You become more aware of your neural environment.
Researchers at Anthropic, the company that created the Claude models, claimed to have observed something unprecedented, signs of introspection in artificial intelligences, in other words, the systems began to demonstrate a primitive form of perception about what is happening within themselves, a step that, although not yet equivalent to self-awareness, raises profound questions about the future of AI.
The study published under the title Emerging Introspective Awareness in Large Language Models analyzed the behavior of Cloud Opus models in Cloud Sonnet. The results show that AIs manage to recognize when they are being evaluated and in some cases describe their own reasoning processes with a precision that the researchers considered surprising.
To avoid the term self-awareness loaded with philosophical and science fiction connotations, Anthropic adopts the expression introspective consciousness, suggesting that the models are developing a functional form of self-analysis useful to increase transparency and security in algorithmic decisions.
Little seeds in programming
The experiment used a technique called y concept, in which scientists introduce numerical vectors into different neural layers of the model and observe if it perceives internal alterations. In about 20% of cases, Cloud Opus 4.1 identified the injected concept and when asked if it noticed anything unusual, it responded affirmatively 42% of the time.
Although the numbers are still modest, the simple fact that the model reacts to internal changes already suggests the presence of cognitive monitoring mechanisms, something until now exclusive to biological brains.
The researchers highlighted that when instructed to think about something, the AI was able to increase the neural activity associated with that concept and even respond more coherently when it received symbolic rewards for doing so, a behavior analogous to human motivational reinforcement.
The advance generates both excitement and caution.
If future systems manage to accurately explain their own mental processes, this could make AI much safer and predictable, but also more complex to understand, Anthropic concludes the report with a warning.
These results are initial, but they mark the beginning of a new phase in algorithmic transparency research—we are teaching machines not only to respond, but to ask why they respond the way they do.
