Thelonious ICYMI — Edition 18 — 6 August 2026
Thelonious ICYMI — Edition 18 — 6 August 2026
Welcome to the 18th edition of ICYMI — your weekly briefing on the developments shaping the AI legal and regulatory landscape, curated from Thelonious.
A week in which the question stopped being whether AI systems infringe and became who answers for it. A German court put the provider on the hook rather than the user. Brussels put disclosure duties on providers and deployers alike. A stockholder went after a board. Read together, responsibility is climbing the chain.
Here’s a snapshot of what’s inside this week:
- GEMA v. Suno Inc. — On 31 July 2026 the 42nd Civil Chamber of the Landgericht München I ruled largely in GEMA’s favour in case 42 O 763/25, granting injunctive relief, disclosure and a declaration of entitlement to damages against the AI music service. The court found that six well-known works, among them Rasputin and Mambo No. 5, were reproducibly contained in Suno’s v3.5 and v4 models. Suno has said it disagrees with the decision, and the judgment is not final. Three findings matter more than the headline. Training conducted in the United States was held capable of infringing where there is a sufficient link to a German claim, which removes offshore compute as a structural defence. Memorisation within the model was treated as a reproduction in its own right, meaning the model is the infringing artefact and not merely the pipeline that produced it. And liability was placed on Suno rather than on the users typing the prompts, on the reasoning that Suno chose the training data and controls the architecture. The chamber also reaffirmed the narrow reading of the text-and-data-mining exception it applied against OpenAI in November 2025. Two rulings from the same chamber in nine months is a position, not a coincidence, and Munich is now the most consequential venue in European AI copyright.
- EU AI Act Article 50 — The transparency obligations became applicable on 2 August 2026, as flagged in Edition 17. Providers of systems that interact directly with individuals must disclose the fact; providers of generative systems must mark outputs in machine-readable form; deployers must label deepfakes and disclose the use of emotion recognition and biometric categorisation. Penalties reach €15 million or 3% of worldwide annual turnover. The Commission adopted guidelines on 20 July and has published a voluntary Code of Practice on Transparency of AI-Generated Content, signature to which carries a degree of presumption of conformity. Two points of precision are being lost in client communications. The obligations apply to in-scope systems regardless of when they were placed on the market, and the only transitional relief runs to 2 December 2026 for the marking and detection duty under Article 50(2), and only for systems already on the market. Everything else applied on Sunday. Anyone who read the Omnibus as a general reprieve has misread which obligations moved.
- EVOX Productions LLC v. Midjourney, Inc. — EVOX filed on 30 July 2026, alleging that Midjourney trained its image model on more than 100,000 of its copyrighted automotive photographs, built a product that competes directly with EVOX’s licensing business, and stripped copyright management information including watermarks. The allegations are untested. The filing is the third EVOX has brought in this space, following an action over dataset redistribution against Stanford and a July claim against Stability AI, Runway, Hugging Face and DeviantArt, and the pattern is the point. EVOX shoots standardised studio images to a repeatable formula, registers them individually with the Copyright Office, and licenses them at a published rate. That combination produces something most AI copyright plaintiffs cannot assemble: a functioning licence market with an identifiable price, which is exactly the evidence the fourth fair use factor demands. The CMI allegation compounds it, because a claim under section 1202 turns on the removal of identifying information rather than on the character of the use, and fair use is not an answer to it. The plaintiffs who trouble model developers are not the ones with the largest catalogues. They are the ones who can show what the work was worth.
- Berliner v. Huang — A verified stockholder derivative action filed on 31 July 2026 in the Northern District of Illinois names NVIDIA’s chief executive, senior officers and directors. The complaint alleges that the defendants knowingly permitted training on pirated books, scraped video and biometric voice data obtained without consent, and then made false and misleading statements in SEC proxy filings that sustained the share price. The allegations are untested and the pleading bar for this species of claim is severe: Delaware’s courts have repeatedly held that imperfect compliance does not establish conscious disregard. What makes the filing worth marking is the mechanism. The complaint leans on the company’s own proxy language about delivering trustworthy models that comply with privacy and data protection law, which converts a governance statement into an evidentiary problem. Copyright exposure is beginning to metabolise into securities and fiduciary exposure, and the route runs through the disclosure. For in-house teams the practical consequence is immediate. Whoever drafts the AI paragraph of the annual report should be reading the underlying dataset provenance records, and should be able to name them.
- Tesla, Inc. v. Granite Vehicle Ventures LLC — Tesla filed a petition for certiorari on 27 July 2026, docketed 29 July as No. 26-136, after the Federal Circuit denied mandamus in February. The question is whether 35 U.S.C. § 314(d) insulates the USPTO’s inter partes review institution criteria from any judicial testing against the statute, and Tesla contends the “Fintiv time-to-trial” criterion is an extra-statutory invention the agency had no power to adopt. It is the fourth such petition now pending. The temptation is to file this under patent procedure and move on, which would be a mistake for anyone advising on AI. The petition’s central complaint is that an agency has built a body of screening rules with real commercial consequence, without notice-and-comment rulemaking, and beyond the reach of any court. Every regulator now assembling discretionary AI frameworks is watching how much room that leaves. If the Court takes the case, the holding will travel well outside the PTAB.