The European Data Protection Board issued a report opinion Addressing data protection in artificial intelligence models. It covers the assessment of AI anonymisation, the legal basis for data processing, and mitigation measures for impacts on data subjects for technology companies operating in the cluster.
It was published in response to a request from the Irish Data Protection Commission, which is the main supervisory authority under the GDPR for many multinational companies.
What are the main points of guidance?
DPC has requested more information about:
- When and how can an AI model be considered “anonymous” – one that is unlikely to identify the individuals whose data was used to create it and is therefore exempt from privacy laws.
- When companies can say they have a “legitimate interest” in processing individuals’ data for AI models, and therefore do not need to obtain their consent.
- Consequences of unlawful processing of personal data at the stage of development of an artificial intelligence model.
EDPB Chairman Anu Talos said in a press release: “AI technologies may provide many opportunities and benefits to various industries and areas of life. We must ensure that these innovations are done ethically, safely and in a way that benefits everyone.
“The EDPB wants to support responsible innovation in AI by ensuring the protection of personal data and full respect for the GDPR.”
When an AI model can be considered “anonymous”
An AI model can be considered anonymous if the chance that the personal data used for training will be traced back to any individual – either directly or indirectly, such as through a claim – is “slim.” Non-identification is assessed by supervisory authorities on a “case-by-case” basis and a “comprehensive assessment of the likelihood of identification” is required.
However, the opinion provides a list of ways model developers can prove anonymity, including:
- Take steps during source selection to avoid or limit the collection of personal data, such as excluding irrelevant or inappropriate sources.
- Implement strong technical measures to prevent re-identification.
- Ensure that data is sufficiently anonymised.
- Apply data minimization techniques to avoid unnecessary personal data.
- Regularly assess redetermination risks through testing and auditing.
These requirements will make it difficult for AI companies to demand anonymity, said Katherine Wynn, a data protection lawyer from Pinsent Masons.
“The potential harm to the privacy of a person whose data is used to train an AI model can be relatively minor, depending on the circumstances, and may be further reduced through security and pseudonymization measures,” she said in a statement. Company article.
“However, the way the EDPB interprets the law will require organizations to meet burdensome, and in some cases unwieldy, compliance obligations with respect to definition of purpose and transparency, in particular.”
When AI companies can process personal data without individuals’ consent
The EDPB opinion clarifies that AI companies can process personal data without consent under a “legitimate interest” basis if they can demonstrate that their interest, such as improving models or services, outweighs an individual’s rights and freedoms.
This is especially important for technology companies, because obtaining approval for the vast amounts of data used to train models is neither trivial nor economically feasible. But to qualify, companies will need to pass these three tests:
- Legitimacy test: A legal and legitimate reason for processing personal data must be determined.
- Necessity test: Data processing must be necessary for the purpose. There cannot be any other alternative, or less intrusive ways to achieve the company’s goal, and the amount of data processed must be proportionate.
- Balance test: The legitimate interest in data processing must outweigh its impact on the rights and freedoms of individuals. This takes into account whether individuals would reasonably expect their data to be processed in this way, as if they had made it publicly available or had a relationship with the company.
Even if a company fails the balancing test, it may not need to obtain consent from data subjects if they implement mitigating measures to limit the impact of processing. These measures include the following:
- Technical guarantees: Implement safeguards that reduce security risks, such as encryption.
- Alias: Replace or remove identifying information to prevent the data from being linked to an individual.
- Hide data: Replace real personal data with fake data when the actual content is not necessary.
- Mechanisms for data subjects to exercise their rights: Facilitate individuals exercising their data rights, such as opting out, requesting erasure, or submitting data correction requests.
- Transparency: Public disclosure of data processing practices through media campaigns and transparency posters.
- Measures for web scraping: Implement restrictions to prevent unauthorized scraping of personal data, such as providing an opt-out list for data subjects or excluding sensitive data.
Technology lawyer Malcolm Dowden of Pinsent Masons said in the firm’s article that the definition of “legitimate interest” has been controversial recently, particularly in the UK’s Data (Use and Access) Bill.
“AI advocates suggest that data processing in the context of AI drives innovation and brings inherent social good and benefits that constitute a ‘legitimate interest’ for the purposes of data protection law,” he said. “Opponents believe this view does not take into account the risks associated with AI, such as privacy, discrimination, or the potential spread of ‘deep fakes’ or disinformation.”
Advocates at the charity Privacy International have expressed concerns that AI models like OpenAI’s GPT series may not be properly vetted under the three tests because they lack… Specific reasons for processing personal data.
Consequences of illegal processing of personal data in the development of artificial intelligence
If the model is developed by processing data in a way that violates the GDPR, this will impact how the model is allowed to work. The competent authority assesses “the circumstances of each case individually” but provides examples of possible considerations:
- If the same company holds and processes personal data, the legality of the development and deployment phases should be evaluated based on the details of the case.
- If another company processes personal data during deployment, the EDPB will consider whether that company has conducted an appropriate assessment of the lawfulness of the form in advance.
- If the data is anonymised after being unlawfully processed, the subsequent processing of non-personal data will not be liable to the GDPR. However, any subsequent processing of personal data will remain subject to regulation.
Why should AI companies pay attention to the guidance?
The EDPB’s guidance is crucial for technology companies. Although it has no legal authority, it affects how privacy laws are implemented in the European Union.
In fact, companies can be fined up to €20 million or 4% of their annual turnover – whichever is greater – for GDPR breaches. They may also be required to change how their AI models work or delete them entirely.
See: EU AI law: Europe’s new rules for AI
AI companies struggle to comply with GDPR due to the vast amounts of personal data needed to train models, often sourced from public databases. This creates challenges in ensuring the lawful processing of data and processing requests for data subject access, corrections or erasure.
These challenges have manifested in numerous legal battles and fines. For example:
In addition, in September, the Dutch Data Protection Authority Clearview AI fined €30.5 million To illegally collect facial images from the Internet without user consent, in violation of the GDPR. In the same month, the Irish Democratic Policy Committee commissioned the opinion after successfully convincing Elon Musk to do so Stop using European users’ public posts to train its chatbot, Grokwithout obtaining their consent.