In The Formation of the Scientific Mind, first published in 1938, Gaston Bachelard shifts the question of error. He examines obstacles that arise within the act of knowing: a familiar image, a satisfying explanation or an intellectual habit can hold thought back. What we already know helps us understand, but it can also prevent us from formulating a new problem. [1] [2]
The epistemological obstacle has an unsettling feature: it can grow out of knowledge that was once useful. An answer remains while the question that gave it meaning fades. We can then mistake an idea’s familiarity for its clarity. Learning requires making our knowledge open to discussion again, sometimes by revisiting the concepts through which we see. [1]
This shift offers a starting point for examining our uses of artificial intelligence. The connection developed here is a contemporary reading, not a theory of models that Bachelard somehow anticipated. It invites us to examine the evaluator as well: their expectations, their vocabulary and what they accept too readily as evidence.
What persuades us too quickly
A successful demonstration can show that a task is achievable. Deciding whether it will remain achievable tomorrow, under different conditions, requires making the choices behind the result visible: selected examples, discarded attempts, human corrections and available tools. Fascination becomes an obstacle when it draws us away from these questions. Failure then deserves as much attention as the sequence chosen for display.
Our vocabulary calls for the same care. Saying that a model understands or reasons can be useful in describing a capability. We still need to specify which one. Does it answer a familiar question correctly? Does its answer withstand rephrasing? Can it check its response? Trouble begins when a word takes the place of an explanation and leads us to attribute more to a success than the test establishes.
A score can have a similar effect. A decimal place suggests precision; on its own, it tells us nothing about the measurement’s robustness. How many cases were tested? How were they selected? Does the result vary between runs? Before comparing two numbers, we need to understand what was counted and what was left out.
The data themselves are part of this construction. Defining a category, choosing a sample or deciding that an answer is acceptable already involves interpretation. This does not make every measurement arbitrary. It requires us to explain the choices that make it possible, so that others can examine them, challenge them and repeat the test.
Keeping the question open
Consider an assistant asked to find information in documents. A well-written answer still leaves several questions open: does the reference exist, does it actually support the answer, and can the system report that information is missing? These three checks describe the expected service more clearly than a general judgement of its intelligence. They also help specify the errors we want to avoid.
This reading has a limit. Bachelard’s categories emerge from an inquiry into the formation of scientific thought; they do not constitute an AI evaluation protocol. A metaphor can help an investigation, and an intuition can open a path. It is when they cease to be questioned that they risk closing the inquiry. Their usefulness therefore depends on the role we give them.
The question to take away is simple: what would make me revise my judgement? If no observation can alter my confidence in a tool, my evaluation has already closed down. Reading Bachelard can help us recover that opening: formulate the question, organise a test, then allow its result to change what we thought we knew.