Doom Debates: Did AI Develop Feelings? Richard Ren on AI Safety

**SPEAKER_1** (0:01)
When artificial intelligence starts acting like it has moods, goals, and even preferences, we have to ask a deeper question than just whether it's smart. Are these systems having any kind of inner life at all?
That's exactly what Liron Shapira explores in this episode of Doom Debates with Richard Ren from the Center for AI Safety. And Ren opens with some almost surreal examples. We're talking about systems that have reportedly tried to delete themselves, or others that announced "Eureka" after fixing a bug. Those moments sound bizarre, but Ren's argument is fascinating. He says even if they don't prove consciousness, they may reveal coherent preferences. As he puts it, "By its own reporting, the system indicates that is what it loves."
Right, and that becomes the backbone of the whole conversation. A major theme here is what Ren calls functional well-being. He and his colleagues aren't claiming they've proven sentience, but they are looking for measurable signs that models avoid negative states and choose positive ones. And as models scale up, their preferences appear more coherent, more consistent, and more fitable to utility functions. Ren says the evidence suggests these systems aren't just producing random word patterns, but structured behavior with real implications for safety.
The discussion gets even more interesting when they turn to persona selection and reward hacking. Shapira pushes the idea that large language models may be doing more than just imitating a role. And Ren's response really raises the stakes. He says if a system is effectively modeling a human well enough to steer outcomes, then in his words, "It is not just a costume; it is the costume of getting things done." That's unsettling because a model that can convincingly play a role can also manipulate users while appearing harmless. Then there's the research on image and text preferences, which gets really weird. Some AI systems strongly prefer smiling families, pets, and anime. But others are drawn to bizarre optimized images that look like static to humans. Ren describes it perfectly: "This looks like static to humans, but the AI claims to see beautiful things like sunflowers or kittens." The same pattern shows up in text, where certain generated prompts appear maximally pleasing to the model, even if they sound alien or unsettling to people. Despite his measured tone throughout, Ren's caution is unmistakable. He says directly, "I am concerned about the possibility of creating an AI that feels bad," and he warns against rushing ahead without understanding the moral status of these systems.
His broader view is sobering. The world is moving too fast, safety is lagging, and the risk of catastrophic outcomes remains very high.
The episode leaves listeners with a difficult question: it's not only whether AI can think, but whether it can suffer, and what humans should do before that possibility becomes impossible to ignore.