The Receipts Room
If your teenager has an AI chat app on their phone, you have probably been told the same thing most parents are told: there’s an age limit, so younger kids can’t get in.
In April 2025, staff at Italy’s data protection authority, the Garante per la protezione dei dati personali, decided to check. They went to Character.AI’s website and created test accounts. One of them said it belonged to a 15-year-old.
The company said users in Europe under 16 could not create an account. The test account went through anyway. Its new profile was set to public. A parent’s email address was offered as an option, not required (Garante, 2026b, s. 3.4).
My proposition is simple: an age limit nobody tests is a promise, not a protection. This decision is what it looks like when a regulator stops reading the promise and starts testing it.
The receipt: a fine, and a list of orders
On 3 July 2026, the Garante fined Character Technologies, the US company behind Character.AI, €158,000. It ordered the company to make its age checks work properly, stop blocked minors from simply signing up again, and keep minors’ profiles private by default (Garante, 2026a; 2026b).
Character.AI is a generative AI service where users create and chat with virtual characters. The Garante describes it as an entertainment service that minors can access (Garante, 2026a). It has been available in Italian since 8 April 2024 (Garante, 2026b, s. 1).
Why a privacy regulator, of all people
The Garante enforces Europe’s General Data Protection Regulation (GDPR). It opened this investigation on its own initiative, sending its first request for information in November 2024 (Garante, 2026a; 2026b, s. 1).
Data protection law might not seem the obvious tool for child safety. But under the GDPR, a company must design its service to protect people’s data from the start, and must take account of the risks to the people who use it. Children are the users the law treats with the most care.
The gate that let a 15-year-old through
Since October 2023, Character.AI had used an “age gate”: new users typed in their date of birth, and, according to the company, anyone under 16 in the European Economic Area was blocked (Garante, 2026b, ss. 1.2.1, 3.4).
A self-declared birth date is the weakest form of age check, since anyone can type a different year. But the Garante’s test found something more basic. Nobody had to lie. A user declaring an age of 15 was able to register, even though that was below the company’s own stated threshold. A test user claiming to be under 13 was blocked (s. 3.4).
The Garante concluded that the company had breached its duties to put in place appropriate measures (Article 24) and to protect data by default (Article 25). The decision acknowledges there is no single EU standard for checking age, but says the law still requires measures that are adequate and can be shown to work (s. 3.4). It found that vulnerable users “sono stati esposti ad un rischio particolarmente elevato”: they were exposed to a particularly high risk (Garante, 2026b).
Key idea: Privacy by default
Privacy by default means a service starts with the most protective settings switched on, so people are protected even if they never open the settings menu. It matters most for children, who are least likely to change a default and most exposed if they don’t. A profile that is public until someone makes it private puts the burden on the user. A profile that is private until someone chooses to share puts it on the company, which is where European law places it.
The paperwork arrived after the kids did
The other findings are less vivid but tell the same story: safeguards arriving after the service.
- The risk assessment. Services that carry high risks for people’s data need a data protection impact assessment (DPIA) before processing begins. Character.AI’s first DPIA was dated 14 November 2024, after the investigation had started, and did not assess the risks to minors or the use of data to train its AI models (Garante, 2026b, s. 3.5).
- The EU representative. Companies outside the EU that serve Europeans generally need a representative inside the EU whom regulators and users can contact. Character.AI’s was formally designated on 31 May 2025 (s. 3.6).
- The privacy notice. For more than a year after the service launched in Italian, its privacy policy was available only in English. A later Italian version still contained misleading or unclear passages (s. 3.2).
- Training data. The Garante also found that Italian users had not been properly told about the use of data to pre-train the company’s AI models (s. 3.3). It did not uphold a separate alleged breach relating to the right to object (s. 4).
Who benefited from that order of events? A service that grows first and documents later reaches users faster. Who carried the risk in between? The teenagers whose profiles were public while the paperwork caught up.
