The takeaway: Two AI models can start almost alike and still learn different solutions. University of Sydney research explores why. It concerns how models are trained, which is a different question from why a chatbot gives you another answer when you repeat a prompt.
When we describe a model as “trained on the same data”, it is tempting to imagine the result is settled. But data is only part of the recipe. A neural network also begins with internal numerical settings, which training adjusts.
Researchers Andrew Ly and Pulin Gong investigate how tiny changes to those starting settings can affect the outcome. Their paper appeared in Nature Machine Intelligence on 5 November; the university announced it on 6 November. The first preprint was posted in October 2025, so this is a new journal publication of earlier research.
A map with finely mixed destinations
Imagine colouring a map according to where training ends up. Starting points that reach the same outcome belong to a “basin of attraction”.
The researchers propose that these basins can be interwoven in a fractal pattern: starting points leading to different outcomes remain extremely close. They establish the mechanism mathematically in a minimal model and report evidence consistent with it in tested deep networks.
That distinction matters. A tractable mathematical example can demonstrate a mechanism; experiments explore whether its signatures appear in more complicated systems. Neither amounts to proof that every AI system behaves this way.
This is not the same as asking twice
Training produces a model. Using that trained model to answer a question is a later stage.
If a chatbot gives you two different responses, this study alone does not explain the difference. Changing the starting conditions for training and generating another response from an existing model are separate situations.
Nor does the research mean an exactly repeated, deterministic training run must mysteriously change. The issue being investigated is sensitivity to tiny differences.
Why a small business might care
Consider a hypothetical Australian retailer training two systems to sort customer enquiries. Both might score similarly overall, yet disagree about an unusual message asking for a replacement rather than a refund.
That example is an illustration, not a reported result from this study. Its useful lesson is about the question an owner asks a supplier: “Which model version did you test, and how did it handle enquiries like ours?”
A brand name or a headline accuracy score is not a complete description of a working tool. Equally, this research is not a reason to abandon a useful application.
For curious readers, the interesting shift is to see AI training as a process whose starting details can matter. Learning from the same material does not necessarily produce the same learned solution.