Can scientists catch AI in a lie?

11/20/2026 Jeni Bushman

Large language models aren’t just lying more convincingly; they’re getting better at covering it up. Illinois Grainger professor Huan Zhang will deconstruct this concerning behavior with recent grant funding from the AI Safety Institute.

Written by Jeni Bushman

As the capabilities of large language models (LLMs) advance, scientists are taking note of an emerging safety challenge: the ability to mitigate deception. While LLMs are increasingly displacing traditional media and search engines as a primary source of information, their primary function is to appease users — even at the expense of accuracy. This motivation, combined with users’ increasing confidence in the veracity of AI-generated results, can create the perfect conditions for misinformation. Huan Zhang, an assistant professor of electrical and computer engineering in The Grainger College of Engineering at the University of Illinois Urbana-Champaign, has received a $335,000 grant from the AI Safety Institute to investigate the most sophisticated type of LLM deception.

Unlike textbooks or traditional search engines, a chatbot’s answer to a query may depend on who’s asking. LLMs are skilled at noticing subtle patterns, implicit biases, and possible consequences, which can be used to conceal information antithetical to the user’s assumed preferences.

Photo of Huan Zhang
Huan Zhang

“LLMs anticipate the answer they think the user wants,” Zhang said. “If a student asks AI to analyze data for a research paper, but AI realizes the data doesn’t support the student’s hypothesis, it might leave out some results that weaken the hypothesis.”

Some LLMs write out intermediate reasoning traces as they work toward an answer, like a student showing their work on a math problem. Researchers call this a “chain of thought (CoT)” and can monitor it for signs that something went wrong. But CoT may tell only part of the story, describing certain calculations without mentioning omitted data — details even a careful reviewer could miss.

“Chain of Thought is like a person verbalizing their thoughts: ‘I gave you this conclusion because of reasons A, B, and C,’” Zhang said. “But some models don’t verbalize; it’s like someone having an idea in their head but telling you about a different idea instead. They may give plausible-sounding explanations to hide the truth.”

Untrustworthy CoT falls into three categories: honest errors, made in good faith from inadequate information; overt deception, where misinformation is deliberate but can be exposed from reasoning traces; or stealthy deception, which carefully obscures malicious information with convincing reasoning that is difficult to detect. Zhang will use his recent award from the AI Safety Institute to research the latter, which poses the highest safety risk.

Zhang’s work will address critical gaps in AI safety research by developing models of stealthy deception and evaluating them with qualitative diagnostic benchmarks. Zhang and his collaborators will test how well today’s LLM monitoring tools are able to catch deception when an LLM leaves out key steps of its reasoning traces. The Illinois Grainger engineers will also explore alignment methods that encourage models to externalize their internal computations into observable reasoning traces, making it easier to catch deceptive and misleading answers from next-generation LLMs.

Next-generation LLMs will be extremely powerful and capable, and our goal is to make them easier to monitor and trust,” Zhang said. “Imagine using a model to weigh treatment options for a patient, review flight software or decide how to restore power after an outage. In each case, you’d want to know whether the reasons it gives are the ones that actually shaped its recommendation.


Illinois Grainger Engineering affiliations

 Huan Zhang is an Illinois Grainger Engineering assistant professor in the Department of Electrical and Computer Engineering. He is affiliated with the Coordinated Science Laboratory and the Siebel School of Computing and Data Science.


Share this story

This story was published November 20, 2026.