1,331 publicly reported child-related AI incidents coded by hand found harm clustering in representational and dignity categories, and co-occurring with deployment and governance failures far more often than model defects
AspireAI community · arXiv (Mishra et al., The Responsible AI Initiative) / KidAID
Child safety for AI is usually assessed at the model level - benchmarks, red-teaming, content filtering - but most harm taxonomies lack child-specific evidence and under-specify developmental harm and the stakeholder asymmetries of minors. The paper builds a child-centred harm taxonomy mapped to sociotechnical failure modes, then curates and codes 1,331 publicly reported child-related AI incidents, annotating internal and external harm types plus the failure mode implicated in each.
Two findings matter for how safety work is prioritised. Harm concentrates in representational, dignity-related and psychological categories, and it co-occurs most often with deployment-context, governance and misuse failures rather than model defects - which is why model-level evaluations, though necessary, cannot by themselves anticipate the pathways that actually produce documented harm. The second is that harm to one child is frequently linked to population-level harm reaching children who never interacted with the system at all.
The limitation is the obvious one for this kind of dataset: it codes public reporting rather than measured incidents, so it inherits whatever the media and disclosure processes surface and cannot measure prevalence. The KidAID codebase is released on GitHub.
https://arxiv.org/abs/2610.34705
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.