Thelonious ICYMI — Edition 12 — 25 June 2026
Thelonious ICYMI — Edition 12 — 25 June 2026
Welcome to the 12th edition of ICYMI — your weekly briefing on the developments shaping the AI legal and regulatory landscape, curated from Thelonious.
Here’s a snapshot of what’s inside this week:
- Ireland’s Regulation of Artificial Intelligence Bill 2026 — The Bill would establish a national AI authority to implement and enforce the EU AI Act (Regulation (EU) 2024/1689) domestically, and amends a series of existing Irish statutes — central banking, communications, consumer protection and freedom of information — to align them with the AI Act’s harmonised rules. The AI Act is a Regulation, directly applicable across the Union, which is precisely why the national implementing measures get so little attention. They should get more. The Act tells you what is prohibited; the national plumbing tells you who knocks on the door, under which powers, and how long it takes. Compliance is not decided by the text of the Regulation. It is decided by the office that enforces it.
- Das v. Meta, Google, OpenAI & X — The plaintiff alleges that the defendants’ AI systems unlawfully ingested, reproduced and commercially exploited two of his copyrighted works for training purposes without authorisation. Most of the training-data litigation to date has come from organised plaintiffs: publishers, record labels, class actions with institutional backing behind them. This is one author naming four of the largest developers in a single complaint. Whether it survives the pleadings is a separate question from what it signals — the barrier to filing has dropped to roughly the cost of a filing fee.
- Jamendo v. Nvidia — Jamendo alleges that Nvidia used the MTG-Jamendo dataset, containing over 55,000 audio tracks, without permission to train its AI audio models Fugatto and Audio Flamingo. The claim seeks damages and an injunction. The detail worth marking is the nature of the source. This is not an allegation about scraping the open web, where the licensing position is contested and the terms are often nowhere. It concerns a curated research dataset with terms attached — written down, published, capable of being read before use. The gap between “publicly available” and “licensed for commercial training” has been the industry’s most comfortable ambiguity. It is a much harder place to sit when someone can point at the licence.
- Shakespeare v. Anthropic — More than 100 authors who opted out of a prior class action settlement have filed a new copyright infringement suit against Anthropic and its executives, alleging the defendants used BitTorrent to unlawfully download, distribute and retain copyrighted books from pirate libraries including Library Genesis and Books3, and used those works to build a permanent internal library for model training and other uses. Settlements are usually the end of the story. Opt-outs are the part everyone stops reading, and this one demonstrates why they are worth reading. Note also what the complaint targets: acquisition rather than output. The alleged wrong is how the library was obtained and retained, not what the model went on to produce — a distinction that has quietly become the most consequential fault line in AI copyright litigation.
- Bipartisan Kids Safety Deal in Congress — A bipartisan agreement out of the House Energy and Commerce Committee, the Kids Internet and Digital Safety (KIDS) Act consolidates over a dozen separate bills into a single vehicle: safety-by-default settings, strengthened privacy rules for children and teenagers, greater transparency from data brokers, expanded parental controls, and accountability obligations for technology companies. Consolidation is the tell. A dozen competing bills is a signal of interest; one merged bipartisan vehicle is a signal of intent. Children’s online safety remains the single lane in which US federal technology legislation actually moves, and the obligations drafted in that lane rarely stay confined to products aimed at children.