Once upon a time, in early spring, an artificial intelligence chatbot, renowned throughout the Internet for its knowledge and ability to answer questions, fell victim to a data breach.

The data breach was publicly notified to the chatbot’s users, and while the news of the breach spread far and wide, a meticulous and attentive group of governmental officials decided to investigate the chatbot’s privacy compliance.

In the wake of the breach and other news on the chatbot’s operations, Italian officials, having identified certain breaches of privacy laws and pending a more thorough investigation, took immediate and urgent action. They indeed ordered the chatbot’s provider to limit the processing of personal data for users in Italy and called on the provider to communicate a plan within 20 days that would address the identified areas of noncompliance.

The chatbot was immediately blocked for Italian users—to protect their personal data while the provider and officials worked to find a solution with the objective of ensuring privacy compliance without stifling innovation.

The provider proposed certain measures, which were duly examined by the officials and, finally, an agreement was reached: If the provider succeeded in implementing the actions agreed with the officials by April 30, 2023, the temporary and urgent limitation on processing the personal data of Italian chatbot users would be cancelled.

The officials closely monitored the provider’s implementation of the proposed measures and, ultimately, the provider succeeded in meeting the requirements to ensure a minimum level of personal data protection and the chatbot was again made available to Italian users!

The above is just a short summary of the recent events involving OpenAI and ChatGPT, on the one hand, and the Italian privacy Supervisory Authority—the Garante—on the other.

Did they live happily ever after? Or, better said: is this a (happy) ending, or is it just the beginning?

To find out, let’s have a closer look at what happened, to understand the moral of the story and to (try to) predict the likelihood of a sequel!

1. What happened? The timeline on ChatGPT’s privacy compliance.

The story of OpenAI and the Garante developed rather quickly. The data breach occurred on March 20, 2023 and was communicated to the public on March 24, 2023; the limitation on OpenAI’s processing of Italian data subjects’ personal data was decided on March 30, 2023; on April 11, 2023 OpenAI agreed with the Garante on the urgent measures it would implement by April 30, 2023; and the limitation was revoked on April 28, 2023.

But what breaches of privacy law did the Garante identify, and what were the measures agreed with the Garante and implemented by OpenAI?

The Garante claimed a breach of Articles 5, 6, 8, 13 and 25 of the GDPR, noting the following failures:

• OpenAI did not provide an adequate privacy policy to ChatGPT’s users, and to all the data subjects whose personal data are collected by OpenAI and processed through ChatGPT.
• There was no suitable legal basis identified for processing the personal data for the purpose of training the algorithms underlying the functioning of ChatGPT.
• The users’ personal data could not be guaranteed to be processed accurately—since the outputs of ChatGPT do not always correspond to real or correct data.
• There was no way to verify the age of ChatGPT’s users, despite—according to ChatGPT’s terms and conditions—the AI system is reserved to individuals who are at least 13 years old.

In response to the Garante’s analysis, OpenAI implemented the following measures to remedy the cited noncompliance with the GDPR; in particular:

• It published, on its website, an information notice addressed to users and nonusers of ChatGPT, describing what kinds of personal data are processed for training the AI system and how and clearly stating that everyone has the right to opt out from such processing.
• The privacy policy was provided to ChatGPT’s users was made accessible from the signup page for new users as well as from the welcome (back) page, where already signed-up Italian data subjects would be directed after the temporary limitation on processing was removed; in addition, with regards to the legal basis for processing, OpenAI clarified that ChatGPT users’ personal data is processed on a contractual basis for the purpose of enabling performance of the AI system, whereas the same personal data is processed for the purpose of training the AI system on the basis of OpenAI’s legitimate interest.
• It created a way (i) for all European data subjects to easily object to the processing of their data for the purposes of training the AI system (and, therefore, to exclude chats and chat history from the data used for training the AI system), and (ii) for all data subjects to obtain erasure of information that is considered inaccurate, as OpenAI stated that it is technically impossible (currently) to rectify inaccuracies.
• An age gate was implemented. On the welcome (back) page prepared for registered Italian users accessing ChatGPT after removal of the temporary ban, users have to confirm that they are 18 or older, or, if age 13 to 17, they have their parents authorization to use ChatGPT. On the signup page for unregistered (first-time) users, users are asked to enter their birthday, and access by users under 13 is denied whereas, for users 13 to 17, access is blocked unless confirmation of parental consent is provided.

2. What is the moral of the story? The Garante’s view and the lessons learned.

As expected, the Garante’s limitation of OpenAI’s processing of personal data of data subjects in Italy triggered a public debate. Certain commentators complained that the Garante, in doing so, was hampering the development of AI and technological innovation and condemning Italy to lag behind other countries; other commentators welcomed the Garante’s timely action with a focus on citizens’ personal data protection.

