GDPR Data Subject Access Request for AI Training Data

You are the Data Protection Officer at Lingua360, a London-based AI company that develops multilingual translation and content generation tools. Lingua360's models were trained on a dataset of 2.1 billion text samples scraped from the public web, licensed datasets, and customer-contributed data from enterprise clients who opted into a data improvement program.

Today you received a formal GDPR Article 15 Subject Access Request (SAR) from Dr. Elena Vasquez, a Spanish medical researcher. Dr. Vasquez has discovered that several of her published academic papers, blog posts, and social media content were included in Lingua360's training dataset. She is requesting:

1. Confirmation that her personal data was used in training (Article 15(1)(a))

2. A copy of all her personal data in Lingua360's systems (Article 15(3))

3. Information about the purposes of processing and recipients of data (Article 15(1)(a-g))

4. Erasure of her data under Article 17 (right to be forgotten)

5. Correction of inaccurate outputs the model generates about her (Article 16)

Each request raises complex legal and technical questions that do not have settled answers.

Article 15 — Right of access: Lingua360 can confirm that Dr. Vasquez's data was in the training set because the company maintains a URL-level index of training data sources. However, providing a "copy" of her data is complicated — her data was used to adjust model parameters during training and does not exist as a discrete, extractable record within the model itself. The Article 29 Working Party guidance suggests that the right of access applies to personal data in any form, but extracting specific training examples from a neural network remains technically challenging.

Article 17 — Right to erasure: Dr. Vasquez requests deletion of her data. Lingua360 can remove her data from the training dataset to prevent use in future training runs. However, the model has already been trained — her data's influence is embedded in the model weights. Full erasure would require machine unlearning, which is an active research area but not yet reliable at production scale. The alternative is retraining the model from scratch without her data, at a cost of approximately 4.2 million euros.

Article 16 — Right to rectification: Dr. Vasquez reports that the model generates inaccurate biographical information about her. Rectification of AI outputs is particularly challenging because the model generates text probabilistically — there is no "record" to correct. Techniques like RLHF fine-tuning or output filtering could reduce but not eliminate inaccurate outputs.

Legal basis for original processing: Lingua360 relies on Article 6(1)(f) — legitimate interest — as its legal basis for processing publicly available data for AI training. Following the Garante's ChatGPT decision and subsequent guidance, this basis requires a documented legitimate interest assessment (LIA) and a balancing test demonstrating that the company's interests do not override the data subject's rights and freedoms.

As DPO, you develop a comprehensive response within the 30-day GDPR deadline:

Immediate response (within 30 days):

- Confirm that Dr. Vasquez's data was included in the training dataset, providing the specific URLs and data categories

- Provide information about processing purposes, legal basis (legitimate interest), data retention periods, and data recipients

- Remove her data from the active training dataset and all data pipelines

- Document the technical limitations of extracting or erasing data from trained model weights, with supporting evidence from technical staff

- Implement output filtering to reduce generation of inaccurate information about Dr. Vasquez

- If the full request cannot be completed within 30 days, notify Dr. Vasquez of a 60-day extension under Article 12(3) with reasons

Systemic improvements:

- Build a data subject search tool that allows individuals to check if their data is in the training set

- Implement a web crawler opt-out mechanism (robots.txt compliance and a dedicated exclusion registry)

- Document the legitimate interest assessment for web scraping, including the balancing test

- Establish a machine unlearning research initiative to develop viable erasure capabilities