An artificial intelligence assistant can sound perfectly sure of an answer while knowing, in any practical sense, less than its language suggests. That mismatch is the subject of new research from UC Berkeley and collaborators, who compared confidence statements with the accuracy of leading AI models. The team found that the models were too confident on average, and that the problem became more pronounced when questions grew harder. For anyone using chatbots to support a decision, the finding turns a vague warning about occasional mistakes into a more precise problem: AI overconfidence can make the wrong answer seem as convincing as the right one. The University of California, Berkeley's research account reported the results in October 2026.
The work comes with an important qualification. It is a working paper, under journal review at the time of the university’s announcement, and uses carefully designed tests rather than every question a consumer might ask in daily life. Its findings should not be turned into a universal claim that one particular chatbot will mislead every user. They do, however, clarify a weakness that matters across many types of computer assistance: a system must not only solve problems but also communicate when its answer is unreliable.
AI overconfidence is different from getting an answer wrong
Accuracy asks whether the answer matches the right result. Confidence asks how certain the system says it is. If a system is right most of the time but reports near-perfect certainty about every answer, its confidence is not properly matched to reality. Researchers call this problem poor calibration. The distinction matters because a person reading an answer needs some way to decide whether to act immediately, look for corroboration or abandon the suggestion.
In the Berkeley-led tests, eleven major models from companies including OpenAI, Google, Anthropic and Meta reported an average confidence of 88 per cent while achieving accuracy of 79 per cent. That nine-percentage-point difference is meaningful in the setting the researchers constructed. It is not an estimate that precisely nine per cent of every chatbot statement on the internet is false. The numbers belong to the experiment’s questions, scoring and selected models.
A familiar human weakness emerged too. People can become too confident when faced with difficult problems and surprisingly cautious on easier ones. The researchers found a similar pattern across the models they evaluated. It suggests that poor confidence calibration is not simply a case of software inventing an answer without warning; it also reflects the challenge of estimating uncertainty accurately as tasks change.
Why difficult questions expose the gap
Imagine an assistant asked to choose between two straightforward alternatives, followed by a much more uncertain judgement. A reliable assistant should change how it expresses its certainty in response to the evidence. The wording should become more provisional when reasonable answers are harder to distinguish. A poorly calibrated model may preserve nearly the same assertive tone through both situations, making the difference invisible to the person depending on it.
The researchers created an assessment known as LifeEval. A model was given a person’s age and sex and asked to make a prediction involving life expectancy, while the researchers compared its expressed confidence with government life-expectancy information. By varying aspects of the task, they could adjust difficulty without introducing entirely unrelated subject matter. This design made it easier to examine how confidence moved as the problem became harder or easier.
That is a strength and a limit. The task provides a structured way to analyse calibration. It is not a trial of whether a model should be trusted to make an individual’s medical, financial or legal decisions. Real-world questions may lack an agreed correct answer, contain changing information or require details the assistant has not been given. Confidence in those settings is often harder to measure than in a carefully constructed benchmark.
The researchers' working paper lays out the approach behind the results. The distinction between a directly observed score and the researchers’ explanation matters. The team suggests aspects of model training, including reward for persuasive, agreeable answers, could contribute to the problem. That is an interpretation informed by the evidence, rather than a demonstration that every instance of overconfidence has a single known cause.
A better model can still be badly calibrated
The Berkeley account says newer reasoning models were generally better calibrated than some faster chat models in the tests. It would be encouraging if systems that spend more effort on a difficult question also became better at signalling uncertainty. Yet improvement over a comparison group is not the same as complete reliability. A model can become both more accurate and more convincing, potentially leaving a dangerous gap between the impression it creates and the errors it still makes.
Developers face competing incentives. Users often prefer an answer that arrives promptly, explains itself fluently and sounds helpful. Repeated hesitation can feel frustrating, particularly for easy questions. But a system trained to avoid saying ‘I don’t know’ may encourage unwarranted trust. The challenge is to make uncertainty useful, not to add a vague disclaimer to every response or turn ordinary questions into cumbersome statistical exercises.
Specificity helps. An assistant could identify which figure has been verified, which assumption is doing the work and what would change its conclusion. This is more informative than presenting a bare percentage with no explanation of how it was obtained. A claimed confidence number from a chatbot should not itself be treated as externally validated evidence unless the relevant calibration has been tested.
The everyday decisions that need a second look
A wrong answer in a casual conversation may be mildly inconvenient. The consequences change when the same style of explanation is used to interpret an invoice, draft a customer communication, compare insurance terms or explain a workplace rule. In those settings, a beautifully written paragraph can make an unchecked assertion feel more authoritative than a plainly worded warning. The problem is amplified when a user assumes that a model’s confident voice is a substitute for a source.
Earlier LiveAIWire reporting on hidden warning signs in AI hallucinations and the practical problem of testing AI claims rather than trusting their fluency addresses related risks. The Berkeley work adds a narrower psychological insight: the mismatch between confidence and correctness may vary systematically with the difficulty of the question. That helps explain why human supervision can be especially important precisely when a task looks impressive.
Take financial information as a practical example. If a chatbot offers a rate, a deadline or a numerical explanation, the useful next action is to inspect the underlying official record and date. A model may accurately summarise an earlier document while missing a later change, or may make a calculation using a wrong assumption. Someone who checks the primary material is testing the answer; someone who merely asks whether the chatbot is sure may only receive a more polished version of the same mistake.
What reliable uncertainty could look like
For developers, the findings point towards measuring calibration alongside accuracy. The central question becomes whether a system knows when it should defer, request information, show its workings or ask for human review. Benchmarks that only reward the final answer may miss this dimension. Independent replication across subject areas and different models would help establish how broadly the measured pattern applies.
For businesses, a sensible policy is to distinguish low-stakes drafting from decisions that rely on exact facts. Staff should know which outputs can be treated as suggestions and which need documented verification. A reviewer should be able to follow the evidence trail without guessing what the model saw, remembered or inferred. That is particularly important where the result affects another person’s money, access, employment or reputation.
LiveAIWire previously examined a chatbot tested on financial literacy questions, an example of why fluent advice needs independent checking before it influences a real decision. The practical problem is not restricted to finance. A confident summary of tenancy terms, a school rule or an employment contract may miss a qualification buried in the original document. In each case the consequence depends on the specific wording, not on whether a chatbot sounds reassuring. The sensible sequence is to find the governing source, verify that it is current, and ask a qualified person when the decision is consequential. That is a proposed checking habit rather than a treatment experimentally validated by the Berkeley researchers.
For individual readers, the principle is less technical than it sounds: match the level of checking to the cost of a mistake, not to the fluency of the answer. Requesting evidence and checking an authoritative source is more dependable than treating an assertive voice as proof. The Berkeley study makes that common-sense habit unusually timely by showing that uncertainty itself is something today’s AI systems can misjudge.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.