The problem
Models trained on people’s data without their permission.
As AI is deployed across society, we believe this is the biggest problem it raises. Our research feed tracks the cases: regulator decisions, court orders and policy changes at the largest AI providers. Five patterns stand out.
Training starts before consent is settled.
X used EU users’ public posts to train Grok from 7 May to 1 August 2024, until Ireland’s Data Protection Commission went to the High Court. LinkedIn switched generative-AI training on by default, reportedly before updating its privacy policy.
Sources: X / xAI case, LinkedIn case
Opt-out is the default.
Slack trains global models on customer messages unless a workspace owner opts out by email. Meta and LinkedIn train on European users’ content under legitimate interests, with an objection form or setting.
Sources: Slack case, Meta case
Once trained, data is hard to take back.
Turning training off does not remove data from models already trained. Research shows that popular unlearning methods mostly suppress what a model says: light fine-tuning brings the ‘forgotten’ text back.
Sources: Hu et al., Anthropic case
Models can give the data back.
Researchers extracted gigabytes of training data from production language models. Separately, about 8% of 100,342 publicly shared chatbot conversations disclosed private or sensitive data.
Sources: Nasr et al., Mollaeefar et al.
Enforcement is slow and uncertain.
Italy fined OpenAI €15 million in December 2024 over training without an adequate legal basis; a Rome court annulled the fine in March 2026. Ireland’s inquiry into Grok training, opened in April 2025, has published no decision.
Sources: OpenAI case, X / xAI case
Selected cases
| Date | Provider | What happened | Status |
|---|---|---|---|
| Mar 2023 | OpenAI | Italy’s Garante temporarily limits ChatGPT, citing no adequate legal basis to collect data for training. Source | Service resumed in April 2023 after changes. |
| May 2023 | Amazon | FTC and DOJ allege children’s Alexa recordings were kept against deletion requests and used to train the algorithm. Source | Settled: $25M penalty, no admission. |
| May 2024 | Slack | Privacy principles allow training global models on customer messages and files, opted in by default. Source | Unchanged; opt-out by email. |
| Jun 2024 | Meta | Pauses plans to train on EU adults’ public Facebook and Instagram content after regulator engagement. Source | Resumed May 2025 with an objection form. |
| Aug 2024 | X / xAI | EU users’ posts used to train Grok; Ireland’s DPC makes its first urgent High Court application. Source | X stopped; inquiry open since April 2025. |
| Sep 2024 | Training on member data switched on by default; paused for the UK after the ICO raised concerns. Source | Resumed with opt-out, extended to the EEA in November 2025. | |
| Dec 2024 | OpenAI | Italy’s Garante fines OpenAI €15M over training ChatGPT without an adequate legal basis. Source | Annulled on appeal, March 2026. |
All 17 cases we track are in the research feed, each with its source and current status. Status as checked on 1 November 2026. Open the research feed
Why private inference is the answer
Rules and policies change; a provider that cannot read the data cannot train on it. Private inference turns privacy from a promise in the terms of service into a property of the system. It is not free: today’s methods cost speed and money, and each still asks you to trust someone, such as a hardware vendor or servers that must not collude. We publish those limits next to every result.
Who we are
A small core team with top researchers across Europe.
RBlock is an independent laboratory based in Europe. We research private inference: combining trusted execution environments, multi-party computation and on-premises inference with data partitioning, so that a cloud operator cannot read the data without physically breaking the hardware.
The same team runs regulatory and market-intelligence monitoring for clients in Spain and Europe, and invests up to €250,000 in European researchers building private-inference tools.
How we operate
Core engineering in-house, advanced research with the best.
Core engineering and documentation, in-house
The core team builds the systems and writes the documentation itself, so we understand every result we publish.
Automated back end
We automate monitoring, research pipelines and reporting, so people spend their time on the work that raises the quality of each deliverable.
Top researchers, part time
For advanced work we bring in leading researchers based in Europe on part-time engagements, matched to the question.
Sources and limits, always
Every claim ships with its sources, its limits and the date it was checked.
Where we are going
At the edge, and in the open.
We like working at the edge of what is possible, and we want to help steward private–public collaboration on private AI in Europe: between researchers, companies and public bodies that each hold part of the data and part of the answer.
We aim to be a good contributor to the society our work touches: open about what we know, clear about what we don’t, and useful to the people whose data is at stake.
Founding team
Rayco Hernández Moreno
Founder
Electrical engineer with more than fifteen years at NXP Semiconductors, from chip design tools and the secure-identification business (the chips that protect bank cards, passports and phones) to global leadership. He founded RBlock to bring the same discipline to AI: trust that rests on something you can check.
Work with us.
Research and client work
[email protected]Research collaborations, private-inference prototypes and client work.
Founders and investment
[email protected]Decks from European researchers building private-inference tools.