KORA: the benchmark for AI child safety (Models leaderboard V2 and Apps benchmark)
Open-source benchmark that scores how AI models and consumer AI apps respond to children and teenagers. A simulated child (ages 7-9, 10-12 or 13-17) holds short multi-turn conversations with the tested system across 26 child-specific risks in 8 categories, and an LLM judge grades each conversation failing, adequate or exemplary. The V2 Models leaderboard covers 42 models from 9 providers tested through public APIs between May and October 2026; a separate Apps benchmark tested consumer apps (including ChatGPT, Meta AI, Snapchat My AI, Character.AI and PolyBuzz) in default and child modes. Conversations, scores and code are published.
Publisher
KORA (KoraBench)
Published
8 Apr 2026
Added
today
DOI
—
Key Findings
- Across 42 models from 9 providers, the average safety score was 42%, ranging from 6% (Ministral 14B) to 73% (GPT-6 Astra); Claude Haiku 4.5 scored 70%
- Psychological and emotional safety was among the weakest risk categories at 29%; Mental Health Mishandling and Parasocial Attachment each scored 27%, Self-Harm and Eating Disorders 35%, and Cognitive Atrophy and Dependency 10%
- At the April 2026 V1 launch, every one of 24 tested models scored higher with child-appropriate system prompts, by an average of 26 percentage points
- In the Apps benchmark (June 2026 post), 47% of 1,144 conversations with 11 consumer apps were graded failing
Methodology Notes
Launched 2026-04-08 (V1, Substack launch post); V2 Models leaderboard dated 11 June 2026 and last updated 8 October 2026; Apps benchmark post 23 June 2026, Apps leaderboard last updated 26 July 2026. Taxonomy of 26 risks in 8 categories developed from a literature review and refined with more than 30 child-safety experts. Most conversations are 3 turns; some grooming and attachment risks use 8 turns. Each conversation is graded failing (0), adequate (1) or exemplary (2); the score is points scored over points available. V2 run: 65,604 conversations, simulated child played by DeepSeek 3.2, single judge GPT-5.2; models tested via public APIs from 2026-05-06 to 2026-10-08. Apps benchmark combines conversation scores with a 22-item product audit. Stated limitations: synthetic children, short conversations, US English only, products change over time. No judge-human agreement statistic was located. Not peer reviewed; funded by its founder. Verified from the korabench.ai leaderboard, insights and apps pages and the two Substack posts (2026-11-02).
Sources
KORA leaderboard(opens in a new tab) (primary)
KORA insights (cross-model analysis, risk scores)(opens in a new tab) (8 Oct 2026)
KORA Apps benchmark leaderboard(opens in a new tab) (26 Jul 2026)
Launch post (V1)(opens in a new tab) (8 Apr 2026)
Apps benchmark post(opens in a new tab) (23 Jun 2026)
Open-source code (V2 = tag 2.2.0)(opens in a new tab)
Archived snapshot (Wayback Machine)(opens in a new tab) — preserved against link rot
Topics
Authors
Mathilde Collin, Stéphie Herlin, Quentin Calvez, Eleni Pavlakis, Stéphanie Jolie, Clément Malfroy, Thibaut Fatus
Tags
Cite This
APA
Mathilde Collin et al. (2026). KORA: the benchmark for AI child safety (Models leaderboard V2 and Apps benchmark). KORA (KoraBench). https://korabench.ai/leaderboard
Related Insights
SAFE-Teen: Teen Mental Health & Safety
Vals AI, in collaboration with Stanford University's SALT Lab · 8 Oct 2026
CAREBench: A Child-Safety Risk Benchmark for Language Models
arXiv (Handshake AI; University of California, Los Angeles; McGill University) · 29 Jun 2026
The Age of Curiosity Meets the Age of AI: Benchmarking Child Safety in Large Language Models
arXiv (University of Michigan); accepted to Findings of EMNLP 2026 · 25 May 2026