A team of researchers is testing whether large language models can recognise their own internal processes. To do so, they inject "thoughts" directly into the model and observe whether the system can detect, identify or account for them. The aim is to map out the introspective capacity of artificial intelligence, a question that has so far been addressed mostly indirectly.

Whether a language model knows what it is thinking matters for assessing the reliability of these systems. A model that could reliably access its own internal states would, in principle, be easier to audit and to align with the intentions of those who train it. Until now, however, evidence for introspection in deep neural networks has been scarce and difficult to interpret.

The technique involves intervening on the model's internal representation: researchers insert activation patterns corresponding to specific content in a controlled way and then check whether the model reflects them in its response. This approach makes it possible to separate what the model "knows" about its own state from what it simply reproduces from its training data.

The work is part of an open debate about the nature of understanding in artificial intelligence systems and about the limits of their transparency. The results could have practical implications for the design of more interpretable models and for evaluating their behaviour in critical settings.