What did 1,331 reported child-AI incidents reveal? Most coded harms involved dignity or psychology—not just model errors
A new unreviewed corpus found reported harms were most often coded alongside deployment, governance and misuse failures. Reporting bias, related incidents and only moderate agreement on failure modes mean the counts are not population rates.
The short version
- What happened
- Researchers manually coded 1,331 child-related records drawn from the OECD AI incident monitor, AIAAIC and the AI Incident Database.
- Who it affects
- The findings direct attention to how child-facing products are designed, distributed and governed, but they do not establish an individual child's risk from any specific service.
- What remains unproven
- Unreviewed preprint
Keep in mind
Unreviewed preprint
Public-reporting and English-language selection bias
Related incidents may not be independent
No child accounts or geographic coding
Only one primary failure mode per record
Moderate coder agreement for failure modes and worst-case flags
Source analysis · 1 linked source. Primary analysed source date: 26 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The study moves the safety question beyond model outputs
An unreviewed US preprint has assembled 1,331 publicly reported incidents in which a person under 18 plausibly experienced harm, or a credible worst-case risk of harm, involving AI. The researchers drew 1,143 records from the OECD AI incident monitor, 141 from AIAAIC and 47 from the AI Incident Database. They excluded speculative claims, cases without plausible child involvement and identified duplicates, then manually assigned one individual harm, one wider harm and one primary failure mode to each record.
The resulting picture is broader than a benchmark of unsafe answers. Representation, dignity and social-identity harms were coded in 475 incidents, including 316 involving sexualization and 131 sexual exploitation. Psychological harms appeared in 409 records, including compulsive use, self-harm, health deterioration and toxic-content exposure. These are counts of records in repositories—not estimates of how many children are harmed, the likelihood of harm or the performance of any particular product.[1]
Deployment and misuse dominated the coding
Human, organizational and contextual deployment failures were the largest family, assigned to 779 incidents. Within that group, contextual-norm violations accounted for 398, governance or regulatory noncompliance for 227 and overreliance or misleading anthropomorphic design for 146. Misuse or abuse was the primary family in 297 records, including 228 coded as malicious use. Model behaviour or technical failure was assigned to 196, led by general reliability problems rather than hallucination.
That distinction matters for schools, parents and regulators. A content filter cannot stop every non-consensual synthetic image, cumulative companion dependency or severe response to a false school-security alert. The paper therefore argues for controls around depiction consent, age assurance, session-level escalation, guardian visibility, distribution and institutional response. Those recommendations are partly informed by governance literature: the dataset does not code provider or deployer identity, so it cannot measure which organizations or interventions account for the patterns.[1]
The distinction that matters
How to read the 1,331 records
The study describes a selected public-reporting corpus rather than a representative survey of children.
Source pool
Combine three public incident repositories
1,331 included records
Manual coding
Assign harm and one primary failure mode
Moderate-to-acceptable agreement
Valid inference
Count patterns inside the corpus
No population prevalence or causality
The corpus surfaces blind spots, not a child-harm rate
Two researchers independently coded a random 150-record subset before resolving disagreements. Agreement was 0.73 for harm categories, 0.67 for failure modes and 0.58 for worst-case flags. The authors call the failure-mode results provisional because 0.67 falls below the conventional threshold for firm conclusions. In 112 records, reporting did not support an individual-harm category. Each record received only one primary failure mode even when several factors contributed.
The repositories rely heavily on English-language journalism, litigation and regulatory reporting. Acute and legally visible events are more likely to appear than private, cumulative harm, and one platform can generate multiple related records. Children’s own accounts are absent. A stronger next step would link related events, add geography, provider and stakeholder fields, include multilingual reporting and test high-severity pathways with better causal designs. Until then, the defensible finding is that documented child-AI harm often sits in product and deployment systems—not that this corpus measures its true scale.[1]