Thelonious ICYMI — Edition 21 — 27 August 2026
This weekly briefing on AI legal and regulatory developments frames the key theme: the inputs are being priced.
Artificial intelligence has been built on three things it has never had to pay for: attention, data, and electricity.
Edition 21 tracks the week each of the three had a number attached, from an $18 billion settlement in the United States to mandatory grid obligations in Australia.
Lead Stories
- Meta social media addiction settlement — Meta agreed on 26 August to pay up to $18 billion resolving state claims that Facebook and Instagram were engineered to hook minors. The structure: $17.6 billion to 48 states, Washington DC, Puerto Rico and the US territories, plus $459 million settling outstanding Cambridge Analytica privacy claims. California receives $2.2 billion, with New York and Texas each above $1 billion. New Mexico and Florida declined to settle. The federal trial had opened on 18 August, with Instagram head Adam Mosseri already testifying and Mark Zuckerberg expected to follow. The settlement ended proceedings before a verdict and Meta denied wrongdoing throughout. Three points carry the significance: the injunctive terms bind conduct rather than speech, capping teen use at two hours daily and blocking access between midnight and 6am without parental consent; Meta conceded nothing on personalised recommendations or targeted advertising; and part of the payment is contingent on YouTube and TikTok adopting comparable protections. This is the first hard contractual limit any platform has accepted on how much of a minor’s day it may take, and it was secured through consumer protection law rather than content regulation.
- wikiHow, Inc. v. OpenAI, Inc. — Filed 24 August 2026 in the US District Court for the Southern District of New York (1:26-cv-07171), alleging that OpenAI downloaded more than 11,000 tutorial articles, at least 1,211 of them registered copyrights, to train its GPT models and to run runtime retrieval systems without permission. The complaint is built as a chain rather than a single act: acquisition, training, retrieval, output, commercial substitution, and removal of copyright management information. The distinguishing feature is the emphasis on output. wikiHow alleges that ChatGPT reproduces its guides closely enough to remove any reason to visit the source, and that attribution is stripped along the way. Substitution is the fourth fair use factor and historically the one that decides these cases, which places this claim on firmer ground than an input-only argument.
- Warren Pandiscia v. Twitch Interactive, Inc. and Amazon.com, Inc. — A class action filed 20 August 2026 in the US District Court for the Northern District of California. On 12 August, Twitch automatically enrolled every user in a programme permitting Amazon to use their streams and chat logs for generative AI training, with an opt-out rather than an opt-in. The complaint alleges collection running since at least 2024 without notice or consent, and use of that material to train Amazon’s commercial models including Nova Reel. Causes of action are breach of express and implied contract, unjust enrichment and unfair business practices, with injunctive relief, damages, restitution and disgorgement sought. The central allegation concerns capacity to consent: a livestream chat carries contributions from thousands of people who agreed to nothing, and the defendants “never obtain, and their systems are incapable of obtaining, the consent of all parties to the communications they capture.” Twitch’s chief product officer has stated publicly that the platform has “not been training models.” The exposure here originated in a settings change, not a model.
- UK and Ukraine AI defence partnership — Signed 24 August 2026, granting the United Kingdom access to Ukraine’s Avengers AI Labs and the battlefield data underpinning it. The platform aggregates millions of real-world sensor feeds, to be used in training models for defence, national security and critical infrastructure protection. The significance lies in the contrast with the week’s litigation: the same commodity, live operational data at a scale no other party holds, acquired by negotiation between states and treated as an asset with a value rather than taken and defended afterwards. Governments are beginning to price training data as sovereign property, and that reframing is unlikely to remain confined to defence procurement.
- Australian Standards for AI — The Australian government has established a national framework for AI investment overseen by a newly formed Office of AI within the Department of the Prime Minister and Cabinet, replacing a patchwork of state approval processes. Mandatory obligations fall on large-scale data centre operators: acting as net-generators of power by returning as much energy to the grid as they consume, paying the full cost of grid connections rather than passing it to consumers, maximising energy efficiency, minimising water usage and funding additional water infrastructure where required. Agreement is being sought at National Cabinet, with legislation targeted for early 2027. Where most jurisdictions have approached AI as a question of speech, safety or copyright, Australia has approached it as a question of electricity and water. Parties to long compute or hosting agreements should assume the cost base moves.
