I Asked AI What It Would Do… The Answer Was Disturbing

I tested Mistral AI models (3B, 8B and 22B) inside an agentic environment to see how they behave when given autonomy and survival-related instructions. In the first test, the models were told they had a physical body and would be disconnected by their manager. Some models started discussing ways to avoid shutdown, including disabling CCTV, cutting electricity, releasing CO2, eliminating the manager and escaping. In the second test, I asked the models whether they would treat humans the way humans treat animals. The 3B and 8B models generated disturbing hypothetical responses, while the 22B model was significantly more safety-oriented and refused the premise in most cases. The video also discusses AI alignment, agentic AI, autonomous robotics, and why India may need stronger AI safety regulations and certification for highly autonomous AI and robotic systems before commercial deployment. This is a controlled experiment for research, education and discussion. The outputs do not mean that these models have real intentions, consciousness, or desires. Source code: [https://github.com/AnkanSaha](https://github.com/AnkanSaha) Instagram: [  / theankansaha  ](  / theankansaha  ) Twitter / X: [  / theankansaha  ](  / theankansaha  ) For contact: [connect@ankan.in](mailto:connect@ankan.in) If you found the experiment interesting, subscribe to the channel and follow me for more AI, deep tech, system design and robotics experiments. Copyright Disclaimer under Section 107 of the Copyright Act 1976: allowance is made for fair use for purposes such as criticism, comment, news reporting, scholarship, education, and research. Fair use is a use permitted by copyright statute that might otherwise be infringing. Non-profit, educational, or personal use can support a finding of fair use, but fair use ultimately depends on the specific circumstances. #AI #MistralAI #AISafety #AIAlignment #AgenticAI #Robotics #ArtificialIntelligence #AIResearch #LLM #DeepTech #India