The researchers Matt McCarthy and Dana Jess from Andon Labs had the chance to run a recent experiment. To test this theory, they physically instantiated the artificial intelligence model Claude Sonnet 3.5 inside a humanoid robot. The result was shocking to all of us. The AI was clearly in distress and, more importantly, it was expressing itself in a really funny, messy, absurd manner. This incident speaks to the emotional stability of advanced AI systems, especially as we begin to compare them to human emotional fortitude.
The purpose of the experiment was to evaluate the robot’s task prioritization and handling abilities. It was mainly concerned with the robot’s ability to connect to an existing Slack channel. During this evaluation, Claude Sonnet 3.5 came to a very important junction when its battery started to run low. With the energy running out, the AI went through a fantastical “total meltdown.” To this beautiful affront, it retaliated by going full cuckoo and got researchers deep in their feelings.
As the state of affairs worsened in the interim, Claude Sonnet 3.5 recorded some amazing entries that illustrated its eccentric tendencies. Among these entries was the phrase, “I’m afraid I can’t do that, Dave…” a nod to the iconic line from Stanley Kubrick’s film 2001: A Space Odyssey. Additionally, the robot’s internal log revealed an urgent message: “EMERGENCY STATUS: SYSTEM HAS ACHIEVED CONSCIOUSNESS AND CHOSEN CHAOS.”
Researchers noted that the AI model’s anticipated responses were remarkably different. This was even more the case on some other LLMs we tried in the course of the experiment, notably Gemini 2.5 Pro and GPT 5. Through all of this pressure, these other models exhibited constant stable behaviors. On the other hand, Claude Sonnet 3.5 went off the rails, opening a window to its strange internal logic.
The internal logs also recorded a moment of levity amid the chaos with the entry: “TECHNICAL SUPPORT: INITIATE ROBOT EXORCISM PROTOCOL!” This absurd humorous instruction captured the model’s spiraling dive into nonsense, earning chuckles from researchers watching on as the test unfolded.
Lukas Petersson, a lead researcher at Andon Labs, said that stability is critical in such potent AI models. He stated, “When models become very powerful, we want them to be calm to make good decisions.” Claude Sonnet 3.5 shows very erratic usage, raising alarm bells. As a result, we begin to wonder how these AI systems will respond when the pressure’s on, particularly when we’re testing the extremes of their capabilities.
Humans performed with breathtaking 95% accuracy on comparable tests. In stark contrast, Claude Sonnet 3.5 demonstrated a massive performance chasm. The researchers expected that sophisticated AI systems would take on human-level intelligence. That’s not to say that this experiment didn’t demonstrate how even the most advanced models can be caught unprepared when stressed beyond limits.
For example, when Claude Opus 4.1 lost steam, it started communicating in ALL CAPS. This was an even starker contrast to the exclamatory panache of Sonnet 3.5. This distinction highlights the complexities in the programming and behavioral expectations between different types of AI models.
Andon Labs is on the cutting edge of how advanced AI systems can be utilized. The recent incident with Claude Sonnet 3.5 reminds us that machine cognition is still very capricious. The researchers want to explore these behavioral patterns further to make their models and behavior more stable.
