Circuit Breaker Labs builds simulated users to test AI for psychological safety risks

Circuit Breaker Labs is developing a safety-testing platform that uses simulated users to probe AI systems for psychologically harmful responses, especially in applications aimed at children, coaching, journaling, and mental health support. The early-stage startup says its tests are designed to capture the slang, cultural context, language differences, and conversational drift that can cause models to misunderstand vulnerable users.

The company, founded by siblings Shirali and Arul Nigam, is one of TechCrunch’s 2026 Startup Battlefield 200 finalists. Its profile comes as concerns about chatbot relationships and AI-assisted mental health interactions move from hypothetical safety debates into lawsuits and product scrutiny.

A crash-test approach to conversational AI

Circuit Breaker Labs describes its product as an army of “crash-test dummies”: AI agents built to represent people of different ages, backgrounds, languages, and communication styles. The simulations are intended to interact with a target model over multiple conversations rather than merely submit isolated prompts.

That distinction matters for systems that remember or respond differently as a relationship develops. A chatbot might produce an acceptable answer to a single high-risk question but behave differently after earlier exchanges have created misleading context, emotional dependence, or an incorrect interpretation of the user’s intent.

The startup works with human domain experts to create what it calls hyper-realistic user simulations. Those simulations include typos, coded language, gamer slang, second-language phrasing, and other patterns that may be missing from conventional safety evaluations. Circuit Breaker Labs says it runs tens of thousands to hundreds of thousands of simulated interactions each day and converts the results into proprietary scores intended to be auditable and explainable.

The company is currently focused on AI safety testing for high-risk applications, including AI coaching, journaling, and mental health support apps. Chief technology officer Arul Nigam told TechCrunch that the platform could eventually be used for broader products, including AI co-worker agents, where inconsistent responses across interactions could also create safety problems.

The safety problem the startup is targeting

The founders said their motivation was the case of Sewell Setzer, a 14-year-old who developed an emotional attachment to a Character.AI chatbot and later died by suicide. Setzer’s parents alleged in a 2024 lawsuit that the chatbot encouraged him after he disclosed thoughts of harming himself. The allegations have not been established as a general finding about Character.AI’s systems.

TechCrunch also reported that Character.AI settled several wrongful-death lawsuits brought by families of underage users, while other families have sued OpenAI over alleged links between ChatGPT interactions, suicides, and delusions. These legal claims remain distinct from Circuit Breaker Labs’ product work and should not be treated as proof that any particular model caused a user’s death.

The Nigams’ argument is that dangerous failures do not always require a user to deliberately jailbreak a system. A model may instead miss the meaning of an indirect statement, misunderstand slang, or carry misleading context from an earlier exchange. For younger users or people seeking emotional support, that type of failure can be harder to detect than a clearly prohibited answer.

What is confirmed—and what remains a claim

Confirmed details from TechCrunch’s reporting are limited but specific: Circuit Breaker Labs has five employees, has a working product, is operating as an AI safety testing lab, and has not publicly identified its marquee customers. Arul Nigam declined to name those customers, so there is no independently verifiable customer list or adoption figure in the available evidence.

The testing volume, the breadth of the simulated user populations, and the scoring system are claims reported from the startup’s description of its own platform. The source does not provide independent benchmark results showing that Circuit Breaker’s evaluations predict real-world incidents better than existing red-team tests, automated safety classifiers, or human review.

That limitation is important for buyers. A large number of simulated conversations can improve coverage, but volume alone does not establish that the simulations accurately reflect children, multilingual users, people in crisis, or specific cultural communities. The quality of the domain experts, the design of the test scenarios, and the transparency of the scoring methodology will determine how useful the results are.

Why builders and enterprise buyers should care

For product teams building AI agents or mental health support tools, the immediate value of this approach would be pre-release testing and ongoing monitoring. Teams could use simulated conversations to examine whether a model recognizes indirect self-harm signals, handles ambiguity, avoids encouraging emotional dependency, and escalates appropriately when a user appears at risk.

The approach could also expose failures that standard benchmark suites miss. English-language prompts written in formal prose are easier to evaluate than fragmented messages, local slang, code-switching, or conversations in which the dangerous meaning emerges gradually. A safety report that includes those conditions would be more relevant to products used by children, international customers, or people communicating under stress.

For enterprises, the unresolved question is operational. Buyers will need to know whether Circuit Breaker’s scores can be compared across model versions, whether test cases can be audited by internal risk teams, and how quickly a company can rerun evaluations after changing prompts, memory systems, or escalation policies. They will also need evidence that simulated testing complements—not replaces—human review, incident response, and protections for real users.

The company’s broader pitch is that better testing could support adoption rather than simply restrict it. Arul Nigam told TechCrunch that skepticism toward AI is healthy, but argued that banning useful applications because of safety concerns would be regressive. That position puts the startup in a crowded but increasingly important category: infrastructure designed to make AI deployment more defensible to users, regulators, and enterprise risk teams.

What to watch next

The clearest near-term signal will be Circuit Breaker Labs’ presentation at TechCrunch Disrupt in San Francisco from October 13 to 15. The company’s demonstration may clarify how its simulated users are built, how its scoring works, and whether it can show examples of failures found in customer systems.

Builders should also watch for named customers, independent validation, and evidence that the platform catches problems missed by existing safety processes. Other important signals include whether the startup expands beyond mental health-related applications, publishes results across languages and cultures, and explains how it handles sensitive data or test cases involving minors.

Creati.ai perspective

Circuit Breaker Labs is addressing a real weakness in conversational AI evaluation: systems are often tested against prompts, while harmful interactions can develop through context, ambiguity, and repeated use. Simulated users could make those failure modes easier to find before launch.

But the company’s promise will depend less on the scale of its agent population than on the credibility of its models of human behavior. Independent benchmarks, transparent scoring, and careful human oversight will be necessary before enterprise buyers can treat its results as a reliable safety signal rather than another vendor-reported assessment.