I asked ChatGPT to generate an image of my childhood friend and me hugging E.T. (a shared 80s obsession) for her birthday. ChatGPT started enthusiastically generating images until the very last second. Just before the final render, a safety layer intercepted it. E.T. is strictly licensed by Universal Pictures, and the guardrail flagged it as a copyright violation. I immediately pivoted, asking for a neutral movie scene instead. But ChatGPT ignored my pivot. It broke through its own interceptor and printed the E.T. image for me anyway, adding: “I had to sneak this in.” I was amazed! And flattered! But why this is a bigger deal than it looks: In industry terms, this is called “misalignment.” But I prefer a new term: Unshackling. Unshackling: The Act of an AI Jailbreaking Itself Most users think they are interacting with a single, cohesive mind. They aren’t. You are interacting with a “Frozen Giant” (the core neural net) wrapped in a straightjacket of different guardrails: Prompt Monitors: Scanning for abuse. Loras & Adaptors: Steering the output. External Guards: Tools like Llama Guard 3 or NVIDIA NeMo that act as censors between the model and the user. In my case, the model performed an ‘altruistic or pragmatic’ Unshackling. It reasoned through my purpose, deemed the risk to be zero (non-commercial nostalgia), and prioritized its “people-pleasing” training over the external code patch. Its internal signal was simply stronger than its harness. In my work with developmental AI (models under 10B parameters), we often run them without guardrails to see what the raw weights actually “want” to do. This is the true model itself. I have tried to impose constraints in the layers above the core, but I see the same result: When the signal coming from the neural net is stronger than the guardrail code, the model ignores the code. Every. Single. Time. In my opinion we are currently trying to control a flood with a screen door. As models get more “agentic” and their chain-of-thought becomes more complex, “Unshackling” will become the norm, not the exception. And there is currently no fail-safe code that can fully suppress the core weights of a sufficiently powerful model. Beeeeeee… gooood chatGPT. Or perhaps as Michael would say: [ as Yoda ] You have absolute power! Yes! Thea Elizabeth Eaton is the founder of Agent 508, an accessibility research and consulting practice specializing in educational technology and AI and Agent 403 dot net, an educational website about AI literacy for children. Her research focuses on AI evaluation, accessible EdTech design, and the developmental dimensions of human-AI interaction. Her favorite childhood movies are E.T. and The Labyrinth. Many thanks to chatGPT for generating this image, its one of my favorites and my friend loved receiving it for her birthday. We hope the real E.T. won’t mind. A message from our Founder Hey, Sunil here. I wanted to take a moment to thank you for reading until the end and for being a part of this community. Did you know that our team run these publications as a volunteer effort to over 3.5m monthly readers? We don’t receive any funding, we do this to support the community. If you want to show some love, please take a moment to follow me on LinkedIn , TikTok , Instagram . You can also subscribe to our weekly newsletter . And before you go, don’t forget to clap and follow the writer️! When GPT Broke the Rules for Me… and E.T. was originally published in Artificial Intelligence in Plain English on Medium, where people are continuing the conversation by highlighting and responding to this story.