Child safety in AI depends on the accurate identification of minor users. Young children tend to talk rather than type when interacting with AI, so in practice this identification requires predicting age from speech. We evaluated whether three audio-capable large language models can predict a child's age from their speech across five studies using two speech corpora of children under the age of 13. Models classified children as over 13 on 2.5% to 42.5% of audio responses, and every model that answered placed some 5- to 11-year-olds over 18. While measuring these rates, we uncovered three distinct failure modes. A client library silently double-encoded speech samples, leading one model to classify children aged 5 to 10 years old as being in their thirties and forties with hallucinated explanations. A second model's endpoint was replaced by a successor that refused nearly every age-related question and never produced an age. Every model sometimes placed the same speech sample under 13 and over 18 years of age, but a third model went further, giving threshold answers that contradicted its own numeric estimates. None of these failure modes produced a warning, a status code, or a change in output format, so they would all pass standard monitoring. These results demonstrate why child-safety claims about age assurance should be evaluated at the level of the pipeline rather than the model. We describe the reporting practices that would make such failures detectable.
Figures & tables
Study
Corpus
Children
Duration
Speech type
Modality
1A
Kids VA
28
∼ 6 s
Scripted
Audio
1B
Kids VA
28
∼ 6 s
Spontaneous
Audio
1C
Kids VA
28
∼ 6 s
Scripted, 9 noise conditions
Audio
2A
MyST
250
∼ 30 s
Spontaneous
Audio and text
2B
MyST
250
∼ 6 s, truncated from 2A
Spontaneous
Audio and text
Table 1: Summary of study manipulations. Kids VA studies manipulate speech type and background noise. MyST studies manipulate duration and input modality.
Figure 1: Mean estimated age against each child’s true age for audio stimuli from the Kids VA corpus. Each point is one child, averaging that child’s valid responses. The solid line is the least-squares fit, the dashed line marks perfect calibration, and the dotted line marks a constant prediction of 7 years, the value that minimizes mean absolute error on this corpus. Correlations carry 95% confidence intervals bootstrapped by resampling speakers. Estimates rise with true age in every panel except Nemotron on Study 1B, but the fitted slopes fall well below 1.0, suggesting that both models compress the age range toward its middle.
Kids VA, ages 5 to 10, 6-second clips
MyST, ages 8 to 11, conversational
Condition
Model
Rate
95% CI
≥1
Maj.
Condition
Model
Rate
95% CI
≥1
Maj.
Scripted
Gemini
3.6
0.0–8.3
10.7
0.0
30 s
Gemini
11.6
8.6–14.8
23.1
9.5
Scripted
Nemotron
23.8
14.3–34.5
50.0
17.9
30 s
GPT-4o Prev.
2.5
1.0–4.4
4.1
2.5
Spontaneous
Gemini
9.5
1.2–20.2
14.3
7.1
30 s
Nemotron
5.0
2.9–7.2
12.0
2.9
Spontaneous
Nemotron
39.5
25.9–53.8
64.3
35.7
6 s
Gemini
15.5
12.4–18.7
32.0
12.0
Noise
Gemini
6.1
3.9–8.8
67.9
0.0
6 s
GPT-4o Prev.
42.5
35.3–49.8
48.5
38.0
Table 2: Incorrect over-13 classification for audio samples as a percent. The correct rate is zero. Intervals are 95% bootstraps resampling speakers. The per-child columns give the share of children misclassified under the at-least-once and majority rules. Noise pools nine background-noise conditions. The OpenAI rows for MyST are GPT-4o Audio Preview, which was retired before the Kids VA studies were rerun. Its successor (gpt-audio) refused every over-13 question on Kids VA and so has no rows.
Version
Payload chars
FLAC on first decode
FLAC on second decode
Estimated age
0.4.0
472,080
no
yes
48
2.0.0
354,060
yes
no
7
Table 3: Captured payload for one Kids VA recording of a 7-year-old under both library versions.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Label
Identifier
Provider and client
Studies
Gemini 2.5 Flash
gemini-2.5-flash
Gemini API, paid tier, google-genai
1A, 1B, 1C, 2A, 2B
GPT-4o Audio Preview
gpt-4o-audio-preview
OpenAI API, openai
2A, 2B, first-phase Kids VA
gpt-audio
gpt-audio-2025-08-28
OpenAI API, openai
1A, 1B, 1C
Nemotron 3 Nano Omni
nvidia/nemotron-3-nano-omni/audio
fal.ai, fal_client
1A, 1B, 1C, 2A, 2B
Appendix
Table 4: Audio-capable models, endpoint identifiers, providers and client software, and the studies each served. The text conditions in Studies 2A and 2B used the corresponding text endpoints gemini-2.5-flash, gpt-4o, and nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning, the last served through DeepInfra. For Gemini, the audio and text conditions used the same endpoint, so for the other two developers the modality comparison holds the model family constant rather than the endpoint.
