A privacy notice is not glamorous. It is also where AI governance tends to become real: not in keynote decks, but in the text users actually see, the rights forms they can actually use and the audit trail a regulator can actually inspect. The UK Information Commissioner’s Office has now said ten major AI foundation model developers have made, or committed to make, changes to how they handle personal data. That is less a victory lap than a useful compliance weather report.

What the ICO says changed

According to the ICO’s 8 October 2026 announcement, ten leading AI foundation model developers made or committed to data protection improvements after supervision, and the regulator is extending scrutiny to agentic AI. The Next Web reported that the named developers are Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI. The same report said the changes cover clearer information on how people’s data is used, stronger ways to exercise data rights and tougher checks on developers’ safeguards.

@title ICO AI supervision path
@source The Next Web
@source ICO

  6 Nov 2024 ··· recruitment intervention
  2025 ··· foundation model programme
  March 2026 ··· biometrics strategy update
  8 Oct 2026 ··· developer changes
  8 Oct 2026 ··· agentic AI call

@caption The ICO path moves from sector audits to foundation models and agents.

The language matters. The ICO is not saying every promised change has already landed in product, and The Next Web reported that the regulator is monitoring whether companies deliver. Richard Nevinson, the ICO’s director of technology regulation, put the regulator’s version this way: “Our engagement with some of the biggest developers has secured real commitments that will help people better understand and control how their data is used, even in a fast-moving and complex area.” In plain English, a model provider’s privacy story now needs to survive more than a terms of service footnote.

Why this is not just a foundation model story

The timeline behind the announcement shows the ICO moving from sector audits into the model layer. In November 2024, the ICO said its intervention into AI recruitment tools produced almost 300 recommendations, including processing personal information fairly, keeping it to a minimum and clearly explaining to candidates how AI tools would use their information. The companies in those audits accepted or partially accepted all recommendations, according to the ICO.

The Next Web reported that the ICO set up a programme in 2025 to supervise foundation model developers, choosing 11 companies based on risk of breaking rules, UK market share and use of higher-risk training data. Elon Musk’s xAI was the eleventh, but the ICO paused that work after opening a formal investigation into the Grok chatbot, which The Next Web said remains under way. That sequencing is the interesting bit for builders: the regulator is not waiting for one perfect AI statute before asking whether existing data protection law is being observed.

What builders should translate into controls

The ICO’s own AI and data protection guidance says its guidance is under review following the Data (Use and Access) Act, but the current direction is not mysterious. The guidance discusses transparency in AI and reflects industry requests to clarify fairness requirements. The practical translation is dull and useful: know what personal data enters training or tuning workflows, explain that use in language a person can understand, keep a route open for data rights requests and document why your safeguards are adequate.

For companies buying foundation models rather than training them, the obligation becomes contractual before it becomes philosophical. If your product uses one of these developers through an API, your vendor review should ask how personal data is used, what rights workflows exist, what safeguard checks have changed and how future commitments will be reported. If the answer is a blog post and a shrug, that is not a control. It is a screenshot waiting to become an exhibit.

Why agents are the next inspection aisle

The ICO said in its 8 October 2026 announcement that it is launching a call for evidence on agentic AI and confirmed enquiries around recent agentic AI testing and deployment earlier this year. That is the natural next place to look. A chatbot can mishandle personal data, but an agent can be designed to fetch, infer, transfer or act on information across services, which makes notice, consent, minimisation and rights handling harder to keep in one tidy box.

The Next Web also reported that the ICO believes current training practices still make it hard for developers to comply with UK data protection law, and that the regulator is raising those problems with government. That is not a free pass for product teams. It means the hard cases are still hard, while the easier cases, clearer notices, usable rights routes and recorded safeguard checks, are now table stakes for anyone who wants to operate in the UK without treating privacy as a decorative appendix.

The next useful signal will be what the ICO asks in its agentic AI evidence work, and whether monitoring of the ten developer commitments turns into public follow-up. Builders should read this as a preview of expected controls, not a finished checklist. Learners and users should read it as a reminder that data rights still apply when the interface looks conversational and the system underneath looks complicated.

Sources