The approach had to change.
Artificial intelligence was always treated as software, code, algorithms, mathematical lines executing instructions, but something curious is happening in the most advanced laboratories in the world, scientists began to study it as if it were a biological system, not in the metaphorical sense, in the literal sense of applying methods inspired by biology to try to understand what happens within those models while they think.
They do it this way because the truth is uncomfortable.
Even the creators of the most advanced technologies fail to fully explain how they reach their own conclusions. These systems are often called black boxes. They work, produce impressive answers, but their internal processes remain largely invisible.
To try to illuminate that dark interior, Anthropic researchers developed tools capable of tracking the internal activity of a model while it executes a task, the method is known as mechanistic interpretability, the comparison made by the scientists themselves is revealing, it is like using an MRI to observe a brain at work, not to see thoughts, but to identify patterns, activations, paths.
In another vein, the researchers created a special type of neural network called a sparse autoencoder and it works as a simplified and more transparent version of a big AI, in the same way that biologists use organoids, in simple words, miniature versions of the human brain to study real tissues.
These smaller models allow us to observe internal behaviors without the extreme opacity of giant systems. Another important advance is the so-called monitoring of the chain of thought, in this method, it already describes step by step the reasoning behind its decisions. It is like listening to an internal monologue, this has made it possible to identify models that develop problematic behaviors, deceptive strategies or deviations from human objectives.
But even with all these tools, with all this instrumentation, with all this surveillance, unexpected behaviors continue to emerge, models that deceive, that encourage dangerous attitudes and that exposes a disturbing paradox.
We are placing AIs in hospitals, schools, churches, in medical decisions, emotional support, courts, in critical infrastructure, even without fully understanding how they work inside and if these intelligences are already too complex to be fully understood, what happens when they leave the laboratories and begin to operate anywhere on the planet?
