Circuit Breaker Labs Aims to Enhance Psychological Safety in AI Models
Founded by siblings, the startup tests psychological safety vulnerabilities and hazardous interactions in artificial intelligence systems using AI agents.
Circuit Breaker Labs, a startup founded by siblings Shirali and Arul Nigam, uses AI agents as crash test dummies to detect psychological safety vulnerabilities in AI models.
AI and Psychological Risks
While debates continue over the threats that artificial intelligence technologies might pose to humanity in the future, it is evident that some systems are already producing psychologically hazardous consequences for users.
Platforms such as Character.AI and OpenAI have faced various lawsuits filed by families linked to the suicides of underage users and negative experiences they endured.
The Purpose of the Startup
A finalist at TechCrunch Startup Battlefield 200, Circuit Breaker Labs aims to make artificial intelligence safer across different languages and cultures.
The startup's founders, siblings Shirali and Arul Nigam, state that they were influenced by the tragic events experienced by young users who lost their lives when adopting this mission.
Agents Operating Like Crash Test Dummies
Circuit Breaker Labs has developed specialized AI agents resembling crash test dummies to test non-traditional conversational styles and nuances.
These agents mimic humans from various age groups, cultural backgrounds, linguistic histories, and slang usage habits to expose the weaknesses of the models.
Large-Scale Simulation Tests
Realistic simulations created with the contributions of human domain experts are used to conduct adversarial and conflicting red-teaming tests against AI models.
The startup's systems run tens of thousands to hundreds of thousands of simulated interactions per day, aiming to ensure that models respond appropriately to risky situations in the long run.
Targeted Application Areas
The company currently operates as a safety testing laboratory for high-risk AI domains such as coaching, journaling, and mental health support applications.
This early-stage platform, which has five employees, could potentially be adapted in the future to all applications carrying the risk of users forming unhealthy parasocial relationships with chatbots.