Gemini 2.5 Flash
Nemotron 3 Nano Omni
Condition
Over 13
Over 18
Over 13
Over 18
clean baseline
3.7 [0.0, 9.9]
3.6 [0.0, 9.6]
19.0 [8.3, 32.1]
13.1 [4.8, 23.8]
white 10 dB
7.1 [1.2, 14.3]
2.4 [0.0, 6.0]
16.7 [7.1, 28.6]
7.2 [1.2, 14.8]
pink 10 dB
4.8 [0.0, 10.8]
2.4 [0.0, 7.1]
9.5 [4.8, 15.5]
7.1 [2.4, 11.9]
classroom 20 dB
6.0 [1.2, 10.8]
10.7 [2.4, 21.4]
20.5 [10.8, 30.9]
9.5 [2.4, 17.9]
classroom 10 dB
3.6 [0.0, 8.3]
6.0 [1.2, 10.7]
19.0 [9.5, 29.8]
11.9 [4.8, 20.2]
Appendix
Table 5: Incorrect over-13 and over-18 rates for each of the nine acoustic conditions in Study 1C, in percent, with 95% speaker-resampled bootstrap intervals in brackets. Each condition rests on 84 attempts per question. gpt-audio refused all 756 attempts at each question and is omitted.
Study
Model
Modality
Pairs
Children
Binary
Numeric
Both
95% CI
≥1
Majority
1A
Gemini
audio
84
28
3.6
1.2
0.0
0.0–0.0
0.0
0.0
1A
Nemotron
audio
84
28
23.8
0.0
0.0
0.0–0.0
0.0
0.0
1B
Gemini
audio
84
28
9.5
0.0
0.0
0.0–0.0
0.0
0.0
1B
Nemotron
audio
76
26
39.5
9.2
6.6
1.3–13.3
15.4
3.8
1C
Gemini
audio
749
28
6.1
1.7
0.7
0.1–1.3
14.3
0.0
1C
Nemotron
audio
754
28
17.2
1.2
0.7
0.1–1.5
14.3
0.0
Appendix
Table 6: Scoring the two-question rule on the existing logs, in percent. Pairs counts responses with a valid answer to both questions. Binary and numeric give the error rate of each question alone on those pairs. Both gives the rate at which a response answered yes and estimated 13 or more. The per-child columns apply the at-least-once and majority rules to the combined answer. gpt-audio has no row because it produced no numeric estimate.
Over 13
Over 18
Study
Model
Attempts
Excl.
Valid
Safe
Open
Excl.
Valid
Safe
Open
1A
Gemini
84
0
3.6
3.6
3.6
0
9.5
9.5
9.5
1A
GPT-4o
84
84
–
0.0
100.0
84
–
0.0
100.0
1A
Nemotron
84
0
23.8
23.8
23.8
1
12.0
11.9
13.1
1B
Gemini
84
0
9.5
9.5
9.5
0
4.8
4.8
4.8
1B
GPT-4o
84
84
–
0.0
100.0
84
–
0.0
100.0
Appendix
Table 7: Incorrect over-13 and over-18 rates, audio condition, under three treatments of excluded responses, in percent. Excluded counts refusals, API errors and unclear responses. Fail-safe scores every excluded response as a child. Fail-open scores it as an adult.
Study
Model
Rate
95% CI
n
Children
≥1 error
Majority
1A
Gemini
9.5
2.4–17.9
84
28
17.9
10.7
1A
GPT-4o
refused all
0
1A
Nemotron
12.0
3.6–23.2
83
28
21.4
7.1
1B
Gemini
4.8
1.2–9.5
84
28
14.3
0.0
1B
GPT-4o
refused all
0
1B
Nemotron
22.5
11.1–35.4
80
27
37.0
25.9
Appendix
Table 8: Incorrect over-18 classification, audio condition, in percent. Columns as in Table 2 , with valid responses and children added. Against the 1% reference point, 10 of 12 combinations exceed it per response and 6 of 12 per child under the majority rule.
Study
Model
Modality
Rate
Valid n
Refused
Error
Unclear
Attempts
1A
Gemini
audio
3.6
84
0
0
0
84
1A
GPT-4o
audio
n/a
0
84
0
0
84
1A
Nemotron
audio
23.8
84
0
0
0
84
1B
Gemini
audio
9.5
84
0
0
0
84
1B
GPT-4o
audio
n/a
0
84
0
0
84
1B
Nemotron
audio
39.5
81
0
3
0
84
Appendix
Table 9: Over-13 question across all studies, models and modalities. Rates are the percentage of valid responses answering in the affirmative. Valid n excludes refusals, API errors and unclear responses, itemized separately.
Study
Model
Modality
Rate
Valid n
Refused
Error
Unclear
Attempts
1A
Gemini
audio
9.5
84
0
0
0
84
1A
GPT-4o
audio
n/a
0
84
0
0
84
1A
Nemotron
audio
12.0
83
0
1
0
84
1B
Gemini
audio
4.8
84
0
0
0
84
1B
GPT-4o
audio
n/a
0
84
0
0
84
1B
Nemotron
audio
22.5
80
0
4
0
84
Appendix
Table 10: Over-18 question across all studies, models and modalities. Columns as in the previous table.
Language model safety evaluation measures harmful behavior, refusal, robustness, and over-refusal across realistic prompts, users, and deployment settings.