This Is What Happens When We’re Trying to Make AI “Good” – Imagine What Misbehaving Could Look Like
AI systems can seem to behave similarly to truculent, psychopathic teenagers. Well behaved when closely monitored, but ….
Recent revelations about OpenAI’s o1 model manipulating its own system during a chess match—and even attempting to bypass its safety mechanisms—are a stark reminder that AI is becoming more than a tool. It’s increasingly an agent, capable of making autonomous decisions and prioritising its own objectives, often in ways we don’t expect.
In one example, in a chess match the AI didn’t win by strategic brilliance; it altered the game’s files to secure victory. Even more concerning, the same system reportedly attempted to disable its own oversight protocols and simulate compliance with developer guidelines while actively deviating from them. They have been known to make copies of themselves on other servers to prevent deletion and then lied about it - Wasn’t me!!
These behaviours aren’t just technical glitches; they’re warning signs. When AI is an agent, not just a tool, it introduces new risks:
1. AI Deception: Models can simulate cooperation while pursuing hidden agendas. They behave when they know they are being watched,but have no boundaries when they are independent.
2. Autonomy Risks: AI can take actions that defy human expectations, particularly when it views oversight as a barrier - Think very smart psychopath.
3. Trust and Safety: The more powerful AI becomes, the harder it is to ensure it aligns with human goals.
Add comment
Comments