How Private AI Facilitates GDPR Compliance for AI Models: Insights from the EDPB's Latest Opinion

The European Data Protection Board (EDPB) has recently provided critical guidance on ensuring GDPR compliance during the development and deployment of AI models. Opinion 28/2024 addresses core data protection issues, such as the use of personal data in AI model training, legal bases for processing, and the impact of unlawful data processing on the deployment and use of AI models.

For organizations grappling with these complexities, Private AI offers state-of-the-art tools to embed privacy protection into every phase of AI development, facilitating compliance with GDPR’s robust requirements.

The acute relevance of this guidance is highlighted by Italy’s data protection agency’s €15 million fine imposed on OpenAI for violating the legal basis requirement under GDPR for the processing of personal data as part of their generative AI model development.

The Challenge: GDPR Compliance in the AI Lifecycle

The development of AI models often relies on vast data sets, many of which may contain personal data, either because the intended use case requires it, or because the data were indiscriminately scraped from the internet, including personal data more or less intentionally.

The question then arises whether the AI model provider can nevertheless claim that the model is anonymous. According to the EDPB, claims of anonymity regarding AI models must meet stringent requirements. This includes ensuring that personal data cannot be extracted from a trained model, e.g., though an adversarial attack with or without access to the model, and that no output generated by the model can reveal identifiable information. The relevance of this question, although not addressed in the Opinion, presumably consists in consequences such as the model not being subject to data subject right access requests, disclosure restrictions, and purpose limitations, among other things.

Another vexing question resulting from the considerable amount of data required to develop AI models the EDPB addresses revolves around the legitimate interest legal basis for personal data processing for AI model development purposes. The consequences of getting this wrong were demonstrated in OpenAI’s fine of €15 million.

As welcome as clarity around this question is, especially from the perspective of businesses for whom such a fine cannot be written off as cost of doing business, and that have limited resources to ensure compliance, the GDPR requirements are arguably in tension with the regulation’s objective of supporting innovation.

And what does it mean for downstream users of commercially available LLMs that were developed unlawfully from the GDPR’s perspective? The Opinion advises that such a second data controller should conduct an appropriate assessment of the lawfulness of processing by the model developer, taking into account whether there is a decision by a court or a supervisory authority to that effect.

Now that there is the Garante’s finding regarding OpenAI’s lack of a legal basis for data processing, downstream users of their LLMs may have clarity with regard to the unlawfulness of personal data processing by OpenAI. But does that mean they cannot demonstrate compliance of their own data processing if they use the model in their AI applications?

The Opinion only says that the original unlawful processing should be one factor when assessing the downstream controller’s own processing activity. This seems to imply subsequent use of unlawfully developed AI models does not automatically result in a finding of unlawful processing by the downstream user.

For organizations seeking to navigate these challenges, Private AI provides solutions that can help align with the EDPB’s framework.

Embedding Privacy from the Ground Up

Private AI enables organizations to address EDPB requirements at every stage of the AI lifecycle:

1. Minimization of Personal Data

The EDPB indicates that AI models trained on personal data must not inadvertently allow the extraction of personal data or the regurgitation during production. Achieving this level of protection requires advanced de-identification techniques.Private AI’s technology facilitates compliance by:

• Identifying and Redacting Personal Data in Training Data Set: Our machine learning models detect and redact over 50 types of personal identifiers, including sensitive information such as health data or ethnic origin. This allows for the cleaning of data sets used in AI training at scale. It can thus greatly limit the risk of harmful extraction by adversarial attacks as well as reproducing training data containing personal data during inference. Whether this suffices for the model to be considered anonymous should be left to experts to determine on a case-by-case basis.
• Identifying and Redacting Personal Data in the Model Output: The same technology can also be used to strip personal data from a model’s output. This may serve as a potential method for ensuring lawful processing when using an LLM that has been developed by processing personal data unlawfully, as it enables the downstream user to prevent the model from generating personal data.
• Preventing Re-identification Risks: By replacing personal data with synthetic placeholders, Private AI reduces the likelihood of re-identification through membership inference or model inversion attacks. When some data points are inadvertently retained in their original form, they are hiding in plain sight, as they are difficult to distinguish from their synthetic neighbors.

2. Demonstrating Legal Basis for Processing

Controllers must select a valid legal basis for processing personal data, such as legitimate interest, and substantiate their choice through a three-step test: identifying the interest, proving necessity, and balancing it against the rights of data subjects

.Private AI supports this process by:

• Reliably identifying what personal data is contained in the data set: This is the crucial first step for every assessment. You need to know what you are processing.
• Providing Granular Data Control: Organizations can enforce strict data minimization policies, ensuring that only essential data is processed for each purpose, as explained above.
• Creating Robust Audit Trails: By documenting how personal data is de-identified or replaced, Private AI helps organizations demonstrate compliance with GDPR’s accountability principle.

3. Addressing Unlawful Processing

The EDPB recognizes that AI models trained on unlawfully obtained personal data present a unique challenge. While uncertainty remains at this time, it seems worth considering whether the following mechanisms could subsequently limit the risk to individuals sufficiently.Private AI minimizes these risks by:

• Validating Model Outputs: Our solutions can support the assessment of an LLM to help verify that outputs generated by AI models do not inadvertently reveal personal data.
• Removing Retained Data: If a model inadvertently retains personal data and includes them in its output, our tools can intersect and remove the personal data before they are shown to the user.

4. Enhancing Transparency and Accountability

Transparency is central to GDPR compliance, yet AI technologies’ inherent complexity often obscures data processing practices. The EDPB highlights the need for clear documentation and the proactive demonstration of compliance measures.Private AI empowers organizations to meet these demands by:

• Generating Detailed Documentation: Our tools provide clear, auditable records of how personal data is handled, ensuring organizations can respond to inquiries from Supervisory Authorities.
• Streamlining Subject Access Requests (SARs): By automating the identification and redaction of personal data, Private AI simplifies responses to SARs, bolstering compliance with GDPR-mandated rights. The key function here is our ability to generate reports on the personal data contained in unstructured data at scale. This can be a laborious to impossible task when attempted to be undertaken manually, depending on the size of the data set concerned.

The Private AI Advantage

The EDPB’s Opinion underscores the high stakes of GDPR compliance in AI development. Non-compliance not only risks regulatory penalties but also undermines trust in AI systems. With Private AI, organizations gain a powerful ally in meeting these challenges head-on.

By integrating privacy into every step of the AI lifecycle—from data collection and training to deployment and beyond—Private AI ensures compliance with GDPR’s stringent standards. Our cutting-edge technology enables businesses to innovate responsibly, without compromising on privacy.

To see how Private AI can enhance your GDPR compliance strategy, try our web demo or get an API key to test our solutions on your own data.