Blog

X suspends Grok training with third-party data gathered by its agent after Irish DPC request

The Irish DPC asked X to suspend training Grok with third-party data gathered by the agent, and X confirmed it. What this means for a legal basis per step.

· By

Large wall screen showing data streams converging to one point with one connection clearly broken, while two people at a meeting table look on.
At the Irish DPC's request, X suspended training Grok with third-party data gathered by its agent.Image: IamVera.ai — original editorial illustration

The Irish Data Protection Commission asked X on 17 November 2026 to suspend training Grok with personal data that the Grok agent gathered from external sources, and X confirmed it. For organisations, this means every gathering and reuse step by an AI agent needs its own legal basis and demonstrable safeguards.

This is not about the ordinary user data of X posts, but about data on third parties that the Grok agent itself gathered and that was subsequently reused for model training. According to Politico's reporting on the intervention, the regulator spoke of significant concerns about this processing and announced further discussions with X. In our estimation the real signal lies here: an agent that gathers data along the way creates a new processing operation that must be justified separately from the original user relationship.

What exactly did the Irish DPC ask of X and what did X do in response?

The DPC asked X to suspend the processing of third-party data gathered by the Grok agent for AI training, after which X confirmed the suspension. This is not a finding that X has definitively breached the GDPR, nor is it a worldwide end to all Grok training; it concerns this specific processing operation, on which the regulator will enter into further discussions.

The intervention fits an existing pattern. In August 2024 the DPC confirmed that X agreed to suspend the processing of public EU and EEA posts for Grok training, following urgent proceedings before the High Court under Section 134 of the Irish Data Protection Act 2018. In September 2024 the DPC confirmed that the proceedings had been definitively concluded and that X had permanently committed to the previously agreed suspension and it referred key questions about legal basis, first- and third-party data and the phases of model development to the EDPB.

Why do regulators treat data gathered by an AI agent as a separate processing operation?

Because an AI agent can gather data about persons other than the user, which can then be reused for training. In our analysis, gathering, combining, storing and training should be assessed as distinct processing activities, with the applicable legal basis and safeguards documented for each. The question of whether a source was publicly accessible does not answer whether each of those steps is lawful.

The European Data Protection Board (EDPB) stated in Opinion 28/2024 on data protection in the context of AI models that personal data from training data can persist in the model parameters and that a model does not automatically qualify as anonymous. According to the EDPB, when assessing whether a model is anonymous, a regulator must consider, among other things, whether personal data cannot be extracted from the model using means reasonably likely to be used and whether model output is not related to the data subjects whose data was used for training. Our reading: publicly available does not mean free of data protection law, and training is therefore never a neutral technical step.

What does this suspension mean for directors, lawyers and CISOs working with AI agents?

Our analysis: because the DPC is examining third-party data gathered by an agent separately from ordinary user data, accountability should extend beyond the user relationship to the relevant gathering and reuse steps; therefore an organisation deploying an agent should be able to show, per step, which data is gathered, on what basis and with what safeguards. Our analysis: because X confirmed a suspension after the regulator’s request, a technical blocking mechanism for reuse is a prudent control; a CISO should consider setting up a switch that can halt training or reuse per data stream without bringing the entire service to a standstill. Because the EDPB states that personal data can persist in model parameters, removing a source afterwards is not enough; Our analysis: during procurement or development, a lawyer should consider including a clause requiring the supplier to document the origin and legal basis of relevant training data and to test whether personal data can be reproduced or related through model output. Because the intervention revolves around data on third parties that the agent itself gathered, a director cannot fall back on the consent or contract of its own user; Our analysis: they should establish in advance, for each agent, which categories of third parties may come into view and separate those streams from user data.

Which steps record the legal basis and safeguards per training phase?

The crux is that gathering, storing, training and model output are each separately auditable. This limited checklist is our practical translation of the DPC intervention and the EDPB opinion and is intended to allow a regulator's suspension request to be followed up without improvisation.

  • Map which data on third parties an agent can gather and separate those streams from regular user data.
  • Record, per stream, the source, the purpose, the legal basis, the retention period and any onward transfer.
  • Block reuse for training until the assessment of basis, necessity and proportionality is complete.
  • Test whether the model output can reproduce personal data from the training data before assuming anonymity.
  • Set up workable processes for objection, deletion and escalation, and keep evidence of suspension for when a regulator asks for it.

Anyone wishing to embed these steps in the wider framework of gathering and reuse by agents will find more depth in our hub on AI privacy and GDPR, in the analysis of why pseudonymisation of training data does not count as anonymisation under the EDPB, and in the approach for a DPIA that assesses the whole agent chain rather than only dataset and model. For the legal-basis question at token level, the piece on token IDs as pseudonymised personal data is also relevant.

Sources and references

  1. Irish regulator stops X training AI with data gathered by agentsPolitico · 2026-11-17
  2. The DPC welcomes X's agreement to suspend its processing of personal data for the purpose of training AI tool GrokIrish Data Protection Commission · 2024-08-08
  3. Data Protection Commission welcomes conclusion of proceedings relating to X's AI tool GrokIrish Data Protection Commission · 2024-09-04
  4. Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI modelsEuropean Data Protection Board · 2024-12-17

Sources: The article draws on the press releases of the Irish Data Protection Commission, Opinion 28/2024 of the European Data Protection Board and Politico's reporting on the November 2026 intervention.

← All articles in this topic ← All articles