovery Tools).

Active questions and open splits.

• Whether the Third Circuit will adopt a market-harm-centered fair use standard for AI training. The panel's oral argument focus on substitution and the existence of a headnote licensing market suggests the court may affirm the non-transformative finding — which would be the first appellate precedent constraining AI training fair use claims .
• Whether the human-learning analogy for AI training survives judicial scrutiny. Anthropic's pending fair use motion is the most direct test of this framing; no court has accepted or rejected it, and the Third Circuit's reasoning in ROSS will bear directly on how the California court rules .
• Whether the GEMA v. Suno ruling survives appeal and how damages are calculated. If the Munich decision holds, it establishes a European liability baseline and will accelerate licensing pressure on AI music firms operating in the EU — and potentially influence how UK and Australian frameworks develop (→ German court finds Suno’s AI music training infringed copyright).
• Whether Bartz settlement approval sets a damages benchmark. The $1.5 billion settlement, if approved, will be the first significant reference point for aggregate AI training copyright damages — shaping every subsequent negotiation and litigation posture .
• How AI authorship suspicion translates into contractual and due diligence obligations. The Falade deal collapse demonstrates that AI-authorship questions are now material deal-breakers without standardized detection or provenance protocols — leaving publishers, agents, and entertainment lawyers without an industry-standard framework .
• Whether Indonesia's and Australia's mandatory compensation regimes create compliance obligations that conflict with US fair use positions. Platforms operating in multiple jurisdictions face a structural tension: training practices defensible under US law may trigger liability or license revocation in Southeast Asia and Oceania .
• UK training-data framework post-opt-out reversal. With the broad exception abandoned, the path forward — licensing regime, narrower exception, or continued gap — remains genuinely open, with significant implications for cross-border AI development .

What to watch.

• Third Circuit ruling in Thomson Reuters v. ROSS Intelligence — the first federal appellate decision on AI training fair use, expected late 2026; the market-harm framing at oral argument is the signal to track.
• Bartz v. Anthropic fairness hearing — whether the settlement is approved and what damages framework the approval order articulates.
• Suno's appeal of the Munich Regional Court ruling — whether the German decision survives and how damages are calculated will define the EU liability baseline.
• Indonesia's copyright bill progression — whether the House sets a timeline for final deliberation and whether the operating-license revocation mechanism survives tech industry pushback.
• Australia's legislative drafting process for its AI copyright framework — compliance mechanisms and enforcement timelines remain undisclosed ahead of the early 2027 target.
• Additional publisher or author suits against Google and Meta as discovery proceeds — what the record reveals about training data provenance and whether piracy allegations surface.