Garante Orders Character.AI to Fix Age Verification Within 120 Days as the OpenAI Fine Falls on Competence

What Has Changed

On 3 July 2026 Italy's data protection authority, the Garante per la protezione dei dati personali, adopted decision no. 487 against Character Technologies, Inc., the Menlo Park company behind Character.AI¹. The headline figure, announced on 9 July, is a fine of €158,000². That number is the least useful thing about the case.

The part that binds is the order. According to the Garante's press release, Character Technologies must make its age-verification systems work correctly, make effective the mechanism that stops blocked minors from simply registering again (the so-called cooling-off period), and set minors' profiles to private by default. Within 120 days of receiving the decision it must tell the Garante what it has done².

Read alongside the Rome court's annulment of the Garante's €15 million fine against OpenAI on 18 March 2026⁵⁶, the pattern for any enterprise that runs, buys or embeds a conversational AI service in Italy is now clear enough to act on. The Garante's reach is strongest against providers with no main establishment in the EU, its sanctions in this sector are small relative to the companies involved, and its real lever is the operational order. Where a provider has an EU main establishment, the Garante's route to a fine has narrowed sharply.

The practical consequence: stop measuring Italian chatbot exposure by fine size, and start measuring it by whether your service or your supplier could satisfy a behavioural order with a fixed deadline.

The Instruments, and What Each Actually Says

Four Garante actions and one court judgment make up the record. They are not equal in authority, and the differences matter.

Character.AI: the instrument and its limits

The decision records that the Garante notified its objections on 26 January 2026 under article 166(5) of the Italian Privacy Code¹. The objections covered five areas: transparency under GDPR articles 12, 13 and 14; the absence of adequate notice and of an objection mechanism for pre-training data under articles 14 and 21; protection of minors under articles 24(1) and 25(2), "at least until 8 April 2025"; a late data protection impact assessment under articles 5(2) and 35(1), the first version of which was dated 14 November 2024; and the late appointment of an EU representative under article 27¹.

The decision establishes jurisdiction under GDPR article 3(2): the service was offered to people in Italy, in Italian, at least from 8 April 2024¹. That is the targeting route. It applies because Character Technologies has no main establishment in the Union, so no other authority leads.

One limit must be stated plainly. The published decision text available for this brief could be read only in part: the operative section, the fine clause and the Garante's conclusion on each objection were not visible. The fine, the three orders and the 120-day deadline are therefore taken from the Garante's own press release², which is an official statement of the authority but is not the decision. In particular, whether the Garante upheld the article 21 objection on pre-training data cannot be confirmed from the text reviewed. The company argued that the European Data Protection Board opinion relied on was adopted after its pre-training and could not apply retroactively¹. Treat the outcome on that point as unverified.

DeepSeek: the strongest tool, used against the most distant provider

The DeepSeek decision is the clearest statement of what the Garante will do when a non-EU provider does not engage. Acting under GDPR article 58(2)(f), it ordered "la limitazione definitiva" of processing of personal data of people located in Italy, with immediate effect from receipt³. The concerns cited included failure to explain the processing (article 31), a privacy notice available only in English (articles 12 to 14), unspecified legal bases (article 6), storage of data in China (article 32) and no EU representative (article 27)³.

Two features deserve attention. The order was made "con riserva": it reserved any further decision until the investigation concluded³. And the decision states that non-compliance engages the criminal sanction in article 170 of the Privacy Code, alongside administrative fines³. For a deploying enterprise, that converts a supplier's regulatory problem into a continuity problem: a service limited in Italy is a service you cannot lawfully keep feeding Italian users' data.

Replika: a fine, and a question left open

The Garante announced on 19 May 2025 that it had fined Luka Inc., the US operator of Replika, €5 million⁴. The press release gives three grounds: as at 2 February 2023 the company had not identified the legal bases for its processing; its privacy policy was inadequate in several respects; and it had no mechanism to verify users' age⁴. It also opened a separate investigation into how personal data are processed by the generative AI system behind the service⁴. The lawfulness of model training was therefore not decided in that case.

OpenAI: annulled on competence, not on the merits

On 18 March 2026 the Court of Rome annulled the Garante's decision against OpenAI, setting aside both the €15 million fine and the order to run a media campaign⁵. The source for that statement is Wilson Sonsini, which represented OpenAI; it is an interested party, though the annulment itself is not in dispute.

The ground matters more than the outcome. According to PPC Land's account of the judgment, which this brief could not check against the judgment itself, the Tribunale di Roma (R.G. 4785/2025) held that the Garante lacked competence: the Irish Data Protection Commission had become OpenAI's lead supervisory authority on 15 February 2024, nine months before the Italian decision⁶. The court examined only that ground and did not reach the alleged GDPR breaches⁶. Neither source reports whether the Garante has appealed.