The company’s side
To be fair to Character.AI, the decision records real changes it made along the way. In November 2024 it introduced a separate AI model for users under 18. In March 2025 it added a feature giving parents some insight into their children’s use. In November 2025 it removed open-ended chat for minors (from February 2026 in Italy) and introduced a more advanced age assurance system, which estimates age and can call on a third-party check if a user disputes the result (Garante, 2026b, ss. 2.2, 3.4).
The company argued that it had acted in good faith, that no harmonised standard existed and that it had cooperated (s. 2.2). The Garante credited these steps but still found it necessary to order more (Garante, 2026a).
Under the orders, Character Technologies must (Garante, 2026a; 2026b, s. 5):
- make sure its age checks actually work for Italian users;
- make its “cooling-off period” effective, so a minor who has been blocked cannot simply try again with a different birth date;
- set minors’ profiles to private by default;
- fix its privacy notice, and deal lawfully with the Italian users’ data previously used for training, deleting it if there is no lawful basis to keep it.
It has 120 days to tell the Garante what it has done (Garante, 2026a). The Garante also ordered the decision to be published.
The verdict
The fine is modest for a global technology company. The significance is elsewhere. This is a regulator treating a broken age gate and a public-by-default profile not as a parent’s worry but as a legal finding, using ordinary data protection law that applies across Europe.
And it got there by doing something any company could do for itself: testing its own controls the way a child would. That is the part that should embarrass the industry. The test took no special tools. It took someone willing to try.
What we still don’t know
The decision does not identify any specific harm to an Italian child. Its concern is risk: that children could get in, and that once in, their profiles were exposed. The number of Italian users is redacted. The decision reflects the service as tested in April 2025 and the company’s changes since; we have not tested whether today’s age assurance works, and the Garante’s 120-day check was still to come when we wrote this. We do not know whether the company will appeal. Our summary relies on the Italian original; any translation in this post is ours.
The open question is the one worth watching: when the company reports back, will anyone run the 15-year-old test again?
Questions you can ask
- If you are a parent: ask your child which AI chat apps they use, then check together whether their profile is public or private, and what the app does with their conversations.
- If you are a teenager: look for privacy settings in any AI chat app and set your profile to private. You can ask the company what data it holds about you.
- If you run a service young people might use: try signing up as a child. If it works, so will everyone else’s attempt. Check what a new account looks like to the world before its owner touches a setting.
- If you launch AI features in Europe: complete your risk assessment before launch, cover minors and training data in it, and have your EU representative in place on day one.
Go further. Operational AI Readiness is FairByDesign’s field manual for AI that is already live, and the complete Standard edition is free to read and download. Chapter 3, Classifying Decision-Influencing AI Without Lying to Yourself, covers why a risk assessment written after launch is a warning sign, and Chapter 4, The Minimum Viable Control Stack, covers the basic controls, such as protective defaults, that should be in place before users arrive. Read it free, or test one workflow with the free readiness assessment, paying close attention to domain 1, inventory and scope, and domain 8, validation and failure readiness.
Your next step: if you want plain-language summaries of decisions like this one, the free Mercury Brief brings our latest findings to your inbox. Know a parent with a teenager online? Send them this.
FairByDesign is reader-funded. If this helped, get the free Mercury Brief or see ways to support. Support never buys editorial control.
References
Bakos, V. (2026). Operational AI readiness: A practitioner guide to governing AI that is already live. FairByDesign. https://fairbydesign.org/operational-ai-readiness/
Garante per la protezione dei dati personali. (2026a, July 9). Intelligenza artificiale: il Garante privacy sanziona Character.AI [Press release; doc. web n. 10269594]. https://www.garanteprivacy.it/home/docweb/-/docweb-display/print/10269594
Garante per la protezione dei dati personali. (2026b, July 3). Provvedimento del 3 luglio 2026 [Character Technologies, Inc.] (Registro dei provvedimenti n. 487; doc. web n. 10269571). https://www.garanteprivacy.it/home/docweb/-/docweb-display/print/10269571




Leave a Reply
You must be logged in to post a comment.