AI Issues: Finally, Solving the AI Hallucination Problem
AI hallucinations are becoming increasingly dangerous as models are trusted to analyze information and make critical decisions.
We all have that know-it-all friend who can't admit what they don't know or gives misguided advice based on something they read online.
Hallucinations by artificial intelligence models are similar to that friend, but these cancerous hallucinations can be prevented.
This is where Themis AI comes in.
This MIT spinout company has succeeded in solving a problem that seems simple in theory but is incredibly complex in practice—teaching AI systems to say, "I'm not sure about this."
AI Tends to Be Overconfident
AI systems generally tend to be overconfident.
Themis's Capsa platform serves as a reality check for AI, helping models recognize when they are relying on guesses rather than certainty.
Founded in 2021 by MIT's Daniela Rus and former research colleagues Alexander Amini and Elahe Ahmadi, Themis AI has developed a platform that can connect to virtually any AI system and flag moments of uncertainty before they lead to mistakes.
Capsa essentially trains AI to find patterns in how it processes information that may indicate confusion, bias, or working with incomplete data that could lead to hallucinations.
Since its launch, Themis has helped a telecom company prevent costly network planning errors, assisted an oil and gas company in understanding complex seismic data, and published research on building chatbots that don't confidently fabricate situations.
Most people don't realize how often AI systems are simply making their best guesses.
As these systems handle increasingly critical tasks, these guesses can have serious consequences.
Themis AI's software adds a layer of self-awareness that didn't exist before.
Themis's Journey to Solving the AI Hallucination Problem
Themis AI's journey began years ago in Professor Rus's MIT lab, where the research team was studying the fundamental problem of making machines recognize their own limitations.
In 2018, Toyota funded research on trustworthy AI for autonomous vehicles, where mistakes could be fatal.
When autonomous vehicles must accurately identify pedestrians and other road hazards, the stakes are enormously high.
The research team found a breakthrough when they developed algorithms that could detect racial and gender bias in facial recognition systems.
This system didn't just identify problems—it actually solved them by rebalancing the training data.
In other words, teaching AI to correct its own biases.
By 2021, the research team demonstrated how this approach could bring new changes to drug development.
AI systems could crucially distinguish between predictions based on solid evidence and those based on educated guesses or outright hallucinations.
The pharmaceutical industry recognized the potential to save costs and time by focusing only on drug candidates that the AI was confident about.
Another advantage of this technology relates to devices with limited computing power.
Edge devices use smaller models that can't match the accuracy of large models running on servers, but with Themis's technology, these devices can handle most tasks locally and only request help from large servers when they encounter difficult problems.
Essential Technology for the Future
AI has enormous potential to improve our lives, but that potential comes with real risks.
As AI systems become more deeply embedded in critical infrastructure and decision-making, the ability to recognize uncertainty may prove to be the most human and most valuable quality.
Themis AI is helping teach this critical skill.
IMPAKERS Blog | Post on the need for AI literacy Read more
Source: Ryan Daws, AI News, "Tackling hallucinations: MIT spinout teaches AI to admit when it's clueless", https://www.artificialintelligence-news.com/news/tackling-hallucinations-mit-spinout-ai-to-admit-when-clueless/, (2025. 6. 3)