MIT Startup Tackles AI Hallucinations by Teaching Systems to Admit Uncertainty

The risks associated with AI hallucinations are escalating as these models are increasingly relied upon to surface critical information and make high-stakes decisions.

We all know someone who acts like a know-it-all, refusing to admit ignorance or offering questionable advice based on something skimmed online. AI hallucinations are like that friend—but in this case, the friend could be designing your cancer treatment.

This is where Themis AI steps in. Spun out of MIT, the company has achieved something that sounds simple in concept but is technically challenging: teaching AI systems to recognize and admit uncertainty.

AI models tend to be overconfident. Themis’ Capsa platform provides a reality check, helping models identify when they are speculating instead of operating with certainty.

Founded in 2021 by MIT Professor Daniela Rus and former research associates Alexander Amini and Elaheh Ahmadi, Themis AI built a platform that integrates with nearly any AI system, flagging moments of uncertainty before they result in errors.

Using Capsa, AI learns to recognize patterns in its own data processing that suggest confusion, bias, or gaps in information—conditions that often lead to hallucinations.

Since its launch, Themis reports helping telecom firms prevent expensive network planning mistakes, aiding oil and gas companies in interpreting complex seismic data, and publishing research on building chatbots that avoid confidently inventing facts.

Many people are still unaware of how often AI systems are essentially making educated guesses. As these systems take on more vital roles, such guesses can carry significant consequences. Themis AI introduces a missing element: self-awareness.

Themis AI’s path to addressing AI hallucinations

The journey to Themis AI began years ago in Professor Rus's lab at MIT, where researchers explored a core question: How can a machine become aware of its own limitations?

In 2018, Toyota funded their work on reliable AI for autonomous vehicles—an industry where errors can be life-threatening. The challenge is especially critical when self-driving cars must precisely detect pedestrians and road hazards.

Their breakthrough came with an algorithm capable of identifying racial and gender bias in facial recognition systems. Instead of just detecting bias, their system corrected it by rebalancing training data—effectively teaching the AI to overcome its own prejudices.

By 2021, the team demonstrated how this method could transform drug discovery. AI systems could assess potential drugs while highlighting when predictions were grounded in solid data versus speculation or total hallucinations. Pharma companies saw the value in pursuing only drug candidates the AI was confident about, saving both time and resources.

Another benefit applies to devices with limited computing power. Edge devices often rely on smaller models that can't match the accuracy of server-based systems. Themis’ technology helps these local models handle most tasks independently, seeking server support only when they encounter something difficult.

AI offers immense potential to improve our lives, but that comes with real risks. As AI becomes embedded in critical infrastructure and decision-making, the ability to recognize uncertainty—and avoid hallucinations—could prove to be its most human and valuable trait. Themis AI is helping models learn this essential skill.

See also: Diabetes management: IBM and Roche use AI to forecast blood sugar levels

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is co-located with other leading events including Intelligent Automation Conference, BlockX, Digital Transformation Week, and Cyber Security & Cloud Expo.

Explore other upcoming enterprise technology events and webinars powered by TechForge here.