AI hallucinations pose increasing risks as models gain trust in delivering information and making critical decisions. We all know that one friend who insists they know everything, even when they don’t, offering unreliable advice based on questionable sources.
In this context, AI hallucinations are especially concerning, particularly if such a system were responsible for creating a cancer treatment plan. Themis AI is stepping in to address this issue.
This MIT spinout has made strides in training AI systems to acknowledge when they are uncertain. Their platform, called Capsa, serves as a reality check for AI, enabling models to identify when they are making guesses rather than providing reliable information.
Founded in 2021 by MIT Professor Daniela Rus, along with her former colleagues Alexander Amini and Elaheh Ahmadi, Themis AI’s Capsa platform can integrate with various AI systems to flag uncertain moments before they lead to errors. It essentially teaches AI to discern patterns in its data processing that may indicate confusion, bias, or incomplete information that could result in hallucinations.
Since its inception, Themis has claimed success in helping telecom companies avoid costly network planning errors and assisting oil and gas firms in interpreting complex seismic data. They’ve also contributed to research on developing chatbots that do not fabricate information confidently.
The journey toward tackling AI hallucinations commenced in Professor Rus’s MIT lab, where the team aimed to make machines aware of their limitations. Their work caught the attention of Toyota in 2018, leading to research on reliable AI for self-driving vehicles, where accuracy is crucial for safety.
By recognizing its own biases and limitations, AI can not only enhance industries like pharmaceuticals but also improve the capability of edge devices that traditionally lack the computing power of larger models. As AI systems play increasingly vital roles in our infrastructure and decision-making processes, cultivating the ability to acknowledge uncertainty is paramount.
Themis AI is helping ensure that this critical skill becomes ingrained in the technology.