COPPA and AI Products Used by Children Under 13

COPPA is the only US privacy statute with a real regulator, real civil penalties per violation, and a rule that says in terms how long you may keep the data. For anyone building a conversational product, that combination makes the retention provisions more consequential than the consent ones, which is the opposite of how the statute is usually summarised.

When the Rule applies to an AI product

The Children’s Online Privacy Protection Act is at 15 U.S.C. §§ 6501–6506, and the operative instrument is the FTC’s COPPA Rule at 16 C.F.R. Part 312. It applies to an operator of a website or online service directed to children under 13, and to an operator of a general-audience service that has actual knowledge that it is collecting personal information from a child under 13.

Both prongs matter to AI products in different ways. Whether a service is directed to children is decided under the multi-factor test at § 312.2 — subject matter, visual and audio content, animated characters, child-oriented activities, music, the age of models, presence of child celebrities, advertising placement, and competent and reliable empirical evidence about audience composition. A homework-help assistant with cartoon characters is directed to children whether or not its terms say users must be 13.

The actual-knowledge prong is the one that catches general-purpose assistants. A conversational product is unusually likely to be told directly: a user types that they are in fourth grade, or that they are eleven. Whether that constitutes actual knowledge under the Rule depends on facts, and the answer is not settled for a system where the statement was processed by a model rather than read by a person. Building a classifier that detects self-declared age is a reasonable safety measure and simultaneously creates the knowledge that triggers the obligations — a genuine tension in the design, and one that the Rule does not resolve for you.

“Personal information” at § 312.2 is broad and includes persistent identifiers usable to recognise a user over time, photographs, video and audio files containing a child’s image or voice, and geolocation sufficient to identify a street and city. A voice assistant collects personal information the moment it records.

What verifiable parental consent requires

Section 312.5 requires an operator to obtain verifiable parental consent before any collection, use or disclosure of personal information from a child, and § 312.4 requires direct notice to the parent and a clear online privacy policy describing what is collected and how it is used.

“Verifiable” is not satisfied by a tick box. Section 312.5(b)(1) sets the standard — any method reasonably calculated, in light of available technology, to ensure the person giving consent is the child’s parent — and § 312.5(b)(2) enumerates approved methods, including a signed consent form returned by post, fax or electronic scan; a monetary transaction using a credit or debit card with notification of the transaction; a toll-free telephone call to trained personnel; a video conference with trained personnel; verification against government-issued identification checked against a database and then promptly deleted; and knowledge-based authentication with sufficient difficulty. The Rule also has a process at § 312.12 for applying to the Commission to approve a new method.

There is a narrower path for schools. The FTC has long taken the position that a school may provide consent on parents’ behalf where an operator collects personal information from students solely for the use and benefit of the school and for no other commercial purpose. That qualifier is the whole condition, and it is why a classroom product that also uses interactions to improve a general-purpose model cannot rely on it. The school-consent path also interacts with FERPA’s separate requirements, discussed in the school official exception page.

The 2025 amendments

The FTC announced final amendments to the COPPA Rule on 16 January 2025; they were published in the Federal Register in April 2025, took effect in June 2025, and gave operators until April 2026 for full compliance with the changed provisions. The amendments are the first substantial revision since 2013, and four changes matter for AI products.

• Separate consent for third-party disclosure. Consent to collect is no longer consent to disclose to third parties for advertising or other purposes; that requires its own opt-in. For a product that sends conversation content to an external model provider, the question of whether that is a disclosure requiring separate consent is a live design issue, and the answer depends on the provider’s role and terms.
• A written children’s data security programme. Section 312.8 now requires a documented programme with designated responsibility, risk assessment, safeguards, testing and vendor oversight, rather than a general reasonable-security duty.
• A written, publicly disclosed retention policy. Section 312.10 requires that personal information be retained only as long as reasonably necessary for the specific purpose for which it was collected, prohibits indefinite retention, and requires the retention policy to be stated publicly.
• An expanded identifier definition, reaching biometric identifiers derived from voice, image or other data, which directly implicates voice assistants and any product deriving a speaker or face embedding.

Retention, deletion and training data

Section 312.10 is where the Rule and machine learning collide hardest. Retention is permitted only for as long as reasonably necessary for the purpose of collection. “We might want it later to train a better model” is not a purpose of collection stated to a parent, and indefinite retention is prohibited outright. Section 312.6 gives the parent a right to review what has been collected and to direct its deletion, and to refuse further collection while allowing the child to keep using the service where practicable.

The FTC has treated this as more than a records question. In its 2023 action concerning Amazon’s Alexa service, the complaint alleged that the company retained children’s voice recordings and transcripts indefinitely and used them to improve its speech recognition, and failed to delete them when parents asked. The resulting order required a $25 million civil penalty and, importantly, the deletion not only of the retained data but of models and algorithms derived from it. The same remedy — sometimes called algorithmic disgorgement — had appeared in earlier FTC orders and appears again in the 2023 order concerning Rite Aid’s facial recognition system, covered in the page on that order.

The engineering consequence is that a deletion pipeline which removes rows from a database is not sufficient to satisfy an order of that shape. If children’s data can enter a training corpus, you need to be able to say which model versions were trained on which corpus snapshots, which is a lineage problem rather than a deletion problem. The cheap way to avoid it entirely is a hard architectural rule that data from users identified as under 13 never reaches any training or evaluation path, enforced at ingestion rather than at export.

Enforcement posture and the companion-bot inquiry

COPPA is enforced by the FTC and by state attorneys general, and violations are treated as unfair or deceptive acts under Section 5 of the FTC Act, carrying civil penalties per violation. The penalty amount is adjusted annually for inflation, and because each affected child can constitute a violation, the theoretical exposure for a consumer product is large enough that settlements have reached hundreds of millions of dollars — the 2022 Epic Games matter included a $275 million civil penalty for COPPA violations.

In September 2025 the Commission opened a study under Section 6(b) of the FTC Act into AI companion chatbots and minors, issuing orders to several major providers seeking information about how such products are designed, monetised and tested for effects on children and teenagers. A 6(b) study is a compulsory information-gathering process, not an enforcement action and not a finding of wrongdoing. It is worth tracking anyway, because 6(b) studies historically precede both enforcement and rulemaking, and because the orders themselves reveal what the staff considers relevant.

Two practical points for anyone shipping a general-audience assistant. First, the deception theory under Section 5 is often easier for the Commission to run than the COPPA theory: if your interface states that the product is not for under-13s while your own metrics show a substantial child audience, that is a misrepresentation claim independent of COPPA’s technical requirements. Second, an age-assurance mechanism that collects government identification to keep children out creates its own privacy exposure and, in several states, its own statutory duties. There is no configuration of this problem that collects no data; the choice is about which data and under which rule.