Thelonious ICYMI — Edition 22— 3 September 2026
The ChatGPT Digital Services Act designation was the defining development of the week, and it illustrates this edition’s central theme: artificial intelligence is being governed by law that already existed.
The Digital Services Act was drafted in 2022 for social platforms and now covers a general purpose assistant. A 2018 child protection statute is being used against an image generator. Section 3LA of Australia’s Crimes Act 1914 appears on a fourteen count indictment concerning compromised AI tooling.
Copyright law is doing what it has done since the first training data claims were filed. Only one item this week is a purpose built AI instrument, and it passed the California Senate without a single vote against.
Where the previous edition tracked the pricing of AI’s inputs, this one tracks the statutes doing the enforcing.
- European Commission Designations Under the Digital Services Act
On 31 August 2026 the European Commission designated ChatGPT a Very Large Online Search Engine, and Reddit and Roblox Very Large Online Platforms, under the Digital Services Act. Each service declared at least 45 million average monthly users in the EU, the threshold that triggers designation. The three now have four months, until January 2027, to meet additional obligations covering systemic risk assessment and mitigation across their algorithmic systems, content distribution and service design, with named categories for illegal content, effects on minors, users’ physical and mental wellbeing, fundamental rights, electoral integrity and public security. The full press release sets out the designation criteria.
The classification of a general purpose assistant as a search engine is the material development. It brings ChatGPT inside a regime built for information intermediaries, complete with independent audit and researcher data access duties, without any new AI specific instrument being drafted. For advisers, the first question about a large consumer facing AI service in the EU is now a user count rather than a capability description, which is precisely the kind of shift regulatory landscape tracking exists to catch before a client asks about it.
- Sony Music Publishing and Warner Chappell Music v. Anthropic
Filed on 28 August 2026 in the U.S. District Court for the Northern District of California, the complaint names Anthropic together with its co-founders Dario Amodei and Benjamin Mann, and alleges a campaign of torrenting, scraping and downloading copyrighted musical compositions, lyrics and sheet music in order to train the Claude models. Anthropic settled the authors’ class action, Bartz v. Anthropic, for $1.5 billion (£1.1 billion) in 2025 and faces a separate claim from Concord and Universal Music Group filed in January 2026. TechCrunch reported the filing the following day.
The pleading against individual founders is what distinguishes this claim from its predecessors. Corporate liability for the acquisition of training data is now well litigated. Personal liability for the decision to acquire it is not, and if that count survives a motion to dismiss it alters the risk position of every executive who has approved a dataset, including retrospectively. We argued a version of this in We Are All Thieves, before the pleadings caught up.
- Doe 1 v. xAI
A class action filed on 27 August 2026 in the U.S. District Court for the Northern District of California alleges that xAI built and operated Grok’s image generation features without sufficient safeguards, and that the resulting system produced unlawful synthetic child sexual abuse material and non-consensual sexualised imagery. The named plaintiff is a survivor whose historic abuse material carries hash values tracked by the FBI. The claim is brought under the federal civil remedy provisions available to victims of child exploitation, commonly known as Masha’s Law. CyberScoop has reporting on the complaint.
The instrument combines a training data claim with a product safety claim, and it does so under a 2018 statute rather than anything written for generative models. Because the pleaded measure of liability is the adequacy of the defendant’s safeguards relative to industry practice, internal safety documentation, model cards and red team findings become discovery material. Model developers should assume that comparison will be drawn, and that the quality of their own human verification records will be part of it.
- Sovereign AI R&D Procurement Scheme, United Kingdom
The UK government announced the scheme and opened its first competitions on 31 August 2026, with £100 million (about $135 million) committed. Four challenges are live: NHS productivity and clinical decision support, compute efficiency, integrating AI across defence mission environments, and agent security and resilience testing. The scheme purchases demonstrator stage technology, meaning solutions ready for testing in real operational environments rather than research. Upfront payments are available so that smaller firms can compete, and successful suppliers keep the intellectual property they create.
The IP retention term carries the commercial weight here. Government is positioning itself as an early customer rather than a grant funder and leaving ownership with the supplier, which turns each contract into a reference customer and a licensable asset rather than work for hire. Across four competitions, the difference between retained and assigned intellectual property is the difference between a pilot and a company, and the model terms are worth reading before a bid rather than after one.
TeamPCP Charges, Australia
Louis Michael Gaebler, 23, and Ruben Ian Thomson, 21, both of Perth, were charged on 27 August 2026 following an Australian Federal Police investigation conducted with the Western Australia Police Force and the FBI. Both are alleged to have held key roles in the TeamPCP cybercrime syndicate. They face 14 charges between them, including dealing with proceeds of crime of A$100,000 or more, possessing, supplying and unlawfully modifying data with intent to commit a serious computer offence, and failing to comply with an order under section 3LA of the Crimes Act 1914 to surrender device passwords. The Record has the charge details.
The alleged campaign compromised open source developer tooling including TanStack, Trivy and LiteLLM, a Python library widely used to route calls between applications and AI models, with downstream effects reported at the European Commission and GitHub. A compromise at that layer is a compromise of the connection between an organisation and its models, which makes this a third party risk matter rather than an AI matter. Most AI governance registers do not record the dependency at all, which is one reason Thelonious tracks enforcement alongside regulation.
- California SB 1119, Adam’s Law
The California Legislature passed SB 1119 on 31 August 2026, by 39 votes to nil in the Senate and 64 to 4 in the Assembly, and forwarded it to Governor Newsom. The bill is named after Adam Raine, a California teenager who died in 2025. Its requirements include age determination, safety risk assessments before new chatbots are released, independent audits, precautions against harmful content, default parental controls, in app crisis support with parental notification, incident reporting to the Attorney General, restrictions on advertising to minors and privacy protections. OpenAI made its support public on 31 August and urged the Governor to sign the legislation into law. Senator Padilla’s office published the vote detail.
An incumbent endorsing a binding safety statute on the day it passes is a statement about cost. Age determination, pre release risk assessment and independent audit are affordable at OpenAI’s scale and are a material barrier at a startup’s, which means the compliance burden will fall unevenly whether or not that was the intention. The near unanimous vote removes any argument that the measure is partisan, and the text should now be read as a template for other state legislatures. Thelonious Research will track its progress through the states.
Six developments, five statutes, and only one of them written for artificial intelligence. If you would like to see how this reads for your own jurisdictions and practice areas, request a Snapshot and we will send you a one page exposure brief.