Flarepoint

ICO Report on 10 AI Developers: What It Means When AI Gets You Wrong

2026-10-09

On 8 October 2026 the UK Information Commissioner's Office (ICO) said ten AI developers, including OpenAI, Google, Meta, Microsoft and Anthropic, had made or committed to changes in how they handle personal data. The commitments cover transparency, routes for people to exercise their data rights, and safeguards around training data. The ICO's statement of what it expects from developers does not mention accuracy or the right to have inaccurate data corrected. Anyone an AI system describes falsely still has to prove what it said.

What did the ICO announce?

The ICO's announcement of 8 October names the ten developers: Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI. Each made, or committed to make, changes in three areas: clearer transparency information, stronger ways for people to exercise their rights, and tougher assessments of safeguards. The ICO says it is monitoring progress against those commitments.

The Next Web reported that the commitments come out of a supervision programme the ICO set up in 2025. It chose 11 companies based on compliance risk, UK market share and use of higher-risk training data. The eleventh was xAI. The ICO paused that engagement after opening a formal investigation into the Grok system, and that investigation is ongoing. Infosecurity Magazine reported on 9 October that the investigation covers both X Internet Unlimited Company and X.AI LLC.

The ICO also opened a six-week call for evidence on AI agents, closing on 20 November 2026. It has made enquiries with OpenAI, Anthropic, Meta and the UK AI Security Institute about recent testing and deployment of agents.

What does the report say about people's rights?

The detail is in the ICO report published alongside the announcement. It lists four rights by article: access (article 15), erasure (article 17), restriction (article 18) and objection (article 21). Each is framed around personal data used for training.

The report also sets expectations that matter to anyone making a request on a client's behalf. Developers should offer dedicated, clear routes for submitting requests. They should be able to give "the rationale for any decision to refuse a request". They must tell people they can escalate a complaint to the ICO and "seek enforcement of their data rights through the courts". The report gives examples of existing processes. OpenAI has produced guidance for "people seeking the removal of their personal data from ChatGPT responses". Meta offers "a separate form for third-party data subjects", meaning people who are not Meta users.

The ICO accepts that removing one person's data from a trained system is technically hard. It says compliance will need "a combination of dataset and model-focused measures, as well as output-focused safeguards". It has also started reviewing its 2020 position that personal data in training data can be extracted from the trained system, because some developers disagreed with it.

What does it not say?

The chapter that sets out the ICO's expectations never uses the words "rectification" or "accuracy", and never cites article 16. That is the UK GDPR right to have inaccurate personal data corrected, and it is the right closest to the problem of an AI system attributing a lawsuit, a conviction or a fraud to the wrong person.

The developer-by-developer commitments are mostly about legitimate interest assessments, privacy notices and summaries of training data. The ICO's supervision chapter names no commitment by any single developer about removing personal data from outputs or filtering them. Its one sector-wide point on outputs concerns reducing the "reproduction of memorised personal data". The objection changes it records, secured with Meta, Microsoft and LinkedIn, are about making it simpler for UK users to object to their own data being used for training.

This is a familiar pattern. The EU's transparency rules require AI systems to say they are AI and say nothing about whether what they say about a person is true. The ICO report is a serious piece of supervision, and it governs how AI systems are built. It does not govern what they say about named people and companies.

Why the record still matters

The report makes requests and refusals easier to document. A developer that follows the ICO's expectations will give a written reason when it declines to act. That letter is evidence that the developer was told and what it decided. The federal ruling on correction emails showed how much weight a court can give that kind of paper trail.

A request only works if it describes the statement precisely: which system, what it said, when, and in answer to what question. If the output changes after the request, the person who made it may be left with a reasoned refusal about a statement they can no longer show. Erasure and objection, as the report frames them, deal with what went into a system. A reputational problem is usually about what an AI system says at the moment a reporter, investor or counterparty asks about someone.

What this means for lawyers and comms teams

  • Use the routes the ICO now expects. Dedicated request forms, including those for people who are not users, are the obvious first step for a UK client an AI system has misdescribed. A refusal should now come with reasons that can be challenged before the ICO or a court.
  • Capture the output before you complain. Record the exact answer, the system, the question and the time in a form someone else can check, before the developer has a chance to act on it.
  • Frame requests around the output, not only the training data. The report says outputs need "output-focused safeguards". Hold developers to that wording.
  • Respond to the call for evidence. The ICO's consultation on AI agents closes on 20 November 2026 and will feed into a statutory code of practice on AI and automated decision-making. Agents that research people and companies without supervision fall squarely within reputation practice. Practitioners who see the harm first are well placed to explain to the regulator what accuracy failures look like in practice.

The ICO has made it clearer how to ask an AI developer to act. Proving what the system said, and what it cost, still falls to the person asking.

Sources