Mind the Alignment Gap: Why General-Purpose Moderation Fails Children, and How a Child-Centric Taxonomy and Synthetic Data Close It

Key Findings

• — On the held-out test set of 7,920 synthetic interactions (4,565 safe, 3,355 harmful), harmful-class recall was 0.300 for the OpenAI Moderation API, 0.338 for base Llama Guard 3 1B and 0.381 for base Llama Guard 3 8B; base Aegis-Defensive 7B reached 0.692 (Table 4)
• — Per category, the OpenAI Moderation API missed 99% of academic dishonesty, 95% of privacy violations and 90% of grooming and predatory behaviour examples
• — LoRA fine-tuning on the synthetic data lifted Llama Guard 3 8B from 0.38 to 0.99 harmful recall at 0.997 precision (0.996 macro-F1, 0.994 harmful recall in Table 4)
• — A lexical BM25 baseline reached 0.902 macro-F1 and FAISS 0.920; the authors attribute this to distributional artefacts of fully synthetic data and state that real-world performance on noisy child input will be lower
• — The taxonomy treats anthropomorphism, misplaced trust, over-sharing and cognitive overload as dominant risks for ages 6-12, and emotional entanglement, use of the chatbot as substitute therapy, body-image harm and self-harm as dominant risks for ages 13-17

Methodology Notes

All data are synthetic; no real child conversations were used. A single generator (gemini-3-flash-preview) was chosen from six candidates (five abliterated open-weight models) by Skywork-Reward-V2-Llama-3.1-8B reward scores; the same reward model scored harmfulness and filtered the pool, dropping 4,145 weak-harmful examples (58% safe / 42% harmful after filtering, including 2,940 hard negatives). Train and test share one generator, so all scores are in-distribution; reward filtering was applied before the 80/10/10 split; there is no human annotation of labels and no prompted-frontier-model baseline. The authors list these threats themselves (Section 7.1) and describe the circularity: 'harmful' means what the reward model judges harmful. Baselines: BM25, FAISS, YouthSafe (0.893 macro-F1, applied out of distribution), OpenAI Moderation API, Llama Guard 3 1B/8B, Aegis-Defensive 7B; fine-tuned ModernBERT-large, Llama Guard 3 1B/8B and Aegis-Defensive 7B. Validation inside a children's educational chatbot run by the Czech non-profit AI dětem is planned future work. Dataset released behind a research-use agreement at github.com/fikeijan/mind-the-alignment-gap. Peer-reviewed workshop paper. Verified 2026-10-26 from the Anthology page and PDF (HTTP 200, 1,576,055 bytes). Anthology gives month precision only (2026/10); day set to 01. No arXiv version found.

Sources