Voice-based age checks for children fail silently in three ways

A preprint tested three audio AI models on recordings of children under 13. It found frequent errors. It also found three faults that no standard monitoring would catch.

6 min readIIBenchmarkedPreprint, not peer reviewed
Save
Two tin cans connected by a taut string on off-white paper, with a tiny, tight blue knot halfway along.AI illustration
Illustration: AI generated

What the study shows

The authors tested three speech-capable AI models: Google's Gemini 2.5 Flash, OpenAI's GPT-4o Audio Preview and its successor gpt-audio, and Nvidia's Nemotron 3 Nano Omni. These models take in audio directly rather than working from a written transcript. The researchers played them recordings from two collections of American children's speech. One had 28 children aged 5 to 10. The other had 250 pupils aged about 8 to 11. This is a preprint and has not been peer reviewed.

For each clip, the models were asked three things. Is the speaker over 13? Is the speaker over 18? What is the speaker's age? Every speaker was under 13, so any 'yes' was an error. The authors set their own pass marks: at most 5% wrong on the over-13 question and 1% on the over-18 question. They note that no law sets an acceptable error rate.

The authors report that the models called children over 13 on between 2.5% and 42.5% of usable answers, depending on model and condition. Only 3 of 12 model and condition combinations met the 5% mark. Every model that answered placed some children over 18. The risk per child was higher than per answer. In one test, Gemini was wrong on about 6% of answers, but about two thirds of children were called over 13 at least once across 27 tries.

The first hidden fault was in Google's software library, the code that packages audio before sending it. One version encoded the audio twice, so the model received gibberish. It still answered confidently, describing 7-year-olds as adults in their thirties and forties. It took the team months to find the cause.

The second fault was a model swap. OpenAI retired GPT-4o Audio Preview during the study and replaced it with gpt-audio. On the same clips, gpt-audio refused every yes-or-no age question and never gave an age. It kept asking for the clip it had already been sent.

The third fault was self-contradiction. A small share of answer pairs said a child was not over 13 but was over 18. Nemotron's yes-or-no answers and its numeric age estimates often disagreed by large margins. None of the three faults produced an error message or a change in output format.

Hearing the audio helped compared with reading a transcript. From transcripts of 30-second clips, Gemini and GPT-4o called more than half the children over 13. From the audio of the same speech, those rates fell to about 12% and 3%.

Our reading

In our reading, the headline lesson is not that these models are poor age judges, though on this evidence they are. It is that the failures were invisible. A broken library, a swapped model and inconsistent answers all looked like normal, well-formed output. A team that checked only status codes and answer formats would have shipped any of them.

The library bug is the most striking case. The model was effectively hearing noise, yet it gave fluent, specific reasons for its wrong answers. That is a warning about trusting a model's explanation as evidence that it processed the input. The fix was not a better model. It was checking that the bytes sent were the audio intended.

A concrete example: a voice tutoring app for primary pupils could validate its age gate in spring, pass a review, and then lose all protection when the provider quietly replaces the model in summer. Nothing in the app's logs would change. The authors' call to test the whole pipeline, and to retest on a schedule, follows directly from what they saw.

Free membership

Read the rest of this story

Membership is free. You get every article in full and the Issue every Sunday.

We send a sign-in link, no password needed. Already a member? Sign in

Readiness assessment

II
Benchmarked
LevelStatusEvidence
I  Lab onlyMetThe work is a research evaluation of third-party models on two existing children's speech corpora. It proposes no new method.
II  BenchmarkedMetResults are reported on two public speech corpora (Kids VA, and MyST from the Linguistic Data Consortium), with full denominators, bootstrap intervals and per-child rates. Caveat: this is the authors' own test with self-set pass marks, not an established leaderboard, and no baseline model was run. It is a fair benchmark of the models, not a benchmark of a new technique.
III  Open sourceMetAnalysis code, coded model responses and figures are at github.com/cmstewart/speech-age-estimation. The audio is not redistributed. The models tested are third-party and mostly closed. The released code supports checking the paper, not deploying a system. The two-question rule is simple but untested outside these logs.
IV  In productionNot yetNo named organisation is documented using any method from this paper in production. The paper shows how failures in existing pipelines would go unseen, but it does not report a deployment.

What would move it up
The work already sits at the level its evidence supports for a measurement study. To be stronger evidence for deployment, it would need adult and teenage speech to measure false negatives, a dedicated speech age model as a baseline, and a held-out test of the two-question rule. It would reach level 4 only with a documented production deployment by a named organisation.

The study

Christopher M. Stewart, Kelsey Kraus, J. Elliott Casal et al.. Three Invisible Failure Modes in Speech-Based Age Estimation for Child Safety. arXiv, 13 November 2026. Read the paper

Preprint, not peer reviewed · Licence arXiv (no reuse)

Notes

  1. Stewart, Kraus, Casal and Brown, 'Three Invisible Failure Modes in Speech-Based Age Estimation for Child Safety', arXiv preprint, November 2026. Accepted at the NeurIPS 2026 Workshop on Child Safety in AI. Not peer reviewed as a journal paper. Source
  2. Analysis code, coded model responses and figures. The audio itself is not redistributed. Source
  3. Exact figures: over-13 error rates ranged from 2.5% (GPT-4o, 30-second MyST clips) to 42.5% (GPT-4o, 6-second MyST clips). Gemini in the noise study: 6.1% per response, 67.9% of children wrong at least once across 27 queries. Over-18 pass mark exceeded in 10 of 12 combinations.
  4. Library fault: google-genai 0.4.0 double-encoded audio. Mean estimate 36.8 years against a true mean of 7.4. One 7-year-old was rated 48 under 0.4.0 and 7 under 2.0.0. Payloads were 472,080 against 354,060 characters, a ratio of four-thirds. Source
  5. Model swap: gpt-audio refused all 924 attempts at each threshold question on the scripted Kids VA clips. Contradictions: 264 of 8,340 answer pairs (3.2%) said not over 13 but over 18. Nemotron said over 13 for 39.5% of spontaneous clips but estimated over 13 for 8.9%.
  6. Refusal handling: GPT-4o on Study 2B over-13, 42.5% valid-only, 14.7% fail-safe, 80.1% fail-open. Two-question rule: below 3% in ten of twelve audio cells; 23.6% of children still passed at least once for Nemotron on Study 2B.

Edited and signed by

Stefan PerdalEditor and publisher86 stories edited

Stefan read the study, set the readiness level and approved this story. An AI system found the paper and wrote the first draft.

How this was made

An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.

More on Safety & evaluation