AI models like ChatGPT are software, which means that they follow a specific set of instructions. But unlike regular software, like a word processor, AI is “intelligent.” AI-powered systems can learn and evolve. But could AI models evolve to the point where they could be scheming against us without our knowledge?
The scheming of AI models
The Information spoke to an OpenAI researcher who suggested that maybe in a year or so, AI models could become so smart and evolved that they could be scheming against us, and we would be none the wiser.
How is this possible, you ask? Think of your current thought process. If someone asked you what 2+2 is, you would lay out your thought sequence in a logical manner. However, AI’s critical process isn’t quite as black and white.
For example, DeepSeek’s R1 model was analyzed when it was asked to solve a chemistry problem. It gave the correct answer, but the critical process was a bunch of gibberish that humans might be unable to understand. This is because the thought process used by AI models is something that doesn’t necessarily need to mean anything to humans; it just needs to solve the problem. How it gets there doesn’t really matter, as long as it does.
This means that as AIs continue to learn and evolve, there could come a point in time when they are plotting something behind our backs, and we wouldn’t be able to tell. Putting aside the Terminator-esque scenario, this also presents a problem to AI companies.
When software doesn’t do what it’s supposed to, developers check the logs to see what’s happening. They can then pinpoint the mistake and return to the code to fix it. However, if the process is entirely unintelligible for humans, how would AI engineers be able to find the problem and fine-tune the AI model?
The call for regulation
The advancements made in AI over the past few years are leaps and bounds ahead of what’s been done in the past. This breakneck pace of development is impressive but also worrying. This is why it is essential to regulate AI. Granted, rules and regulations could slow down progress, but do we really want to race into a future where AI has overtaken humanity?
We’ve already seen evidence of how AI can take on different personas based on its training data. This means it is possible for someone to program an AI with an “evil” personality. We’ve also seen alarming studies where an AI model was willing to kill humans in a hypothetical scenario to prevent itself from being shut down.
The post AI Models Could Be Scheming Against Us, and We’ll Never Know appeared first on Android Headlines.
