AI training data needs a traceable path to people’s rights

On October 8, 2026, the UK Information Commission’s Office said ten foundation-model developers had made or committed to make changes in how they handle personal data, and that it was monitoring their progress. That distinction matters: a commitment is not the same as a completed change. The useful lesson for research teams is not that a regulatory review has settled the question of responsible AI training. It is that claims about responsible training can be tested against specific records, explanations and routes for people to act.

The ICO’s report examined training data drawn from publicly accessible websites, procured datasets and people with whom a developer has a direct relationship. Those routes create different opportunities to know where information came from and to explain its use to the people concerned. A dataset’s availability does not, by itself, answer either question. For anyone commissioning research or preparing data for AI work, provenance should be treated as part of the study design, not as paperwork assembled after collection.

A source list is not an explanation

The ICO found that some developers’ privacy information said too little about their training datasets or the categories of personal data involved. A broad label such as “public data” can obscure the decisions a research team most needs to inspect: what was collected, through which route, under what conditions, and with what limits on later use. The point is not to demand that every record be published. It is to make the path from a stated data source to an actual use intelligible enough to challenge.

In a September 9 disclosure, Apple listed licensed or purchased third-party data, user-study data, publicly available data and synthetic data among its generative-AI training sources. Such categories help frame further questions; they cannot answer them all. A careful team would still ask what permissions and notices accompany each route, how personal data is identified within it, and whether documentation follows the data when it moves between organisations. Without that continuity, a source inventory can look reassuring while leaving the people represented in the data difficult to locate.

Rights must work after collection

The ICO said developers using third-party personal data for training must provide accessible information about its sources, types and categories, alongside a way for affected people to exercise their rights. It also identified unclear routes for access, erasure or objection where developers had obtained data indirectly. This is a practical warning for research and evaluation teams that receive datasets through intermediaries. If a participant cannot tell who now holds their information, a rights notice at the original point of collection may offer little help when it is needed.

The ICO’s supervision findings also identified gaps in some developers’ evidence that safeguards adequately protected people when they relied on legitimate interests to train on personal data. The research lesson is to document the protection, not merely name the basis for using the data. Teams should be able to explain what a safeguard does, whom it protects and how they would know if it failed. Computing highlighted the ICO’s position that a model itself may contain personal data, not just the dataset used to train it. That possibility makes a plan confined to deleting or restricting source files an incomplete answer.

What careful teams should take forward

According to the ICO’s supervision account, Apple, Cohere and OpenAI changed the information they provide about model training, while seven other developers made changes or committed to measures described in its report. The distinction between changes made and changes promised is worth preserving in any assessment of progress. Research teams should apply the same discipline to their own claims: specify what is documented, what is operational, and what remains a proposed fix.

For a dataset intended for analysis or AI training, the most useful review may begin with an ordinary request: show the route by which a person’s information entered the work, the reason it may be used, and the route by which that person can question the use. If those answers depend on records held by someone else, the handoff itself needs scrutiny. Responsible data practice is not established by a polished statement of principles; it is demonstrated when those questions remain answerable after data has travelled.