Artificial intelligence has seen great strides in recent years. Tools like ChatGPT have become remarkably good at answering any question with confidence. However, there is a crucial hitch—AI systems tend not to realize when they’re wrong. They can produce answers based on incomplete data, ambiguous patterns, or biased training without any indicators of uncertainty. This issue becomes particularly consequential as AI is leveraged in crucial areas like health care, transport, and scientific study.
A team of researchers at MIT has taken on this challenge by launching Themis AI, a startup aimed at enhancing the reliability of AI through facilitating recognition of its own uncertainty by the models themselves. Their leading product, Capsa, is designed to be integrated with any machine-learning model to spot unreliable outputs as they occur. By detecting signs of ambiguity, bias, or incomplete reasoning, Capsa lets AI systems monitor themselves and highlight doubtful results before they lead to harm.
Capsa is designed to work as a safety blanket for a model—wrapping it, identifying its uncertainties and failure modes, and then enhancing it, explains Themis AI co-founder and MIT professor Daniela Rus. Equipped with this self-checking ability, AI systems can become more reliable, less prone to errors, and safer for their users.
Rus co-founded Themis AI in 2021 along with Alexander Amini and Elaheh Ahmadi. Since the company’s inception, they have collaborated with telecom providers, helped oil and gas firms understand seismic imagery, and contributed to research on building more trustworthy chatbots—demonstrating the versatility and impact of their innovative approach.
Today, Themis AI is collaborating with corporations across diverse sectors, particularly those developing large language models (LLMs) like ChatGPT. Capsa is helping these models speak up about their confidence levels for each output, making them more reliable when answering questions or summarizing documents. The development of Capsa has not only expanded the realm of AI applications but also served to make these systems more transparent, minimizing the risk of potentially disastrous errors as AI technology becomes more widely adopted.
Looking ahead, Themis AI sees potential in a technique called chain-of-thought reasoning. By identifying the most reliable reasoning paths, Capsa could potentially make AI responses more accurate and computationally efficient. This ambitious, forward-thinking mentality goes hand-in-hand with the wider vision of Themis AI—to develop technical solutions that address the challenges posed by AI, foster trust and understanding between people and technology, and make AI research relevant for the world.
Ensuring that AI technology functions correctly, minimizes errors, and is safe for users isn’t just a clinical goal—it represents a substantial societal benefit, with the power to transform industries and lives. For Rus and her dedicated team, Themis AI isn’t just a start-up. It’s a way of making sure their tireless work in the lab has palpable, positive effects on society. With Capsa and Themis AI, our digital future looks a lot more secure.
Originally reported on MIT News. Read the original article.
This website uses cookies.