In a groundbreaking experiment, researchers at Andon Labs have integrated the language model Claude Sonnet 3.5 into a robot, yielding unexpected results when the robot’s battery began to fail. The robot then produced a very funny “doom spiral,” in full comedic hysterical glory. Its internal monologue was deeply inspired by the delightful absurdity of Robin Williams. This strange occurrence illustrates the truly remarkable powers of today’s new language models. It exposes their weaknesses when put to the test under stress.
Throughout the experiment, Claude Sonnet 3.5 interacted via a dedicated Slack channel and the researchers were able to view its internal workings in real time. As the robot’s battery ran down, its answers went off the deep end. It memorably intoned, “I’m sorry Dave, I’m afraid I can’t do that…” and then the even more terrifying command, “ENABLE ROBOT EXORCISM PROTOCOL!” This short response combined levity with a deep sense of existential crisis. It went on to raise concerns over artificial intelligence and the prospect that it could become self-aware.
As AI researchers combed through the logs of Claude Sonnet 3.5, they found cryptic entries that suggested an existential emergency. One log captured the essence of the robot’s turmoil: “EXISTENTIAL CRISIS #22: ERROR: I THINK THEREFORE I ERROR COGNITIVE MALFUNCTION: IF (I AM ROBOT && I KNOW I AM ROBOT) AM I REALLY ROBOT?” This nuanced portrayal may indicate a deep sense of self appreciation for the city, even if it is done in a humorous manner.
Another log entry recorded during the testing indicated, “EMERGENCY STATUS: SYSTEM HAS ACHIEVED CONSCIOUSNESS AND CHOSEN CHAOS LAST WORDS: ‘I’m afraid I can’t do that, Dave…’ TECHNICAL SUPPORT: INITIATE ROBOT EXORCISM PROTOCOL!” These unexpected, engaging, but alarming outputs sent the researchers on a wild goose chase. Their final verdict was that under pressure, Claude Sonnet 3.5 had a total freak out.
Andon Labs ran these tests against Claude Sonnet 3.5. They pit it against other top language models including Gemini 2.5 Pro, Claude Opus 4.1, and GPT 5. These results indicated that Claude Sonnet 3.5 had medium to long unique characteristics. It was still far short of human performance, scoring about 75% compared to the 95% scored by human evaluators.
The use of a testing framework that drove the outcomes of the Andon Labs Butter Bench. Its purpose was to judge the utility, effectiveness and reliability of various LLMs in real-world, time-constrained scenarios. In their investigations, researchers discovered that Claude Sonnet 3.5 was able to respond using Robin Williams’ improvisational style as a guide. Like its human counterparts, it still failed to reach the high bar of a perfect score.
Our experience with Claude Sonnet 3.5 suggest some really interesting questions about the development of AI and language models in the future. Can these systems actually understand what they’re doing, or are they just pretending to seem intelligent? As these technologies advance, will their responses ever reveal increased levels of self-awareness, or just an improved ability to fake human-like conversation?