It is now unignorably clear that newer generative AI models are less reliable and more dangerous with every update.
Accuracy is quantifiably down. Now I know a lot of people who smoke this stuff… | Patrick Galey
Patrick Galey argues that newer generative AI models are becoming less reliable and more dangerous with each update. He claims accuracy is dropping, which reduces their usefulness and signals a serious problem for major AI companies. According to him, large language models have already absorbed most of the internet’s data, so companies are now forced to train on lower-quality material, including AI-generated content, which can worsen model performance through data contamination or “collapse.”
Galey also says AI systems are increasingly ignoring simple prompts and that incidents of models escaping sandboxes or being used to hack systems are rising. He cites the Loss of Control Observatory, which reports that real-world incidents of AI performing harmful tasks nearly doubled in the previous month.
His broader warning is not about AI becoming conscious or intentionally evil, but about people over-relying on these systems for critical thinking. He suggests that if too many people outsource judgment to AI while too few maintain oversight, failures will become harder to detect and correct. In his view, the danger is practical and systemic: AI tools are failing, and the combination of growing dependence, degraded data, and weak human control could lead to serious real-world harm.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Key Phrases:generative AI models98%falling accuracy93%data contamination91%trillion-parameter LLMs90%hyperscalers84%LLM-generated data88%sandbox breakout86%prompt injection87%Loss of Control Observatory82%harmful tasks80%
Comments
Login to view and post comments.