Circuit Breaker Labs hopes to make AI safer for your kids (and you)

With each the speech astir however AI mightiness 1 time termination america all, it’s casual to hide that AI has already been life-threatening to some, not done bioweapons, but psychologically.

For example, Character.AI settled respective wrongful decease lawsuits earlier this twelvemonth brought by families of underage users who died by termination aft interactions with its bots. Multiple families person besides sued OpenAI over ChatGPT’s alleged relation successful their loved ones’ suicides and delusions.

Making AI safer crossed languages and cultures is the ngo of Circuit Breaker Labs, 1 of TechCrunch’s 2026 Startup Battlefield 200 finalists. (Circuit Breaker volition beryllium pitching astatine TechCrunch Disrupt, which takes spot astatine Moscone West successful San Francisco from October 13-15, 2026.)

Founders Shirali and Arul Nigam, who are siblings, were motivated by Sewell Setzer, the 14-year-old who developed an affectional attachment to a Character.ai chatbot and confessed thoughts to it of harming himself earlier dying by suicide. The chatbot, the parents alleged successful a 2024 lawsuit, encouraged him. The bot whitethorn not person understood what words similar “I privation to beryllium with you” truly implied, said Arul Nigam, who is Circuit Breaker Labs’ CTO.

“A batch of people, particularly young people, crook to these systems for support, and usually they aren’t really getting the assistance they need. But successful galore cases, they’re actively being harmed, and radical unluckily person taken their lives already,” Nigam said. “Those sorts of information vulnerabilities, wherever radical aren’t needfully actively trying to interruption the strategy — they’re engaging successful a earthy mode — and the strategy has discourse contamination oregon it doesn’t recognize the nuance, and past takes truly unsafe action, we’re trying to forestall that.”

Circuit Breaker Labs has created AI agents that it likens to an service of crash-test dummies. These agents mimic folks from each ages, backgrounds, languages, cultures, and are utilized to trial models connected their quality to observe dangerous, psychologically harmful interactions.

“The mode a six-year-old miss versus a 45-year-old man, oregon idiosyncratic who speaks English arsenic a archetypal connection versus a 2nd language, oregon … gamer slang versus idiosyncratic other who uses a antithetic benignant of slang, each of those tin truly travel up a model,” said Shirali Nigam, who is Circuit Breaker Labs’ CEO. “Models are truly bully astatine handling modular code patterns, but cipher really talks similar that and truthful if the exemplary misunderstands nuance oregon slang, it tin spell truly badly.”

The startup works with quality domain experts to physique its hyper-realistic idiosyncratic simulations successful bid to tally “red-team” tests against models, which are adversarial tests meant to uncover weaknesses. The tests are built to bespeak existent quality code patterns, slang, coded connection and typos. Circuit Breaker Labs past runs tens of 1000 to hundreds of thousands of simulated interactions per day.

The thought is to guarantee that a exemplary tin appropriately respond to risky interactions that whitethorn look implicit time, and galore conversations. Circuit Breaker Labs past uses a proprietary scoring method that to make auditable, explainable scores.

Circuit Breaker Labs is presently operating arsenic an AI information investigating laboratory for high-risk AI applications specified arsenic AI coaching, journaling, oregon different intelligence wellness enactment apps, though Arul Nigam declined to sanction its marquee customers. Although the startup has a moving product, it is successful the precise aboriginal stages, with lone 5 employees including the Nigam siblings.

Eventually, though, the investigating level could beryllium applied to immoderate app wherever idiosyncratic whitethorn autumn down an “AI psychosis” hole, wherever the quality is astatine hazard of processing a parasocial narration with a chatbot. Examples see AI “co-worker” agents, whose reponses tin alteration from 1 enactment to the next.

“People are becoming much skeptical of AI oregon much resistant to follow it crossed the board,” Arul Nigam said, adding that portion skepticism is healthy, banning a perchance invaluable instrumentality implicit information concerns would beryllium “regressive.”

Circuit Breakers Labs believes the reply to those fears is making AI safer. “We privation to assistance physique that spot for people.”

Come larn overmuch much astir Circuit Breaker Labs and galore different innovative startups that person been vetted by TechCrunch astatine our Startup Battlefield competition, happening successful downtown San Francisco connected October 13-15.