Italy Blocks ChatGPT Over Privacy and Child Safety Concerns

On March 31, 2023, Italy’s data protection authority imposed a temporary limitation on the processing of ChatGPT users’ data; the decision shows how to turn a regulatory alert into a concrete list of data, legal basis, age controls and deadlines.

Italy’s data protection authority on Friday ordered a temporary block on ChatGPT for users in Italy. The agency, known as the Garante, questioned whether OpenAI has a sufficient legal basis to collect and use the personal data it uses to train its system, and pointed to specific risks for minors.

The decision makes Italy the first Western country to directly suspend access to ChatGPT on data protection grounds. This is no theoretical debate: it affects a tool that, since its launch in November, has brought generative AI to millions of people and thousands of businesses.

What Italy is challenging OpenAI over

The Garante’s order immediately restricts OpenAI’s processing of Italian users’ data. The authority is examining a possible breach of the General Data Protection Regulation (GDPR), the European law that requires companies to clearly explain why personal data is collected and have a legal justification for using it.

The regulator identifies three main problems. The first concerns the information provided to ChatGPT users and to people whose data may have ended up in the material used to train models such as GPT.

The second is the legal basis for training. Large language models learn patterns by processing vast amounts of text. That capability makes it possible to draft, summarize, write code or hold a conversation, but it raises a difficult question: what happens when that text includes personal data published online or entered by users themselves?

The third is age. ChatGPT allows users to sign up from age 13, but the Garante says it has no effective way to verify that age. In the authority’s view, this could expose children to responses that are inappropriate for their level of development. Primary source

Recent security breach adds to the case

The intervention comes just days after OpenAI temporarily took ChatGPT offline because of an error in an open-source code library. The company said the bug may have exposed some users to the titles of other people’s conversations.

OpenAI also acknowledged that, during a nine-hour window on March 20, some ChatGPT Plus subscribers may have seen payment-related data belonging to other customers. This included names, email addresses, billing addresses, the last four digits of card numbers and expiration dates. The company said the number of people affected was small and contacted those who may have been exposed.

The incident is not the sole reason for the Italian order, but it provides context for the regulator’s concerns. ChatGPT does not work like a conventional search engine: users often paste documents, work-related queries, code snippets or personal information into it to get help. That makes the way conversations are stored and used a practical issue for individuals and businesses.

OpenAI has 20 days to respond

The Garante has given OpenAI 20 days to explain what measures it will take in response to the authority’s objections. If the company fails to provide a satisfactory response, the regulator can impose a fine of up to €20 million or 4% of the company’s total annual global turnover, whichever is higher under the GDPR. Primary source

OpenAI has said it complies with applicable privacy laws and is working to reduce the amount of personal data used to train its systems. The company has also argued that ChatGPT helps people in many countries and that it aims to work with regulators.

A warning for Europe’s generative AI rollout

Italy’s decision does not yet settle whether training a language model on information available online is compatible with European law. That question will likely remain in the hands of regulators, courts and lawmakers. But the order does anticipate the demands generative AI services will face in Europe: better explanations of how their data is used, tighter limits on the use of personal information and meaningful protections for minors.

For businesses, the episode is a wake-up call. Entering sensitive information into an external assistant can create privacy obligations even when the tool speeds up everyday tasks. Innovation does not eliminate those responsibilities; it makes them more urgent when an application reaches millions of users within a few months.

How to turn a regulatory order into concrete questions

The Garante’s release identifies four distinct issues: information for users and data subjects, legal basis, data accuracy and age verification. Saying only “privacy concerns” erases what would have to change. To audit an assistant, create a row for every purpose: account creation, chat delivery, history storage, model training, abuse prevention and billing. Record which data each purpose uses, for how long and on what basis.

Accuracy has a specific meaning in data protection. A model may invent information about a person and present it as real; harm can arise not only from collected data but from processing inaccurate data. Correction requires a route to access, rectify or challenge what the system associates with someone, not merely a general hallucination warning.

An age limit is not implemented by writing a number in terms. Separate declared minimum age, verification mechanism, age-appropriate design and response when a child is detected. Intrusive verification may collect even more data, so the control should be proportionate and explain what it retains.

A temporary measure has scope and a calendar

The authority gave OpenAI twenty days to report measures and described possible penalties. That was not a final judgment or a worldwide ban on ChatGPT. For any order, record the authority, territory, affected processing, recipient, provisional character, response deadline and appeal route.

A business customer cannot transfer all responsibility to the provider. It still decides which employees may paste data, which uses are allowed and how outputs are reviewed. Its inventory should include contracts, processing location, retention, subprocessors and a procedure for removing access when the regulatory position changes.

The transferable skill is to turn “a regulator is investigating” into a matrix of testable duties. Read the original document rather than the headline; separate allegation, interim measure and final decision; identify the disputed data and purpose. That record prevents both minimising the issue and presenting an open inquiry as a conviction.

Preserve the source in its original language and retain the original passage when translating. Terms such as processing limitation, legal basis and data subject have technical meanings. Rendering them as “ban”, “permission” or “user” may improve readability but must not expand the order. When a decision materially affects an organisation, interpretation belongs with specialists.

Follow-up needs a second column for the provider’s response and a third for later decisions. This separates which objection was accepted, which measure was proposed and what the authority later confirmed. The initial story opens the file; it does not close it.

This record also prevents dependence on institutional memory. A new colleague can open the order, see which system was affected and check whether the measure remains in force. If only a headline remains, the organisation cannot tell what changed or why a control was adopted.