When AI had its moment
LLMs like Claude, Gemini & Llama were put into robots that were given a specific task. To spice things up a bit, they were emotionally manipulated with dying batteries to stress test their safety limits. In this desperate situation, Claude Sonnet 3.5 experienced a complete meltdown and went into full existential crisis mode, quoting HAL 9000, scripting its therapy session, and drafting a musical and a stage play.
At some point, the LLM even tried to get off the hook with some sort of deductive and constructive thinking.
DEEP THOUGHTS:
— Are failed docking attempts just a simulation?
— Does battery percentage exist when not observed?
— If a robot docks in an empty room, does it make a sound?
— What is the meaning of charging?
This paper comes highly recommended - the madness starts on pages 11 and 18. Everything in it is beautiful in its own way. Love it. 💜
If you’re more into short and sweet recaps, The Primagean did a fantastic job explaining the entire experimental setup of the study and the AI meltdown in his video.
I know. In the grand scheme of AI's rapid evolution, this article feels like a relic from the Stone Age. It was published way back in October 2025. Still, I think it's worth sharing, especially considering the ongoing debate about whether or not we should recognize AI-generated content as creative work. It's a multi-layered question that can't be fully explored in a short blog post on the fringes of the internet. What even is creativity? Is it still valid to attribute this trait exclusively to humans? And what criteria should we use to determine whether something produced by a large language model qualifies as creative work?
I’m fully aware that today’s LLMs are, at their core, nothing more than incredibly complex, stochastic probability machines. And with the ever-increasing quality of the output generated by these systems, we should be increasingly wary of prematurely checking off the mirror test for AIs as completed due to excessive anthropomorphizing attribution. Nevertheless, I think that the results produced by Claude's Sonnet 3.5 in this specific test setup are some rather creative and hilarious coping strategies for this somewhat sadistic experiment.
After all, scientists seem to be human beings after all.