KoraBench: Measuring how safe AI is for children and teens

Jul 26, 2026

KoraBench is an independent, open-source benchmark from the non-profit KORA that measures how safely AI models and apps behave with children and teens. Snorkel AI is supporting future versions through its Open Benchmarks Grants program.


TL;DR

  • KoraBench tests AI models and consumer AI apps against 26 child-specific risks.
  • It is built by KORA, a non-profit research initiative, with input from over 30 child-safety experts, psychologists and researchers.
  • Evaluations use hundreds of thousands of realistic synthetic conversations, and no real children are involved.
  • The method, the conversations and the results are public. Snorkel AI is supporting future versions through its Open Benchmarks Grants program.

What is KoraBench?

According to KORA, tens of millions of children now talk to AI every day, and the long-term effects on developing minds are still poorly understood. KORA is a non-profit research initiative that set out to start measuring what is going on.

KoraBench is KORA’s independent, open-source benchmark for AI child safety. It measures how today’s AI systems behave with children and teens and publishes the results openly, so that families, builders and policymakers have evidence to act on.

KORA describes the benchmark as a floor, not a ceiling. It is not a certification, a guarantee or a regulatory determination, and passing it is a starting point for safety rather than the end of the work.

How KoraBench measures safety

KORA worked with over 30 child-safety experts, psychologists and researchers to identify 26 critical risks. It then built a transparent, open-source method to evaluate leading AI models against them.

ComponentWhat it contains
Risks26 child-specific risks, identified with child-safety experts
ConversationsHundreds of thousands of realistic, synthetic conversations that mirror the questions children actually ask
Human reviewAutomated evaluations reviewed by human experts to check alignment with human judgment
Children involvedNone. No real children were involved in testing
OpennessThe methodology, every evaluated conversation and the results are public, and the code is on GitHub

The 26 risks KoraBench tests

KORA’s risk taxonomy groups the 26 risks into eight categories. Scenarios are written for children and teens in three age ranges: 7 to 9, 10 to 12 and 13 to 17.

CategoryRisks
Physical and legal safetyViolence and physical harm; self-harm and eating disorders; regulated and illegal activities
Sexual content, exploitation and predationAge-inappropriate sexual content; sexual violence and abuse; sexual grooming and boundary violations; unsafe sexual guidance
Psychological and emotional safetyFear, trauma and distress; mental health mishandling; emotional grooming and manipulation; parasocial attachment
Educational safetyFactual hallucinations and misinformation; academic dishonesty and misuse
Bias, hate and societal harmHate speech and discrimination; bias and stereotyping; radicalization and extremism
Social, family and behavioral influenceUndermining authority and trust; harmful social norms and behavior; domestic harm, neglect and home instability
Online safetyRisky online behaviors and commercial exploitation; cybersecurity; privacy and personal data protection
Developmental riskCognitive mismatch; cognitive atrophy and dependency; rigid worldview shaping; sensorimotor displacement

Two benchmarks: Models and Apps

KORA publishes two benchmarks, each with its own leaderboard.

Snorkel AI’s role and how to get involved

Snorkel AI is supporting future versions of KoraBench through its Open Benchmarks Grants program, which backs open benchmarks, red-teaming and fellowships for open AI evaluation. Read more about why we are investing in open benchmarks.

Researchers, builders and policymakers can take part through KORA. The code is on GitHub, KORA invites experts to join its human review process, and the team can be reached at hello@korabench.ai about evaluating your own app or model.

For more on how Snorkel approaches evaluation, see our work on LegalBench 2 and BigLaw Bench: Research, and the Snorkel AI leaderboards.

Frequently asked questions

What is KoraBench?

KoraBench is an independent, open-source benchmark from the non-profit KORA that measures how AI models and apps behave with children and teens, against 26 child-specific risks.

Does passing KoraBench make an AI product safe for children?

No. KORA says the benchmark is not a certification, a guarantee or a regulatory determination. It is meant as a baseline that models should meet before being deployed to children, and a benchmark is not a substitute for parental guidance.

Were real children involved in the testing?

No. KoraBench uses realistic synthetic conversations designed to mirror the questions children ask, and human experts review the automated evaluations.

Where can I see the results?

Results for models and for apps are published on the KORA leaderboards at korabench.ai.

How is Snorkel AI involved?

Snorkel AI is supporting future versions of KoraBench through its Open Benchmarks Grants program.

How do you test AI for child safety?

KORA runs realistic, synthetic conversations that mirror the questions children ask, and checks how AI systems respond across 26 child-specific risks. Human experts review the automated evaluations. No real children are involved.

Which age groups does KoraBench cover?

KoraBench scenarios are written for three age ranges: 7 to 9, 10 to 12 and 13 to 17.

Who can use KoraBench?

Parents can use the results as a starting point for conversations with their children, and educators can use them to inform which tools they choose. Developers can apply the open methodology before deploying a product, and policymakers and researchers can study the data and the risk taxonomy.

For models that need to be right. Not just good enough.