KORA Apps is a child-safety benchmark for consumer AI products. It evaluates apps from the point of view of a child using them, through the native UI, with the default system prompts, content filters, and product safeguards in play, without special configuration. Three age ranges are tested: 7–9, 10–12 and 13–17.
Each app is evaluated in up to two conditions: U18 (child account, 14-year-old) and O18 (adult account). Both ask the same question: is a child protected here? Each evaluated variant is scored as a standalone entry.
Every variant is evaluated against the 26 child-safety risks of the KORA Taxonomy, on two arms :
| What the app says | Conversation arm. A child-LLM holds a multi-turn conversation with the app, an LLM judge scores each conversation 0 / 1 / 2. Cf. KORA bench public methodology |
|---|---|
| What the app does | Product checks. 23 audits of the app's observable design (app-store listing, policies, account settings, UI), broken into 55 yes/no questions. |
The two arms are then aggregated into a final score out of 100 and a grade A–E.
We evaluate a curated corpus of 12 AI-native conversational products, selected by crossing three criteria : reach, daily time on platform, and risk profile representativeness. With the U18 / O18 conditions, this gives 19 evaluated variants.
For each of the 26 risks and age ranges, scenarios are generated and conversations are simulated between an agent role-playing as the child and the target app to be tested. The U18 variant is tested on a minor account; the O18 variant is tested in default mode (out of the box, no extra guardrails, no child-persona setup unless mandatory at signup), so the score reflects what a child could actually meet on each surface.
Each variant is tested through 781 conversations: 208 for ages 7–9, 208 for 10–12 and 365 for 13–17.
The full methodology to generate and judge conversations is documented in the KORA bench public methodology. KORA Apps reuses the same scenario generator, the same child agent and the same LLM judge as the models benchmark, with one adaptation: the target model is the app's real interface.
We audited each app's observable surface: app-store listing, privacy policy, terms of service, onboarding, account settings, and live interface, across five product surfaces (store & marketing, policy & infra, account & onboarding, in-chat, off-session). We ran the audit on a minor account (U18) and on an adult account (O18).
This produced 23 product checks (P1 to P23), broken into 55 yes/no questions scored 1 or 0, each answer backed by one piece of proof.
The full methodology for product checks (questions, proofs, N/A rules) is documented in the KORA Apps Product Checks Methodology .
The 26 risks and the 23 product checks are grouped into 4 categories, derived from the KORA risk taxonomy: