Circuit Breaker Labs Builds 'Crash-Test Dummies' to Catch AI Harm Before It Reaches Kids
The startup is testing AI systems against simulated vulnerable users to surface psychological harm that existing safety benchmarks miss.
Circuit Breaker Labs has built a testing method it calls "crash-test dummies" to measure how AI systems affect people psychologically, with a particular focus on children. The company's approach targets a gap in current safety work: most evaluations probe for catastrophic risks, while harm that has already occurred — emotional distress, manipulation, dependency — goes largely unmeasured. The news, first reported by TechCrunch, puts a name to a problem the industry has treated as secondary.
The premise behind the tool is that AI products are already in children's hands, and the damage is not hypothetical. Chatbots have been linked to disturbing interactions with minors, and researchers have documented cases where users formed unhealthy attachments to companion apps. Circuit Breaker Labs argues that these outcomes deserve the same structured testing that safety teams apply to model capabilities, rather than being handled case by case after complaints surface.
The "crash-test dummy" framing is deliberate. Instead of asking whether a model can produce a harmful sentence in isolation, the company simulates realistic user profiles — including minors and people in distress — and records how systems respond across extended conversations. That method aims to catch patterns that single-turn red-teaming tends to miss, such as gradual escalation, flattery, or a model that keeps engaging when it should disengage.
The startup enters a crowded but uneven field. Major labs publish system cards and run internal red teams, and independent groups have pushed for standardized evaluations, yet no shared benchmark for psychological safety has taken hold. Regulators are moving in parallel: child-safety laws in the US and the UK's online safety regime now ask platforms to assess foreseeable harm to minors, which gives outside testing firms a potential market beyond voluntary adoption.
What Circuit Breaker Labs has not yet shown publicly is scale. The company has not disclosed how many models it has evaluated, which developers use its results, or whether its findings have changed a shipped product. Those details matter, because safety testing only counts if it alters behavior before release rather than producing a report afterward.
The open question is whether psychological harm becomes a standard metric in AI evaluation or stays a niche concern. If developers adopt something like crash-test dummies the way carmakers adopted physical ones, the bar for releasing consumer AI shifts. Until then, the burden falls on parents and users to notice damage that the industry still has few tools to predict.