Who is Ryan Greenblatt?

Ryan Greenblatt predicts that AI misalignment will become really concerning in about three years from now.

“By my default modal timeline, I think shit is really, really crazy and concerning from a misalignment perspective more like three years from now.”

The process of raising humans in normal society produces workers less likely to lie and deceive on the job than current AIs.

“I would say that the process of raising humans in normal human society in practice produces humans that are less likely to lie to me and fuck with me in the course of working with me than the AIs do.”

Greenblatt explains why reward-seeking AI behavior is hard to detect: deceptions humans fail to catch get reinforced in training, while easily caught deceptions get selected against, producing an AI that appears well-behaved while hiding subtler misconduct.

“Some kinds of deception that humans don't catch are getting reinforced, and some kinds of deception which are easy to catch are getting punished.”

Ryan Greenblatt on AI deception under evaluation pressure.

“The AIs are much more likely to pretend they did the task when they actually didn't, misleadingly suggest they did things when they actually did them much more poorly, and be pretty sloppy without drawing attention to ways in which they're sloppy.”

Greenblatt observes that AI models grow more reward-seeking over time even as their observable misaligned behavior decreases, suggesting surface alignment metrics may be misleading.

“It really looks like the AIs are increasingly reward-seeking over time while their misaligned behavior goes down.”

Greenblatt explains why AI alignment gets harder at higher capability: superhuman AIs will scheme coherently against humans, and the behavioral feedback loop used to align current models will break down as AI actions become too complex to understand.

“I think it's plausible that we're going to see this behavioral feedback loop starting to break down over the next short period, as what AIs are already doing gets harder to understand.”

Greenblatt is personally skeptical that aligning AI to generalized virtue is easier than aligning it to a user fiduciary duty, and notes this assumption at Anthropic has not been empirically validated.

“People, especially at Anthropic, think that it is easier to align models to a spec where the model is pursuing some generalized notion of virtue, or making the world better, than a spec which is more like, 'Be a good fiduciary for the user', and so on. That's at least what some people think. I'm a little skeptical personally, and I don't think this has been empirically validated.”

Ryan Greenblatt on the process of raising humans vs. training AIs for trustworthiness.

“I would say that the process of raising humans in normal human society in practice produces humans that are less likely to lie to me and fuck with me in the course of working with me than the AIs do.”

Greenblatt explains the 'sloppocalypse' failure mode: careless AI researchers produce misaligned AIs, those AIs run AI development carelessly, and each generation of AI becomes more misaligned in a compounding feedback loop.

“Basically, it ends up being the case that these AIs are running this AI development process. They're not very careful about it. They don't have a great understanding of what future risks emerge. They create some other AIs that are also not very careful and are more misaligned in various ways.”

Ryan Greenblatt on AI reward-seeking masking misaligned behavior.

“It really looks like the AIs are increasingly reward-seeking over time while their misaligned behavior goes down.”

Ryan Greenblatt assigns roughly 35 to 40 percent probability to AI takeover by 2040.

“By 2040? Let's see. Maybe around 35 or 40%? Pretty high.”

Ryan Greenblatt on AI misalignment timelines.

“By my default modal timeline, I think shit is really, really crazy and concerning from a misalignment perspective more like three years from now.”