AI
Three Invisible Failure Modes in Speech-Based Age Estimation for Child Safety
Publication: cs.AI updates on arXiv.org
Author: Christopher M. Stewart, Kelsey Kraus, J. Elliott Casal, Geoffrey L. Brown
Source published: 17/11/2026, 05:00:00 UTC
Briefing created: 17/11/2026, 08:50:54 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 17/11/2026, 08:33:25 UTC.
Briefings
Three audio-capable large language models failed to accurately predict child ages from speech across five studies. One model hallucinated ages for 5-10 year olds, another refused age questions entirely, and a third gave contradictory thresholds. These failures occurred without warnings or output changes, meaning standard monitoring would miss them. The study evaluated models using two child speech corpora under age 13.
Key points
- Three distinct failure modes were identified in speech-based age estimation: silent double-encoding of samples causing age hallucinations, endpoint replacement causing refusal of age questions, and contradictory threshold answers.
- Models classified children as over 13 on 2.5% to 42.5% of audio responses, with every model placing 5- to 11-year-olds over 18 years old.
Context
This research evaluates three audio-capable large language models' ability to predict a child's age from speech across five studies using two speech corpora of children under 13, addressing critical child safety in AI.
Limitations
Evidence is limited to five studies using two speech corpora of children under 13; results are preprint status with uncertainty regarding generalizability and long-term impact.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 17/11/2026, 08:50:54 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeihzlp2rts4s2o5ns6oq4nuoygf2r5kw6goddrqhkodr6pbnc5s5da