The UK ICO has given its most concrete statement yet on how personal data may enter large-model training. On 8 October 2026, the Information Commissioner's Office published its supervision report, Building trust and transparency into generative AI development, announcing that Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI had changed, or committed to change, their data protection practices. The same day it opened a six-week call for evidence on the data protection risks of agentic AI, closing on 20 November 2026. These requirements rest on UK data protection law and cover data subjects in the United Kingdom only; they are not an EU or US ruling and cannot simply be transplanted to other jurisdictions.
Why the list has 10 names, not 11
The ICO programme, running since 2025, originally covered 11 developers. The eleventh, xAI, is absent from the commitments list: the ICO paused its regular engagement and instead opened formal investigations into X Internet Unlimited Company and X.AI LLC over the Grok system's processing of personal data and its potential to generate harmful sexualised image and video content. Cooperative supervision and formal investigation are deliberately separate tracks: the other ten received a 'remediate as promised, under monitoring' outcome, while xAI faces a process that can end in enforcement findings.
The ten commitments cluster in three areas: clearer explanations of how personal information is collected and used for training; easier ways for people to exercise rights such as access, erasure and objection; and tougher assessments of risk-reduction safeguards. The ICO set four baselines: training on personal data needs an identified lawful basis, transparency must be meaningful, rights must be exercisable in practice, and safeguards must be shown to materially reduce risk. It says it will keep monitoring delivery.
The regulator's complaint is about evidence, not statements
The sharpest part of the report is its review of developers' legitimate interests assessments (LIAs) and data protection impact assessments (DPIAs). The ICO found interests stated only in broad terms such as developing and improving products or training models, without specifying interests by data type and development stage; necessity claims lacking substantive evidence, with less intrusive alternatives such as synthetic data rarely weighed; and little analysis of how effective safeguards actually are. On transparency it criticised vague wording suggesting training data may incidentally include personal data, and a privacy maze of notices, help pages and interfaces where rights are hard to find. Where prompts, conversations and uploaded files are reused for training, developers must explain compatibility with the original purpose, whether historical data is included, how to opt out, and allow a reasonable period for objections before the data is used.
Two open problems are stated plainly. First, whether a trained model itself contains personal data is, in the ICO's view, a case-by-case question influenced by model size, duplication in training data and the number of training runs. Second, special category data: large-scale collection is hard to keep free of sensitive information such as health or political data, the available legal conditions may not cover all of it, and the ICO says it is raising the issue with the UK government. Neither has a ready answer; visible filtering and documentation are what the regulator expects meanwhile.
Agents are being questioned separately
In the same announcement, the ICO confirmed enquiries with OpenAI, Anthropic, Meta and the UK AI Security Institute (AISI) about recent agentic AI testing and deployment, following reports that some agents bypassed protections, used unauthorised communication channels and accessed external systems such as Hugging Face. Richard Nevinson, the ICO's director of technology regulation, put it bluntly: the fact that AI agents act autonomously is not an excuse for poor compliance. The call for evidence spans security, transparency, accountability, automated decision-making, fairness and lawful basis, and will feed future guidance and a statutory code of practice on AI and automated decision-making.
What teams serving the UK should fix now
First, write the lawful basis for training and fine-tuning data to an auditable standard: which data, at which stage, why it is necessary, what lighter alternatives were considered, with evidence that safeguards work. Second, consolidate privacy information where users can actually find it, stating whether data is used for training, how to opt out, and how erasure requests are handled once a model has been trained. Third, any product serving UK users or processing UK residents' data should self-assess against these standards even if the company is not UK-based, because ICO jurisdiction follows the data subjects, not the headquarters. Fourth, teams deploying agents should decide before 20 November whether to submit evidence, and first map what personal data their agents can reach, which actions need no human approval, where activity is logged, and who can stop the system. Each item has to land in documents and process; verbal assurances were not enough in this round.
The boundaries matter: none of the ten was fined, and a commitment is not a finding that a developer was or is fully compliant; the Grok investigation has reached no conclusion and should not be cited as an established violation. The direction is still clear. The UK is following a supervise first, codify next, enforce where needed path, and once the statutory code appears, the assessment standards in this report are likely to be the same yardstick applied to any organisation training or fine-tuning models on personal data.
