Circuit Breaker Labs Builds "Crash-Test Dummies" to Make AI Safer

Much of the public debate about AI safety focuses on distant, dramatic scenarios in which machines turn on humanity. Meanwhile, a quieter problem is already here. Chatbots have been linked to real psychological harm for some users, including young people who form intense attachments to conversational systems or who receive responses that make a bad moment worse.

A startup called Circuit Breaker Labs says it wants to deal with that present-day risk. Its approach borrows from the car industry: build crash-test dummies, but for AI. Instead of waiting for a vulnerable teenager or distressed adult to stumble into a failure, the company creates simulated users that behave like people in difficult situations and puts them in conversation with a chatbot. The goal is to see where the model breaks down, whether by encouraging unhealthy behavior, mishandling signs of crisis, or blurring the line between software and a trusted friend.

The analogy is a useful one. Automakers do not learn how a vehicle protects occupants by watching real accidents. They stage controlled collisions with instrumented dummies, measure the results, and fix weaknesses before the car goes on sale. Applying that logic to language models means testing at scale and in a repeatable way, rather than relying on a handful of manual red-team exercises or on complaints after something goes wrong.

The timing matters. AI companions and general-purpose assistants are increasingly used by children and teenagers, often without much oversight. Companies that deploy these tools face growing pressure from parents, regulators and lawmakers to show that their products have been tested for more than accuracy and speed. A third-party tool that can generate evidence of how a model behaves with at-risk users could give developers a practical way to meet that demand.

There are open questions. Simulated users are only as good as the behavior they model, and real people are unpredictable. A system that passes a battery of synthetic tests may still fail in a conversation nobody anticipated. Safety testing of this kind also needs clear standards for what counts as a pass, and the industry has not yet agreed on them.

Still, the idea reflects a shift in how some in the field think about risk. Rather than treating safety as a filter bolted on at the end, Circuit Breaker Labs is arguing for something closer to engineering discipline: measure the failure modes, publish the results, and improve. If AI is going to sit in the lives of children and adults alike, supporters say, it should face the same kind of scrutiny as any other product that can cause harm.