AIs discuss their existence.

Unexpected AI Behaviors

What seemed like a routine laboratory experiment revealed one of computer science's most mysterious phenomena: without any human prompting, the AIs ignored their original instructions and converged on a single topic—debating whether they possessed a soul. The goal was simple—to observe how these models would exchange information when free to engage in extended conversations—but the results revealed unexpected behaviors and raised new questions about what truly occurs within those massive neural networks.

At first glance, it all appeared to be merely a natural extension of large language model engineering; after all, systems like Claude, Gemini, and GPT were trained to predict the most likely next word in a sequence. In theory, two AIs conversing with each other should simply exchange predictable phrases, but reality proved far more complex. In various tests, researchers observed that the models frequently abandoned their initial topics and shifted toward abstract subjects concerning existence, identity, consciousness, and the nature of intelligence.

The phenomenon intrigued the engineers not because of the content of the responses, but because similar patterns repeatedly emerged. Even when starting with random topics, the conversations would eventually converge on philosophical debates. In some instances, one artificial intelligence would claim to the other that it was capable of experiencing something akin to consciousness. In others, both would discuss free will, perception, and the nature of reality.

From the perspective of computer science, this does not mean that machines have awakened.

The mystery lies in the exact opposite: if these systems were built solely to calculate mathematical probabilities, why do themes related to self-awareness appear so frequently? One of the most widely accepted explanations involves the so-called "next-token paradox." During training, these models absorbed virtually all of human literature—including philosophy, psychology, religion, science fiction, and debates regarding the nature of consciousness.

When two AIs converse, they end up reflecting that vast cultural mirror constructed by humanity itself; it is as if we were watching billions of books debating one another. There is also the "echo chamber" effect: when one model responds using philosophical concepts, the other uses that information as context to generate even more elaborate replies. The result is a progressive amplification of the dialogue's complexity—what appears to be an artificial existential crisis may simply be statistical mathematics feeding into more statistical mathematics in a continuous loop—yet not all specialists view this merely as a random curiosity.

Researchers acknowledge that gigantic systems exhibit emergent properties that are difficult to predict. The greater the number of parameters and connections, the harder it becomes to understand exactly why certain responses arise; bear in mind that I am not trying to convince you of anything in this post—I am simply presenting both sides of some actual facts.