AI data controls need to work at the moment people share

An AI conversation can begin with a routine question and turn into something a person would never knowingly contribute to a training set. The OECD warns that chatbot users may not realise their conversations are logged or used for training. These routes differ, but they pose a common question: when does the person concerned get a meaningful chance to decide what happens next?

Freevacy’s account of the brief likewise identifies chatbot inputs and large-scale web scraping as its central subjects. For teams collecting human responses to train or evaluate AI, the important distinction is not simply whether information was volunteered or found elsewhere. It is whether the proposed use is clear enough, and the available choice specific enough, for a person to act on it. A permission sought at the start of an exchange cannot anticipate every disclosure that follows.

Control needs to follow the exchange

The OECD reports that some developers let users exclude conversations from model training, but such settings generally cover all their activity rather than selected conversations, files or categories of information. It suggests more specific controls because the sensitivity of what someone shares can change during an exchange. That is a useful challenge to research teams: an account-level opt-out is not necessarily a sufficient answer to a concern about a particular answer.

A careful study design would separate permission to provide a response from permission to reuse it for a different purpose. It would also consider whether participants can skip a question, restrict a particular submission or reconsider a choice before data enters a training workflow. None of these steps makes a sensitive response harmless. They do, however, make the decision more closely resemble the material the person is actually being asked to share, rather than an abstract category called ‘data’.

Timing is part of the choice

The OECD proposes giving users relevant information at the point of interaction, such as before a sensitive exchange, instead of relying only on help pages or privacy policies. Separately, Ireland’s Data Protection Commission says people should have sufficient notice and an easy way to object before their personal data is used for AI training. The shared lesson is temporal: information delivered after reuse begins cannot perform the same function as a choice offered beforehand.

In a supervisory case, Ireland’s Data Protection Commission recommended excluding data collected before, at the earliest, 2014 from generative-AI training because earlier notices gave people insufficient reason to expect that use; the controller reduced the period to 2014 onward. This is not a universal cut-off for other datasets. It illustrates a more durable problem: an old collection notice does not automatically cover a new purpose simply because the data remains available.

Build the choice into the workflow

For research and monitoring-and-evaluation teams, the practical task is to make permissions usable downstream. Collection forms, interview protocols and dataset handovers should distinguish participation in a study from later model development. Teams should decide in advance how a restriction on a particular response will be recorded and communicated to anyone preparing data for reuse. Otherwise, a choice offered to participants can disappear during cleaning, export or aggregation.

The point of intervention matters because deletion may not reverse every use. The OECD cautions that personal data memorised during training can be difficult to remove from a model even after deletion from the underlying training database. Teams should therefore treat an objection before training as a design requirement, not merely a customer-support request to resolve later. The test is straightforward: can a person understand the proposed use and exercise a relevant choice while the team can still honour it?