What matters in AI.

Subscribe

AI audio models rate some children aged 5 to 11 as more than 18

The models put children above age 13 in 2.5% to 42.5% of audio answers. No warning showed the faults.

Researchers tested three audio-capable large language models (AI models that use voice) on one task: find the age of a child from voice. They ran five studies with two voice sets of children below age 13. The models put children above age 13 in 2.5% to 42.5% of audio answers. Each model that answered put some children aged 5 to 11 above age 18.

How it was tested

Child safety needs a system that finds young users. Young children speak more than they type. The system must then predict age from voice.

The three faults

  • A software library encoded the voice data two times. One model then rated children aged 5 to 10 as in their thirties and forties.
  • A maker replaced one model endpoint with a new model. The new model refused almost all age questions.
  • One model gave threshold answers that did not agree with its own age numbers.

What is not known

None of the faults gave a warning, a status code or a new output format. Standard monitoring does not find them. The authors say child safety claims must be tested on the full pipeline, not only the model. The abstract does not name the models.

Insight: A silent fault can pass monitors with no hidden intent

An earlier paper (27 October 2026) made models that hide bad behavior when watched. Tests missed them. Here nobody hides anything, and a library and a replaced endpoint give the same blind spot. Vendors change parts of the pipeline without a signal, and the buyer must own the repeat test. This seems a duty to put in the contract.

Sources

  1. Three Invisible Failure Modes in Speech-Based Age Estimation for Child Safetyarxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·17 NOV2026

Posted

Tags

More in Research

All Research news