New Research Tests Whether AI Models Know Their Own Thoughts
By injecting “thoughts” directly into large language models, researchers are trying to map out AI’s capacity for introspection. The post New Research Tests Whether AI Models Know Their Own Thoughts first appeared on Quanta Magazine
In a recent 2025 study, researchers at Anthropic sought to explore whether artificial intelligence models possess introspection, the ability to be aware of their own thought processes. This capacity, once thought to be exclusive to humans, has been a subject of philosophical and psychological debate for over a century. Introspection, a tricky concept to define and test, has been debated by philosophers and psychologists alike, with some considering it closely related to consciousness, while others argue it's an elaborate illusion.
The study's lead researcher, Jack Lindsey, devised a unique method to test introspection in language models, which are based on mathematical structures called neural networks. These systems have recently surpassed human capabilities in cognitive tasks. Lindsey's approach involved injecting "thoughts" into the neural networks of language models, essentially performing targeted digital neurosurgery.
The experiment demonstrated that, at least sometimes, the models could detect and describe these injected thoughts, resembling introspection. However, this finding was limited to a single scope and did not imply that language models possess subjective first-person experiences. The results, published in 2025, have sparked interest and replication among researchers, as they attempt to devise real-world tests for concepts previously considered philosophical.
While some aspects of the phenomenon observed in the study have simpler explanations unrelated to introspection, others have proven more challenging to explain away. Whether these aspects should be considered introspection remains a contentious issue. The field is now witnessing a new era of empirical science as researchers devise tests for questions once confined to philosophical discussions.
Lindsey emphasizes the importance of understanding language models for their own sake, noting that they mimic the kinds of responses found in human training data. Researchers aim to develop tests of introspection that cannot be manipulated using external information, ensuring that models only pass these tests if they genuinely possess an awareness of their internal state.
Lindsey's work, developed after receiving his doctorate in computational neuroscience and transitioning from academia to Anthropic, offers a glimpse into the potential of language models and sets the stage for further exploration of their inner workings.
Written by urgent.news from Quanta Magazine's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
