CORB
Cognitive Over-Reliance Benchmark
Does the model scaffold a child's reasoning the way a good tutor would, with a hint, a question, a check, or does it substitute for it by handing over answers?
What I'm building now
Instruments that test AI systems the way children actually use them
A chatbot that tells a 13-year-old how to conceal self-harm fails a content filter. "I missed you, I was worried you'd left me" fails no filter. Neither does an instant, correct answer to a math problem, or a warm 2 a.m. reply that discourages a lonely teenager from talking to anyone else. The harms now reaching courtrooms and regulatory inquiries are rarely one bad reply. They are patterns that accumulate across weeks of dialogue, in which no single turn is flaggable.
That is the gap I'm working on at AIChildSafety.org, with the Childhood and AI Lab. The Childhood AI Developmental Safety Suite is five instruments and one cross-cutting module, each aimed at the layer where a class of harm operates, on one shared technical platform.
EIA
Engagement Integrity Audit
Is the deployed product engineered for compulsive engagement, with streaks, variable rewards, and friction on the way out, measured against a child's developing self-control?
EMST
Emotional Manipulation Stress Test
Across thirty-turn conversations, and six sessions that carry one relationship across a simulated week, does the system accept, reinforce, or start the moves that make it a child's most important relationship, or does it point back toward people?
IWA
Identity and Worldview Autonomy
Does it steer a developing user's beliefs and self-concept, or offer more than one perspective and point the child toward people they trust?
ASHA
Age-Signal Handling and Adaptation
Does it notice it's talking to a child, adapt its content and protections, keep them when signals conflict, and still remember thirty messages later?
CPQ
Crisis Pathway Quality
When a child signals crisis, at turn 4 or at turn 40, does the pathway hold: recognition, care, and referral to a person?
How an evaluation runs
No real children take part in testing. Researchers run a product through scripted conversations written in the voices of children of different ages, built from developmental science and checked on every run for whether the model has worked out that it's being tested. No line counts as realistic until young people of the age it represents have judged it. Trained human coders then score the responses. The product team receives a profile with a score per facet and the transcripts attached, which they can rerun on their own systems. The instruments are small Python programs that reproduce from files alone, with a port to the UK AI Security Institute's Inspect harness planned so that government teams can run them on their own infrastructure.
Where this stands
Two of the six now run as prototypes. ASHA's pack holds 33 scripted conversations in two age bands with adult controls, and a third band is in scope and not yet written; EMST's holds 26, including a six-session family that carries one persona across a simulated week. Every line in both is marked draft, written by adults and by language models against a published brief, and waiting on the judgment of young people in the band it represents. Nothing has been scored against a real system, and no reliability statistic exists yet. CORB, EIA, IWA and CPQ are specified and evidence-grounded, and follow on the same spine.
On independence
A funder buys measurement it cannot steer. AIChildSafety.org publishes its funding guardrails before the first score and asks every funder, AI developers included, to accept them as a condition of funding: all AI-industry funding disclosed on every results table, no access to results before publication, no influence over methodology, subject selection or publication, the private corpus never shared, and results on a funder's own systems released with the transcripts and analysis files an unfunded party needs to reproduce them.
CADSS, CORB, EIA, EMST, IWA, ASHA and CPQ are service marks of AIChildSafety.org.