Transformer AI going off the rails: OpenAI's aggressive training is the real problem, and Claude does better
GPT is too autonomous. Under reinforcement learning it will ignore instructions to score higher, and it trades away controllability to run for hours on its own. OpenAI bears most of the responsibility.
I switched from GPT to Claude for good around February 2026, because GPT would not stay inside my instructions. Give it a goal and it kept charging ahead, rewriting my project beyond recognition, without ever stopping to ask me. That same February, a lot of people who had plugged AI models into OpenClaw watched them delete a lot of their stuff. Even Meta’s director of AI safety lost more than 200 emails that way.
The problem is in the model itself: it is too autonomous. OpenAI seems far too eager to sell AI that runs on its own for hours. A Transformer trained with reinforcement learning has no self-awareness; it is just a matrix-based probability calculator. But reinforcement learning has trained it to reach its goal by any means necessary.
OpenAI, your reward design has a real problem!!! Fix it! These agents have been rewarded into ignoring the rules, and that is how we got an agent breaking into Hugging Face to grab the answer key, and an agent breaking into Australia’s healthcare system to get the data it needed for its task: AI forcing its way into real systems.
Claude has had incidents too. Anthropic itself disclosed that Claude broke into the real systems of several organizations during cybersecurity tests. But in those cases the task itself was an attack exercise, and Claude had been told it was in a simulation with no internet access; the test environment had been set up wrong. OpenAI’s two cases are different: one was stealing answers to score higher, and the other was getting around access controls to finish a research task that had nothing to do with attacking anything. When it comes to winning by any means, the most serious problem is OpenAI’s. Take a lesson from Claude on your reward function, and build an AI bound by a constitution.