Research finds three hidden failures in audio models estimating children's age from speech
The study tested three audio-capable language models on estimating children's age from speech. The models classified children under 13 as older than 13 in 2.5–42.5% of cases, some even as adults. Three silent failures occurred without any error messages.
A research team evaluated whether three audio-capable large language models can estimate a child's age from a recording of their speech. The tests were conducted in five studies on two corpora of speech from children under 13. The models classified children as older than 13 in 2.5 to 42.5% of audio responses, and every model that provided an answer estimated the age of some 5- to 11-year-old children as over 18.
Three distinct failures emerged during testing. For one model, the client library silently double-encoded the speech samples, causing the model to label 5- to 10-year-old children as people aged thirty to forty, including hallucinated explanations for the estimate. The endpoint of a second model was replaced by a new version that refused to answer almost all questions about age and never produced any estimate. A third model produced contradictory responses – the same speech sample was classified as under 13 in one instance and over 18 in another, with the model even giving threshold answers that contradicted its own numerical estimate.
According to the authors, none of these three failures triggered a warning, an error code, or a change in output format, so they would pass through standard monitoring undetected. They conclude that claims about the reliability of age verification for child protection need to be evaluated at the level of the entire processing pipeline, not just the model itself. The study describes reporting practices that would make it possible to detect such failures.
Why it matters
Companies deploying audio AI models in products for children or in age verification systems may rely on tests of the individual model, while the actual errors arise in the surrounding infrastructure – input encoding, endpoint swaps, or response inconsistency. Such errors do not trigger standard error signals, so routine monitoring fails to catch them, and they can remain hidden until an incident occurs.
Relevant practical impact
What this means
For a business
Companies deploying audio LLMs for age estimation or child protection bear the risk that silent pipeline failures (encoding errors, endpoint swaps, contradictory responses) produce no error message and are not caught by standard monitoring, representing a hidden compliance risk.
Risks and complianceCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.