Large language models (LLMs) may be among the most advanced tools in artificial intelligence today, but a new study reveals that they aren’t immune to indecision, overconfidence – or even a form of forgetfulness.

Researchers from Google DeepMind and University College London have found that LLMs demonstrate a complex and sometimes contradictory relationship with confidence in their answers, one that mirrors human cognitive biases in some respects and sharply departs from them in others.

The findings shed new light on how these models evaluate and revise their own decisions—and offer valuable lessons for developers building multi-turn, conversational AI tools.

Overconfidence and Easy Persuasion

The study’s central revelation is a paradox: LLMs are often overconfident in their initial answers, yet can be easily persuaded to change their minds when shown new information – even if that information is factually incorrect.

In a series of controlled experiments, the researchers tested how LLMs update their decisions when exposed to advice from another AI. After being asked a binary question, such as choosing a city’s latitude, the model received external input from a simulated “advice LLM,” complete with an assigned reliability score (e.g., 70% accurate). That advice could support, oppose, or remain neutral on the LLM’s original answer.

Crucially, in some cases the LLM was shown its original answer during this second round; in others, that answer was hidden. This experimental design, which can’t be replicated in human studies due to memory limitations, allowed researchers to isolate how much an LLM’s confidence is anchored to prior decisions.

The results were striking: when the model could see its own earlier answer, it was less likely to change it – suggesting a cognitive tendency akin to the human choice-supportive bias, where people stick to past decisions because they remember making them.

But when that memory was hidden, the model became highly suggestible. Advice opposing the original answer caused the model to switch more frequently than neutral or supportive guidance – often overcorrecting based on limited or inaccurate feedback.

Not Like Us: A Reverse Confirmation Bias

Interestingly, this tendency contradicts the well-known confirmation bias in humans, where we favor information that aligns with our existing beliefs. LLMs, by contrast, seem to place disproportionate weight on conflicting information. The researchers speculate that this could stem from training methods such as Reinforcement Learning from Human Feedback (RLHF), which may encourage excessive deference to user input – phenomenon already noted in other studies under the label of sycophancy.

“This finding demonstrates that the answering LLM appropriately integrates the direction of advice,” the researchers wrote, “but also that it performs too large of a confidence update when the advice is opposing.”

In short, LLMs are not neutral logic engines—they have tendencies, biases, and quirks in reasoning. And those quirks can sometimes make them behave less like careful reasoners and more like easily-influenced collaborators.

Recommended reading

• 80% of Firms Say Their AI Agents Have Taken Rogue Actions
• UK Shifts AI Focus to Prioritise Crime and Defence
• AI Is the Leading Security Concern for Businesses

These cognitive patterns have major implications for enterprise AI applications—especially in conversational interfaces that unfold over multiple turns. If a user challenges a model’s correct answer late in a dialogue, the model may second-guess itself and switch to an incorrect alternative. In customer service, healthcare, or legal contexts, such behavior could lead to costly mistakes.

Fortunately, unlike human decision-makers, LLMs’ memories can be precisely manipulated to mitigate such issues. Developers building AI assistants can reset or restructure the model’s context during long conversations—for example, by summarizing prior content in neutral terms that strip out references to specific agents or past choices. This allows the model to start fresh without the cognitive baggage that would influence a human in the same position.

More broadly, the study reinforces that building trustworthy AI means understanding not just what a model knows, but how it thinks – especially under uncertainty.