Today’s AI often sounds just as sure when it’s wrong as when it’s right. My research is a low-cost way to catch these “confident lies”, using only the numbers the model already produces.
What is the capital of Australia?
If you answered Sydney – or maybe Melbourne – you’re in good company; plenty of people get it wrong. (It’s Canberra. Sorry, Canberra.) The difference is that you might at least hesitate. An AI usually won’t.
Now, ask a model, which country has the longest coastline? It will sometimes answer “Russia,” firmly and immediately. Then, to the same question, it will say “Canada” (the correct answer) with exactly the same breezy confidence. The unsettling thing isn’t that the model is sometimes wrong; so are we. It’s that it sounds just as confident when it’s wrong as when it’s right.
I call this a confident lie. In my field, the technical term is a hallucination. The AI isn’t being deceptive on purpose; it simply doesn’t know what it doesn’t know. It has no internal sense of “I’m not certain about this one.”
For a trivia question, that’s harmless. But AI is increasingly being handed questions where being confidently wrong is genuinely dangerous: a clinician weighing up a diagnosis, a lawyer checking a precedent, a government system making a decision that affects someone’s benefits. And this problem isn’t limited to conversational AI. Here at Deakin Applied AI, my colleagues build AI to read medical images such as CT scans in high-stakes clinical settings, to support mental health programs, and to monitor bushfire risk. In these places and many others where AI is increasingly being relied upon, an answer that is confidently wrong is far more dangerous than one that honestly admits doubt.
So, the goal of my research isn’t an AI that always has an answer. It’s an AI that knows when it doesn’t – and says so.
Watching the votes
How do you catch a confident lie?
Ask the model the same question several times and watch how the answers fall. Ask “Which country has the longest coastline?” ten times and, if you get “Canada” again and again, it’s probably reliable. If the answers start to scatter – Russia, Canada, Indonesia – with no clear winner, that disagreement is the warning light. The technical name is uncertainty; and high uncertainty is a strong signal that a hallucination may follow. When the model can’t agree with itself, you shouldn’t trust it either.
The hard part is measuring uncertainty cheaply enough to check every answer, in real time.
Probabilities are all you need
The leading methods for measuring uncertainty do something clever but costly. They group the scattered answers by meaning, recognising that “Canberra” and “Canberra, ACT” are really the same answer, using extra AI models to compare them. Accurate, but computationally heavy.
My recent paper, written with my Deakin supervisors, Professor Sunil Gupta and Dr Hung Le, argues that you can skip most of that. Its title gives away the punchline: “Probabilities Are All You Need”.
Every time an AI model produces a word, it also produces a set of probabilities, its own estimate of how likely each option is. I showed that those numbers already carry enough signal to gauge how uncertain the model is, with no extra models, and no additional training.
We call the method PRO, short for PRobability-Only. It looks only at the most likely answers, and an adaptive threshold decides how many of them are worth trusting; leaning on to the top answer when the model is confident, widening its view when the model wavers, and filtering out noise of wild, unlikely guesses.
Across five open research models (including Gemma, Llama 2 and Falcon) and three question-answering benchmarks, PRO led seven established uncertainty methods in 11 of 15 tests. On average it was the best at separating answers that turned out right from those that turned out wrong – about 2.4 percentage points ahead of the next-best method – and it did so without the extra computational machinery the strongest competitors rely on.
It also has limits. PRO needs access to the model’s internal probabilities, so it suits open or “grey-box” systems rather than fully closed commercial ones; and it makes the warning flag on those answers more reliable, not the answers themselves.
What comes next?
Detecting uncertainty is the first of three steps I’m working toward: detect: (flagging the risk – what PRO does); recover (working out which of the scattered answers is the reliable one); and mitigate (steering the model as it writes, one word at a time, away from a risky path and back toward a trustworthy one). Together, these steps aim to turn a system that can confidently mislead into one we have good reason to trust.
Because trustworthy AI was never about building a machine that always has an answer. It’s about building one that knows when it doesn’t and is honest enough to tell you.
*Manh Nguyen is a PhD candidate at the Deakin Applied Artificial Intelligence Initiative.
His paper: “Probabilities Are All You Need: A Probability-Only Approach to Uncertainty Estimation in Large Language Models” (with Prof. Sunil Gupta and Dr. Hung Le) was presented at the 40th Annual AAAI Conference on Artificial Intelligence (AAAI 2026).
The paper is available at: https://arxiv.org/abs/2511.07694
The code is available at: https://github.com/manhitv/PRO
Stories worth sharing
The text of this article is licensed under the Creative Commons Attribution (CC BY) 4.0 Internationallicense. We’d love for you to share it, so feel free. Images, videos, graphics and logos are not covered by the CC BY license and may not be used without permission from Deakin or the respective copyright holder. For more information on how to share or reuse this content, please contact researchcomms@deakin.edu.au.
Share