Themis AI Teaches Smarter AI: MIT Spinout Helps Machines Know When to Say “I Don’t Know”

In an age where artificial intelligence is making life-changing decisions—from medical diagnoses to infrastructure planning—there’s a growing need for machines that know when they’re out of their depth. One MIT spinout is tackling that problem head-on by teaching AI systems a rare but vital skill: admitting uncertainty.

Founded in 2021, Themis AI is the brainchild of MIT Professor Daniela Rus and her former research collaborators, Alexander Amini and Elaheh Ahmadi. Their startup is behind Capsa, a platform designed to integrate with existing AI systems and help them recognize when they’re veering into guesswork. It's like giving AI a built-in gut check—something most current systems sorely lack.

AI “hallucinations”—a term for confident but incorrect outputs—have become more dangerous as models are entrusted with complex and sensitive tasks. Whether it’s suggesting a faulty drug candidate or misidentifying a pedestrian in autonomous driving software, the risks are real. Themis AI’s solution? Train models to flag uncertainty before a small error becomes a major problem.

Capsa works by identifying subtle patterns that signal when an AI model is confused, biased, or dealing with incomplete data. This meta-awareness allows AI to defer or ask for human intervention when it senses it’s in over its head. Think of it as a machine version of saying, “I’m not sure about this—better double-check.”

The technology builds on years of foundational work at MIT. Back in 2018, with funding from Toyota, the team was researching how to build trustworthy AI for autonomous vehicles. Their work led to breakthroughs in spotting and correcting bias in facial recognition systems. Later, they demonstrated how their approach could reshape drug discovery by helping AIs indicate which results were based on solid evidence versus flimsy correlations.

Since its launch, Themis AI has already proven its worth in the real world. Telecommunications companies have used the platform to avoid costly planning mistakes. In the oil and gas industry, it’s been applied to improve seismic data interpretation. And in AI-driven customer service, it has helped reduce the frequency of chatbots confidently stating falsehoods.

Capsa also offers a major edge for devices with limited computing power. Smaller models that run on local hardware typically struggle with accuracy. By recognizing when they need help, these edge devices can operate more independently without compromising reliability.