California demands information from OpenAI over incidents involving its agents

LAW · LAWYERS · CALIFORNIA · approved 30 September; effective 1 January 2027

California sets obligations for lawyers using AI

Governor Gavin Newsom approved SB 574 on 30 September. It takes effect on 1 January 2027. It does not ban AI as an assistance tool, but prevents lawyers from delegating the practice of law to it. Users must take reasonable steps to check outputs, correct errors and personally verify citations submitted to court. It also sets conditions for protecting confidential and non-public information and requires disclosure of generative-AI use in court filings. Arbitrators may not delegate their decisions to AI either.

Why it matters AI governance is moving from general professional ethics into explicit statutory duties. The broader principle travels beyond California even where the specific law does not: efficiency does not transfer responsibility for sources, confidentiality or the final professional judgement to the model.

SPAIN · EMPLOYMENT · AI GOVERNANCE · 30 September–1 October

CaixaBank’s AI pilot reaches the Labour Inspectorate before full rollout

UGT at CaixaBank has filed a complaint with Spain’s Labour Inspectorate, alleging that the bank failed to provide sufficient information about AI systems KIRA and Agente RED. The union wants to know what information is collected, what happens to employees’ voices, who can access stored data and whether the systems could be used to compare or evaluate workers. These are the union’s allegations and questions; the Labour Inspectorate has not found that CaixaBank violated the law.

CaixaBank told El País that the tools entered a limited pilot in early September, are intended to reduce administrative work and support commercial staff, and do not constitute mechanisms for monitoring, evaluating or controlling employees. The bank says it was preparing answers to UGT’s requests.

Why it matters “Pilot” does not necessarily mean “outside governance”. When AI touches employee data, work allocation or potential assessment, questions about information rights, privacy and worker participation can arise before a system reaches full production.

INFRASTRUCTURE · FINANCE · 1 October

Broadcom may lend Anthropic $42 billion to finance compute bought from the same ecosystem

According to Anthropic’s IPO prospectus reviewed by Reuters, Broadcom has agreed to provide up to $42 billion of financing for Anthropic’s infrastructure spending. The facility could finance roughly one-third of Anthropic’s $125.2 billion five-year commitment for TPU compute capacity. The debt instruments may convert into Anthropic equity. Anthropic is also expected to become Broadcom’s largest compute customer in 2027.

Anthropic itself identifies Broadcom’s combined role as hardware supplier and financing partner as a source of potential conflicts of interest, warning that Broadcom’s pricing and hardware decisions could affect its ability to secure sufficient compute.

Why it matters AI economics is becoming increasingly circular: infrastructure suppliers are helping finance the customers buying that infrastructure. That may accelerate investment, but it also creates concentration, vendor dependence and harder questions about where the financial risk ultimately sits.

BUSINESS · PRODUCTIVITY · published 1 October

AI creates value, but extending it across the business remains difficult

On 1 October, BearingPoint published a survey conducted in August 2026 among 1,050 senior executives in Europe, the United States and China. Among organisations that have implemented AI, 74% report measurable revenue or cost impact. Thirteen percent say they have scaled their initiatives fully in line with the original business case: that does not mean the rest have made no progress. These are executives’ responses, not audited financial results.

Sixty-two percent estimate that AI has released at least 10% of workforce capacity in certain functions. This is an assessment of available capacity, not evidence that the same proportion of employees is unnecessary.

For agentic AI specifically, more than three-quarters of organisations are still learning, prioritising or exploring pilots; 13% report a defined strategy with active initiatives and 10% say they are scaling agentic architecture across the enterprise.

Why it matters Access to a capable model may no longer be the main obstacle. Data quality, integration with legacy systems, organisational design, governance and what management actually does with released employee capacity may now be the real bottlenecks.

ANTITRUST · CONTENT · decision 30 September; reported 1 October

Google wins dismissal of Chegg and Penske antitrust cases over AI Overviews

US District Judge Amit Mehta in Washington dismissed Chegg and Penske Media’s antitrust lawsuits on 30 September. The companies alleged that Google used their content in AI summaries, reduced traffic to their websites and forced them to accept AI Overviews to remain indexed. The judge rejected the different antitrust theories pleaded. Among his reasons: an expectation of receiving traffic from Google does not itself constitute an agreement with the search engine.

The judge also said he was not unsympathetic to creators facing economic harm, but that antitrust law is not a substitute for legislators deciding how innovation-related economic harm should be addressed.

Why it matters The ruling does not establish that every use of content in AI Overviews is lawful or settle copyright law. It instead illustrates the difficulty of using existing antitrust doctrine to resolve the broader economic distribution problem between AI search systems and the publishers whose material feeds the information ecosystem.