Children's voice data and AI training: why this is the highest-risk corpus

Four risk factors stack on the same corpus

Every data category has a risk profile. Children's voice concentrates four factors that are usually spread across different categories:

• The data subject cannot consent, so the consent has to come from someone else and has to be verified as coming from that person.
• A recording of a child's voice is treated as personal information by definition in several regimes, without any analysis of whether it could identify anyone.
• Voice supports identification and imitation in ways that a photo or a text sample does not, and the imitation is now cheap to produce.
• A synthetic voice built from children's recordings creates a capability — an adult speaking in a child's voice — that the rules were not written to address.

The rules, stacked

Three layers apply at once, and they do not line up with each other. In the United States, the children's privacy rule covers collection from children under a defined age, treats an audio file containing a child's voice as personal information, and requires verifiable parental consent — a standard a signature collected on site does not automatically meet.

In the EU, the age at which a child can consent to online services is set by each member state within a range, so a corpus collected lawfully from thirteen-year-olds in one country may not be lawful in the next one. Several US states set their own thresholds as well, and those thresholds differ from the federal one.

The EU AI Act adds something different in kind: a prohibition rather than a governance duty, aimed at systems that exploit vulnerabilities connected to age. That is a ban on a use, it applies without any risk classification, and it means the intended use of a children's voice model is in scope before any paperwork question is even reached.

The operational problem is age, not forms

Consent forms are the easy part. Knowing who is a child is the hard part:

• Self-declared age is not verification, and a corpus built on it cannot be documented in either direction.
• School-mediated collection produces a permission slip from an institution, which is not the same thing as a guardian's consent to AI training.
• Dubbing and animation sessions may involve child performers whose guardian signed a production contract that never mentions training or synthetic voice.
• User-contributed recordings arrive with no age data at all, which makes the whole batch unclassifiable rather than merely uncertain.

The discipline that follows, and what the guardian must be told

Record the verification method, not only the form. A file that contains signed guardian consents but no description of how the guardian's identity was checked cannot answer the question a buyer will ask, and the answer cannot be reconstructed later.

Treat an unverifiable age as an exclusion rather than a risk to be noted. Where the supplier cannot show how ages were checked, the correct treatment is to leave that material out of the delivery, not to describe it as probable adults.

Keep children's material in its own batch and its own corpus. Mixing it with adult recordings means an age problem contaminates an otherwise clean delivery, and unpicking a mixed corpus after the fact is the most expensive form of remediation there is.

None of that is worth much if the guardian was not told what they were agreeing to. A guardian cannot consent to a use that was never described, so the description needs to cover, in language a non-specialist can act on:

• What the recordings will train, and whether a synthetic voice will be created from them.
• Which organisation will hold the recordings after the session, and whether they can be passed on, sublicensed or sold.
• How long the recordings will be kept, and what happens at the end of that period.
• Whether the child will be identifiable in connection with the model, or whether the voice is used with no name attached to it.
• How to withdraw, and what withdrawal can and cannot achieve once training has happened.

The commercial argument for strictness

The usual reason given for caution is regulatory. The stronger reason is commercial, and it holds even where no regulator is looking:

• A corpus with unverifiable ages cannot be warranted, and an unwarrantable corpus is priced as a liability rather than as an asset.
• A discovered consent defect in children's data is not fixable by topping up paperwork, because the affected people may no longer be reachable. The remedy is removing the data, which means a retraining event.
• Buyers increasingly ask for the age-verification method as a standard due diligence question, so a supplier that cannot answer loses the deal before any legal question is reached.

The honest summary

A children's corpus assembled properly — verified ages, guardian consent that names the use, a stated retention period, a working withdrawal route — is a legitimate and scarce product. Child speech is genuinely needed for speech recognition aimed at children, for reading assessment tools, and for accessibility work, and those uses do not become illegitimate because the data is hard to document.

The scarcity is a documentation problem rather than an ethical one. The documentation is produced at collection or it is not produced at all, which makes this the category where the earliest decisions matter most and where a project that starts clean stays licensable for years.

This is an operational overview rather than legal advice. Children's data rules differ by jurisdiction and by the age of the child, and applying them to a specific corpus is work for counsel.