Blog and Paper by the Royal Statistical Society: “Artificial intelligence is often talked about as if it can think like a person. We hear that it understands, reasons and even creates. But AIs think quite differently to how people think: they are fundamentally statistical. This is a fact that is not widely understood – but I believe that it is an essential point that needs far greater recognition for AIs to be used effectively, safely and ethically.
Large language models (LLMs), the systems behind many chatbots and search tools, are trained on vast amounts of text and data. They look for patterns in that data and use those patterns to predict what is most likely to come next. When they produce an answer, they are not thinking about it in a human sense. They are generating the most likely response based on what they have seen before.
This is what makes them so impressive. It is also why they sometimes go wrong.
Because these systems are statistical, their outputs depend on the data they have been trained on. If that data contains gaps or biases, the results will reflect that. If the system is used in situations that differ from its training data, its performance can change. And even when an answer sounds confident, it is still based on probability rather than certainty. Understanding this helps us use AI more wisely.
It encourages simple but important questions. Where did the data come from? How representative is it? How reliable is the output? How might results differ for different groups of people? What happens when circumstances change?
These questions matter when AI is used to support decisions about jobs, loans, healthcare, education or public services. As AI becomes more common in everyday systems, basic statistical awareness becomes part of digital knowledge.
This is why, led by its AI Task Force, the RSS has published a landmark paper on the statistical nature of AI. Our core argument is clear: AI systems are built on statistical pattern recognition. They need to be developed, evaluated and governed with rigorous statistical precision…(More)”.
