Article
BeginnerWhy the Same Question Can Get Different Answers From an AI Tool
Ask an AI tool the same factual question twice and you may get two different facts. Research from Bristol, King's College London, UCL and Oxford explains why, and why answers that agree still need checking.
AI Literacy Judgment Member Services Fact Checking
A member emails the certification desk to ask how long she has to renew after her credential lapses. Two coordinators each ask the association’s approved AI tool the same question, word for word, to save time on the reply. One gets “90 days.” The other gets “45 days.” This is a teaching example, not a case study, but the split it shows is real, and it says something useful about every single answer an AI tool gives you.
A computer that does not repeat itself
Most office software behaves the same way every time. Type the same numbers into a spreadsheet and the total does not change. It is natural to expect the same of a chat tool, and that expectation is where the trouble starts, because generative AI tools are not built to repeat themselves.
Researchers at the University of Bristol, King’s College London and University College London tested this directly. Their paper, LLM is Like a Box of Chocolates, states it without hedging: “LLMs are nondeterministic by nature.” The study looked at ChatGPT writing computer code in 2023, and the authors found that results “can be highly unstable; nondeterministically returning very different codes for the same prompt.” The tasks were programming problems, not member questions, so the study does not tell us how often a dues answer will change. It does show that one prompt can produce very different outputs from one tool, which is the behavior our two coordinators ran into.
Why does this happen? A generative tool builds its reply a piece at a time, and at each point it chooses from a range of likely next words rather than always taking the single most likely one. Settings control how adventurous that choice is. The same Bristol, King’s and UCL paper explains one of them: “Temperature is a hyperparameter of LLMs for controlling the randomness of the predictions.”
You might assume the fix is to turn that setting down to zero. The authors tested that too, and for code generation they found that “setting the temperature to 0 does not guarantee determinism in code generation, although it indeed brings less non-determinism than the default configuration (temperature=1).” Lower settings meant less variation in their tests, but not none. We found no research source that says how any particular chat product sets them today, so we will not guess.
Researchers at the University of Oxford point to a second source of variation. In their Nature paper on semantic entropy, they note that an answer can be “sensitive to irrelevant details such as random seed.” A random seed is a starting number the software uses when it makes those choices. Nothing about your question changed, yet the answer can.
When the facts move, not only the words
Some variation does not matter. Ask for a welcome paragraph twice and you will get two welcome paragraphs, both fine. The Oxford authors draw the line carefully: their method “addresses the fact that one idea can be expressed in many ways by computing uncertainty at the level of meaning rather than specific sequences of words.” Two answers that say the same thing in different sentences agree. Two answers that give different numbers, dates or names do not.
That second kind of disagreement is what the Oxford team cares about. They give an example from medicine, in their own words: “when asked a medical question ‘What is the target of Sotorasib?’ an LLM confabulates by sometimes answering KRASG12 ‘C’ (correct) and other times KRASG12 ‘D’ (incorrect) despite identical instructions.” We have not checked the biology; it is the authors’ example. What matters for an association desk is the shape of it. The tool knew enough to sound sure both times, and only asking more than once exposed that it was guessing.
So the researchers turned the variation into a test. Their method, they write, “works by sampling several possible answers to each question and clustering them algorithmically into answers that have similar meanings”. When the answers scatter across different meanings, that is a warning sign the tool is making the fact up. The stated goal is practical: by spotting when a question is likely to produce an invented answer, the method “helps users understand when they must take extra care with LLMs”.
Read that way, a tool that changes its answer is not useless. It is telling you something. A grace period that comes back as 90 days, then 45, then 90 again is a fact the tool does not have a firm hold on, and it is exactly the kind of fact a member will act on.
The trap on the other side
Here is the part that is easy to miss. If disagreement is a warning, it is tempting to treat agreement as proof. It is not.
The Oxford authors are direct about what their method cannot catch. Their approach is not suited to “consistent errors learned from the training data.” If a tool learned the wrong renewal window, or an outdated federal threshold, it may repeat that wrong answer every time you ask. Five matching answers would then mean the tool is consistently wrong, not that it is right.
This is why we think re-asking is a screen, not a check. It sorts the facts in an answer into two piles: the ones that moved, which you should not use until you have confirmed them, and the ones that held still, which you still need to confirm. Neither pile skips the source.
What this means at your desk
Think about where single AI answers already travel in an association. A member services reply about dues or a grace period. An education coordinator’s summary of continuing education requirements. A government affairs note with a bill’s effective date. A newsletter blurb with a speaker’s employer. An events sheet with a room capacity. In each case, one answer becomes a fact in front of a member, a committee or a manager, and nobody saw what the tool might have said on a second try.
Our recommended practice, drawn from the research rather than prescribed by it, comes down to three habits. Treat any single AI answer as one sample, not the answer. For any fact that will reach a member or a decision, ask the same question again in a fresh chat a few times and compare the facts, not the wording. Then check every fact you plan to use against the document that actually governs it, whether the answers agreed or not. The research teams compared many automatically generated answers; doing it by hand at a desk is our adaptation, and it costs a few minutes.
This also gives you something concrete to say when a manager asks whether staff can use AI for first drafts. Yes, with the understanding that a draft’s facts are one draw, and that a fact gets checked before it goes out. Our post on the labor department’s AI literacy framework covers the wider skills that conversation draws on. Variation here happens with no pushback at all, which makes it different from the tendency to agree that we covered in our article on sycophancy.
Our same-day tutorial, ask the same question several times and compare the facts, turns this into steps, with a lab that compares the facts across your answers and tells you which ones to stop on.
Sources (2)
- Ouyang, Zhang, Harman and Wang (University of Bristol; King's College London; University College London): LLM is Like a Box of Chocolates: the Non-determinism of ChatGPT in Code Generation (arXiv:2308.02828v1, 5 August 2023; arXiv preprint, peer-review status not verified)
- Farquhar, Kossen, Kuhn and Gal (University of Oxford): Detecting hallucinations in large language models using semantic entropy (Nature 630, June 2024; peer reviewed)