Speech technology

What are the ethics of AI and devices that listen to children?

Published · Updated

The short answer

Devices and AI that listen to children raise questions beyond the law: who agreed, when the microphone is on, what is kept, who can hear it, and whether it works fairly for every child. Guidance from UNICEF and the UK’s children’s code points the same way: collect the minimum, keep it briefly, explain it in words a child can follow, and put the child’s interests first. A parent’s consent is often legally required, but the child’s own understanding and agreement matter too.

In the US, an app or device covered by COPPA generally needs a parent’s verifiable consent before collecting a child’s voice, because an audio file containing a child’s voice counts as personal information (FTC, 2025). The legal detail is in children’s voice data privacy; this page is about good practice on top of it.

A parent’s consent isn’t the whole picture. In pediatric medicine, the American Academy of Pediatrics describes parental permission and the child’s assent as an active process, with the child’s view counting for more as they mature (Katz et al., 2016). UNICEF’s guidance says children’s data should be used only with their informed consent, and that children should be supported to access, understand, control and delete their data in line with their age and maturity (UNICEF, 2025).

The UK’s Age Appropriate Design Code (the children’s code) applies to online services likely to be used by under-18s (ICO). For higher-risk uses of data, it suggests prompting children to talk to an adult before switching on a new use, and to stop if they’re unsure (ICO).

Always listening, or only when asked

Some devices listen all the time for a wake word; others record only while a button is held or an activity is running. Always-on listening carries a risk of recording by accident. In a study that played TV dialogue from 12 shows near smart speakers in the US and UK, the speakers woke without their wake word about once an hour (0.95 times), and on some devices 1 in 10 of these accidental recordings lasted at least 10 seconds. The authors caution that the results apply to the audio they used (Dubois et al., 2020).

The children’s code says connected devices should make it obvious when they’re recording, for example with a light, should not collect personal data in listening mode, and should be easy to switch off, ideally with a physical control (ICO).

COPPA has a narrow exception for audio used only to respond to a child’s specific request and then deleted immediately (FTC, 2025). In 2023, the FTC and the Department of Justice alleged that Amazon kept children’s Alexa recordings indefinitely and used them to improve its Alexa algorithm; under a proposed settlement, Amazon agreed to a $25 million civil penalty (FTC, 2023).

Push-to-talk, where the child or an adult presses to start, makes listening a choice the child can see. It doesn’t remove every risk, but it’s easier to explain and easier to trust.

Collect less, keep less, share less

UNICEF calls for purpose-specific, minimal data collection, no “invisible” data processing, and keeping children’s data for the shortest feasible time so it doesn’t follow them into adulthood (UNICEF, 2025). The children’s code says to collect only what’s needed for the parts of a service a child is actively and knowingly using, and to give children separate choices over each part (ICO).

Explaining it, and the line between support and surveillance

Children often don’t know who can hear them. In interviews with nine parent–child pairs (children aged 6 to 10) who played with two connected toys, some children didn’t realize the toys might be recording, and six didn’t think their parents could hear what they said to the toy unless the parent was nearby. The researchers recommended recording indicators and letting children control and listen to their own recordings (McReynolds et al., 2017).

UNICEF says a child who interacts with an AI system, such as a toy, has a right to an age-appropriate explanation of how it works and how it uses their data (UNICEF, 2025). The children’s code asks for short explanations at the moment a use of data is switched on, with cartoons, video or audio for younger children (ICO).

Monitoring has costs too. The children’s code notes that children under persistent parental monitoring may have a diminished sense of their own private space, and asks for a clear, obvious sign when monitoring is active (ICO). In an analysis of 736 app-store reviews written by children aged 8 to 19, 76% gave parental control apps one star, and children described them as invasive of their privacy and harmful to their relationships with their parents (Ghosh et al., 2018).

A practice tool that sends every recording to a parent or speech-language pathologist (SLP) can feel like being watched. Sharing summaries, and telling the child what’s shared, helps keep trust intact.

Does it work fairly for every child?

Speech recognition is still significantly less accurate for children than for adults, partly because children’s speech varies more and there’s far less child speech to train on (Shivakumar & Georgiou, 2020). Accuracy also differs between groups of speakers: in a US study of five commercial systems, the average word error rate was 0.35 for Black speakers and 0.19 for white speakers (Koenecke et al., 2020). That study was of adults; we found fewer comparable studies of children.

For speech practice the risk is specific: a tool could mark a dialect feature as an error. ASHA asks clinicians to tell the features of accent and dialect apart from speech sound disorders (ASHA). UNICEF says AI shouldn’t discriminate against children, including children who speak languages not prioritized by commercial markets, and that dataset descriptions should be explicit about how well children are represented (UNICEF, 2025).

Ask any tool who it was tested on: ages, dialects, languages and kinds of speech errors. See automated speech scoring for SLPs and bilingual children’s speech.

Disclosure

Saylo, which publishes this page, is a speech-practice wearable in development. It hasn’t been tested with children yet, and we’ll publish validation results before making any claims about accuracy or outcomes.

Common questions

Is it ethical for AI to listen to children?

It can be, with care: a parent’s consent where the law requires it, an explanation the child understands, listening only when needed, keeping as little as possible, and checking that it works fairly for children like yours.

Should a children’s device always be listening?

Always-on listening raises the risk of accidental recording. The UK children’s code says connected devices shouldn’t collect personal data in listening mode and should show clearly when they’re recording.

Does my child need to agree, as well as me?

In the US, COPPA requires a parent’s consent, not the child’s. But UNICEF’s guidance and pediatric medical ethics both give a child’s own understanding and agreement growing weight as they mature.

Is speech AI accurate for children with different accents?

Not reliably. Speech recognition is less accurate for children than adults, and a study of adults found large differences between groups of speakers, so ask who a tool was tested on.

Sources

Join the Early Adopter Program