This week, AI’s integration into legal practice shifted from experimental to structural: California passed landmark attorney-specific AI regulation, Sidley Austin became the first London firm to mandate an AI training seat for all trainees, and courts continued cracking down on fabricated citations. The dual pressure of regulatory codification and professional re-skilling is reshaping what it means to practise law in 2026.
Top story: California unanimously passed a bill regulating attorneys’ and arbitrators’ use of generative AI — the most direct legislative intervention into legal practice in the US to date.
California Passes First US Law Regulating Lawyers’ AI Use
Sullivan & Cromwell / Law360 · Regulation
A bill regulating attorneys’ and arbitrators’ use of generative AI passed California’s State Legislature unanimously and is now awaiting Governor Newsom’s signature, with Newsom having until 30 September to decide the fate of over 30 AI-related measures. The bill would embed statutory duties around AI use directly into the Business and Professions Code, covering verification obligations and disclosure requirements. For legal practitioners, this marks the first time a US state has legislated specifically on how lawyers must handle AI-generated work product — a model other states are likely to follow.
Sidley Austin Mandates AI Seat for Every London Trainee
Non Billable / Legal Cheek · Strategy
Sidley Austin has overhauled its London training contract, making a four-month AI seat compulsory for every trainee — expanding the programme from four seats to six. The firm says it is the first in the City to build a dedicated AI rotation into every training contract, with trainees working in the firm’s AI Knowledge Lab on live legal assignments alongside its Knowledge Management, Research and Data & AI teams. The move signals a broader shift in how BigLaw firms are beginning to treat AI literacy as a core professional qualification rather than an optional add-on.
https://www.nonbillable.co.uk/news/sidley-overhauls-london-training-contract-mandatory-ai-seat
DC Court Strikes Brief After Counsel Used Unverified AI Search
SafeLegalAI · Risk
The D.C. Court of Appeals struck an appellee brief after counsel used Google’s generative AI search tool without verifying four nonexistent cases cited in the document. The incident adds to a rapidly growing global tally: as of 8 September 2026, 566 US court documents on generative AI show 507 fabricated-citation rulings. With courts across multiple jurisdictions now issuing standing orders on AI use and ordering remedial CLE for offending counsel, the professional risk of unverified AI outputs in filings has never been higher.
EU AI Act’s High-Risk Obligations Pushed Back to 2027–28
Regulation-AI.eu · Regulation
Following 2026 Digital Omnibus amendments, the EU AI Act’s high-risk AI obligations have been significantly extended: standalone high-risk systems now face compliance deadlines of December 2027, and those embedded in regulated products face August 2028. For legal teams advising on or deploying AI in justice, legal research, or court-facing contexts — all classified as high-risk — this delay offers a planning reprieve but not an excuse to pause compliance work. The transparency rules that came into force in August 2026 remain in effect and apply now.
Lightfield Emerges from YC to Automate In-House Legal Work
Y Combinator · Tools
Lightfield, a Y Combinator-backed legal AI startup, has launched with a proposition aimed squarely at in-house counsel: the platform connects to a company’s email, chat, and business systems to complete legal work before it even reaches the legal team, gathering context, applying company policies and past decisions, and presenting finished work for lawyer approval. The approach represents a new category of agentic legal AI that targets the triage and first-draft layer of in-house operations rather than law firm workflows. For corporate legal departments facing pressure to do more with fewer resources, tools like Lightfield signal a fundamental shift in how legal operations may be structured.