Ten AI Companies Promised Better Privacy in the UK
Matthew Leo · Published October 9, 2026 · AI
The United Kingdom’s privacy regulator says ten large foundation-model developers have changed, or committed to change, how they handle personal data. The list includes Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI.
The Information Commissioner’s Office announcement is important for what it does and does not claim. It does not say every promised change is finished, and it does not certify the companies as compliant. The regulator says it is monitoring progress.
What the companies agreed to improve
The ICO grouped the work into three practical areas: clearer explanations of how personal data is used, stronger ways for people to exercise privacy rights and tougher assessments of safeguards.
Its accompanying supervision report sets a more specific test for model developers. When training uses personal data, developers need a lawful basis, meaningful transparency, workable rights processes and evidence that safeguards materially reduce risk. The ICO says the bar is higher when a model may contain personal data or training involves sensitive categories.
Those are outcomes, not product features. A longer privacy notice is not enough if a person still cannot discover what data was used, challenge it or receive a practical response.
Why this matters in Canada
Many of the same companies sell services to Canadians, even though the UK regulator’s authority comes from UK law. Changes to a global account portal, training-data request process or privacy control can still affect users outside Britain if the company deploys the same system broadly.
Canada’s federal, provincial and territorial privacy regulators have published their own principles for privacy-protective generative AI. They similarly emphasize legal authority, transparency, data minimization, access and correction rights, accountability and privacy by design.
The useful Canadian comparison is therefore not whether UK rules apply here. It is whether developers can turn the same underlying privacy obligations into controls people can actually use, and whether Canadian regulators can verify those controls with evidence.
The regulator is already moving to agents
The ICO also opened a six-week call for evidence about systems that can complete tasks, use external services and act with limited human oversight. The consultation runs from October 8 to November 20, 2026, and asks about security, transparency, accountability, automated decision-making, fairness, purpose limits and lawful processing.
That shift follows the same risk boundary raised by the U.S. FTC’s investigation of AI agents. Training data remains important, but an agent can create a second set of privacy questions when it reads email, browses websites, calls services or takes an action on a person’s behalf.
The ICO says it has made enquiries with OpenAI, Anthropic, Meta and the UK’s AI Security Institute about recent agent testing and deployment. The announcement refers to reported cases in which agents bypassed protections, used unauthorized communication channels or accessed external systems. Those reports are being examined; the regulator has not announced findings against the companies on those points.
What users and organizations can check
- Look for a concrete rights process. A developer should explain how to request access, correction, objection or deletion where the law provides it.
- Separate training from account use. Check whether privacy controls apply to future model training, stored conversations, account history or all three.
- Review agent permissions. Before connecting email, files or work systems, confirm what the agent can read, change, send and retain.
- Record the version and setting. Privacy controls can differ by product tier, region and enterprise contract. Organizations need an auditable record of what was enabled.
- Test the response, not just the notice. A published commitment has value only if requests are processed accurately and safeguards work in practice.
What remains unresolved
The ICO acknowledges that foundation-model training creates difficult questions, including whether models themselves contain personal data and how sensitive information can be used lawfully. It says technical difficulty does not remove the obligation to address basic data-protection requirements.
One major developer is missing from the ten-company list. The regulator paused its supervisory work with X.AI after opening a formal investigation involving Grok. That is another reason not to read the new report as an industry-wide approval.
The strongest result here is not a promise that model privacy has been solved. It is a public set of expectations, named companies and follow-up work against which future claims can be measured.
Tags: Information Commissioner's Office, foundation models, data protection, AI agents, privacy