While the debate on this continues, there is one fact that almost no one has denied: Privacy compliance is not easy to reach for generative AI, considering its intrinsic characteristics and training methodology. Ingesting large amounts of data, collected (or, better said, scraped) from the Internet based on the belief that, since it is public data, it can be used for whatever purpose (and even for purposes different from those for which the same data is made public…) can hardly be consistent with—just to name a few—the purpose limitation principle, the data minimization principle and the storage limitation principle in accordance with Article 5 GDPR.

In this context, the measures implemented by OpenAI and welcomed by the Garante can shed light on some of the main issues to be solved when applying privacy laws to AI, as the same measures can be used as guidelines for developing generative AI in a way suitable to “reconcile technological advancements with respect for the rights of individuals”.

The Garante’s orders are of particular interest as they seem to confirm that the legitimate interest of the provider of an AI system can concretely be considered as a legal basis for processing personal data for the purpose of training an AI system. Obviously, on the fundamental assumption that the provider is and remains accountable for its choice of the legal basis applicable to the processing activities carried out.

Based on the analysis of Garante’s orders towards OpenAI and ChatGPT, we drafted a five-point checklist for developing privacy-mindful AI systems. Our checklist can be found here.

3. Will there be any sequel to this story? ChatGPT’s quest for a full regulatory compliance.

If we were to answer the very first question in this article, we’d say that obviously everyone likes a happy ending and, indeed, many commentators have already rejoiced at learning that the measures agreed by OpenAI and the Garante were implemented within the given deadline and that, consequently, ChatGPT is available again in Italy.

Does this mean that ChatGPT is now privacy-proof? No; or, at least, not yet.

The Garante ordered a temporary limitation on ChatGPT as an urgent remedy, to foster the prompt achievement of a minimum level of compliance while carrying out a more thorough inspection of OpenAI’s compliance with European privacy laws.

Such minimum level of compliance seems to have been reached thanks to the measures agreed by OpenAI and the Garante, but some of those measures remain to be implemented. This is the case, in particular, for the age verification mechanism (this must be presented to the Garante by May 31, 2023 and put in full operation by September 30, 2023) and for the “non-marketing campaign” (that must be run on all main Italian media by May 15, 2023). This campaign is to inform data subjects that their personal data has likely already been collected for the purpose of training algorithms, that an ad-hoc detailed information notice has been published on OpenAI’s website, and that everyone can request and obtain the erasure of their personal data using the tool available on the same website.

In addition to the measures already implemented, and to the other measures already agreed that remain to be implemented, OpenAI may still be required to make more efforts.

The Garante is indeed continuing its investigation of OpenAI’s privacy compliance and likewise other European Supervisory Authorities have also announced that they’ve started their own analyses; this led to the European Data Protection Board’s decision to launch a dedicated task force to promote cooperation and to exchange information on possible enforcement actions conducted by national Supervisory Authorities against OpenAI and ChatGPT. All these actions may result in further conditions to be met by OpenAI—and, as applicable, by any AI systems.

Last, but not least—the AI Act.

In previous articles, we discussed about the Commission’s draft and the Council’s common position (general approach), which is now being examined by the European Parliament with a view to prepare the text that will undergo the trilogue process. It’s well known that the Parliament’s discussions became more difficult due to the rise of ChatGPT, as ChatGPT made more evident the risks of using general purpose AI systems and of foundation models and urged the need to ensure that both categories fall within the scope of the AI Act and are duly regulated to protect individuals.

According to recent press news, the European Parliament has just reached a provisional political deal on a draft that, despite being still subject to certain adjustments, should go to a plenary vote within the end of June.

Such parliamentary draft is expected (i) to put stricter obligations on foundation models, which are part of general purpose AI systems and which also include also generative AI systems like ChatGPT, and (ii) to provide for general principles applicable to all AI systems, which include, among others, privacy and data governance.

Privacy compliance is therefore highly likely to be (explicitly) required not only under the GDPR, but also under the AI Act. If so, it might be necessary—among others—to understand how the two sanctions regimes (the one under the GDPR, and the one under the AI Act) will intertwine, and how the regulatory authorities will cooperate in their respective fields of competencies. Obviously, if privacy Supervisory Authorities are put in charge of monitoring application of the AI Act, coordination will likely be easier.

“But this is another story and shall be told another time” (Thank you Michael Ende!).

If you are curious about generative AI, check out the recording of our Digital Breakfast @Dentons “Let’s talk about artificial intelligence” (in Italian only) and, to avoid missing any updates, remember to follow us and don’t hesitate to contact us with any questions!

This article is updated as of May 2, 2023