Who Is Bound, From When

The obligations are the GDPR's, in force since 25 May 2018. What the Garante's decisions show is how they are applied to chatbot providers.

• Providers with no EU main establishment that target Italian users are directly within the Garante's reach under article 3(2). Character Technologies, Luka and DeepSeek all fall here. For them the Garante acts as a full authority, with power to fine, order and limit.
• Providers with an EU main establishment fall under the one-stop-shop. On the Rome court's reasoning as reported, the Garante cannot itself decide a cross-border case once a lead authority is in place⁶. Italian complaints still exist, but they travel through the lead authority.
• The order's addressee, in the Character.AI case, has 120 days from receipt of the decision to report the measures taken². The receipt date is not public, so the deadline cannot be fixed from outside.
• Deploying enterprises are not the addressees of any of these decisions. Their exposure is indirect: through supplier continuity, and through their own GDPR duties as controllers wherever they embed a chatbot in their own service.

The Garante's chatbot record turns on one question: does the provider have a main establishment in the EU? Without one, the provider faces the Garante's full toolset. With one, the Rome court has, as reported, held that the Garante cannot decide the case itself.

What Compliance Concretely Requires

Across the four Garante actions, the same operational demands recur. They are what a chatbot provider facing Italian users must be able to show, and what a buyer should be able to verify.

1. Age assurance that works in practice. Replika had none⁴. In the Character.AI case the order is not to adopt age verification but to make it function correctly, and to make the cooling-off block on rejected minors effective². A self-declaration screen is not the standard being enforced.
2. Protective defaults for minors. Minors' profiles must be private by default².
3. A privacy notice Italian users can read. DeepSeek's English-only notice was a specific concern³. The Character.AI objections covered the 2023 and 2025 privacy policies¹.
4. An identified legal basis for each processing operation. This was the first ground against Replika⁴ and a concern in DeepSeek³.
5. A DPIA before launch, not after scrutiny begins. Character.AI's first DPIA dated from November 2024, after a 2022 beta launch¹.
6. An EU representative under article 27 for non-EU providers. Both DeepSeek and Character.AI were challenged on this¹³.
7. Engagement with the regulator. DeepSeek's limitation followed its failure to explain its processing to the Garante (article 31)³. Character.AI's lower fine came, by the press release's own account, after the Garante took note of measures adopted during the investigation².

What Enforcement Has Actually Done

The fines are modest. €158,000 for Character.AI and €5 million for Replika are not sums that change a well-funded provider's economics. The one large fine, €15 million against OpenAI, no longer stands⁵.

The orders are what bite. An order to fix age verification within a fixed window, with a reporting duty to the authority, requires engineering change and leaves a written record that the Garante can test. A definitive limitation, as with DeepSeek, removes the Italian market altogether and carries criminal exposure for non-compliance³.

Where It Does Not Reach

• EU-established providers. After the OpenAI judgment, the Garante's direct route against a provider with an Irish or other EU main establishment is closed for cross-border processing, unless an appeal changes the position. Whether one has been lodged is not known from the sources⁶.
• The merits of OpenAI's alleged breaches. These have not been tested in court⁶.
• Training lawfulness. The Replika training investigation is open⁴, and the Character.AI article 21 outcome is unverified here.
• Other EU authorities. These are Italian decisions. They do not bind other national authorities, although the GDPR provisions they apply are the same everywhere.

The Judgement

Three corrected assumptions follow, and one test.

First, a small fine is not a small risk. The binding part of the Character.AI decision is an operational order with a reporting deadline. Assess Italian exposure by whether the order could be met, not by the size of the penalty.

Second, the supplier's legal structure is part of your risk. A chatbot provider with no EU main establishment faces the Garante's full powers, including limitation of processing with immediate effect. If you deploy such a service to Italian users, a DeepSeek-style order would cut off your service, not just theirs.

Third, the OpenAI annulment is not a verdict that the conduct was lawful. It decided who was competent, and nothing else.

“A small fine is not a small risk; the binding part is the order, and the order has a deadline.”

The test for any procurement or internal review of a conversational AI service used in Italy: can the provider show, today, working age assurance with an effective re-registration block, private-by-default settings for minors, an Italian-language privacy notice, a documented legal basis per processing operation, a pre-launch DPIA and, if it is outside the EU, a named article 27 representative? Those are the points the Garante has repeatedly raised. A provider that cannot evidence them is the one most likely to receive an order it cannot meet in 120 days.

This brief describes regulatory decisions and their published effect. It is not legal advice.