The BaseVideoCenter for Humane TechnologyOctober 22, 202635 min

A New Way to Measure AI Safety for Kids

In a 35-minute interview, Imran Khan of the Center for Humane Technology talks with Mathilde Collin, founder of the open-source Kora benchmark of AI models and apps for child safety. Collin reports results, including that telling an AI it is talking to a child raises its safety score.

Open the source ↗

Why it matters

The same thing, for different readers.

What it means for you

Collin says two thirds of US teenagers use AI chatbots, and that the uses children rely on most, emotional support and schoolwork, are where the tested models scored lowest.

For lawmakers

Collin asks AI labs to share usage data and calls for research funding, and the benchmark finds apps rarely verify age, which regulators could target.

For researchers

Kora scores whole multi-turn conversations with child personas using language-model graders reviewed by humans; Collin says the score is not a certification and that the framework is imperfect.

The question it raises

Collin names a paradox: a good teacher aims to be unneeded, while AI products are optimised for the child to return.

In their own words

Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.

From our own transcript of the episode’s audio, made by machine. Speaker names come from what is said, not from a voice match.

“the more an AI will tell a child, go see a human, the safer it is”

Mathilde Collin, Founder of the Kora benchmark; founder of FrontDescribing the two behaviours most correlated with safety in the benchmark: redirecting children to humans, and not agreeing with everything.

Found word for word in the sourceHear it at 10:24 ↗Also in a second, independent transcriptSpeaker known from context, not confirmed by listening

“right now zero out of nineteen apps verify the age of children”

Mathilde Collin, Founder of the Kora benchmark; founder of FrontExplaining why the safety gain from telling an AI it is talking to a child is easy to capture but not being used.

Found word for word in the sourceHear it at 19:12 ↗Also in a second, independent transcriptSpeaker known from context, not confirmed by listening

“we're running this uncontrolled experiment on our children and there is a need to act now and to build now and to educate now”

Mathilde Collin, Founder of the Kora benchmark; founder of FrontCollin on what parents and educators should know, after noting that two thirds of teens use AI regularly.

Found word for word in the sourceHear it at 22:39 ↗Also in a second, independent transcriptSpeaker known from context, not confirmed by listening

How the finding was reached

In our words, from the source. Weigh the finding by how it was made.

What was done: Kora benchmark: 26 risks in 8 categories, simulated multi-turn chats with child personas, graded 0, 1 or 2 by language-model judges reviewed by humans, for leading models and for apps.
What it found: Collin says no tested app earns an A and one earns a B; telling the AI it is chatting with a child adds 24 points of 100; behaviours linked to safety are redirecting children to humans and not agreeing with everything; education and emotional support scored lowest.
What it does not show: Collin says the scores are not a certification label, the framework is imperfect and some risks, such as agency, are hard to measure.
Status: The company’s own report about its own work

The record

Questions: How is AI changing childhood, health and daily life? · Are AI companies putting stronger safeguards into practice?
People: Mathilde Collin (Speaks, Founder of the Kora benchmark; founder of Front) · Imran Khan (Speaks, Researcher and strategist leading humane evals work, Center for Humane Technology)
Also covered by: Your Undivided Attention · Center for Humane Technology
What we read: We read the whole source.
Found by: The Radar, run of October 23, 2026, via yt-center-for-humane-technology (tier 1: the source’s own channel or an official record)
Cite this: Center for Humane Technology, “A New Way to Measure AI Safety for Kids”, October 22, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/center-for-humane-technology-2026-10-22-kora-kids-benchmark

Our notesSpeech-to-text with no speaker labels; the answers are clearly from the guest by context, but speakerCertain is false. The transcript has errors (for example 'Cora', 'Corabench'). Kora's own results were not checked on its site. Status company-report is a rough fit: Kora is described as a nonprofit.