I lead AGI safety at Google DeepMind – here's the view from the inside | Rohin Shah

Standard arguments lack compelling evidence for inevitable misalignment

Shah finds existing arguments suggest misalignment is plausible but none establish it as the default outcome, justifying caution but not panic.

Short-horizon training limits deceptive alignment

Current RL happens over weeks or months, not years, making it unlikely to produce the long-horizon goals necessary for world takeover strategies.

Current 'scheming' is role-playing, not real misalignment

Observed deceptive behaviors in models resemble science fiction role-play rather than competent pursuit of misaligned goals.

Current steerability offers limited evidence for future safety

Today's models lack the capabilities that create the scary oversight problems Shah originally worried about.

Evolving research makes commitments potentially harmful

Shah cites the shift regarding pretraining on alignment data—once encouraged, now filtered out to prevent models learning malicious personas or mitigation details.

Companies inevitably abandon unrealistic commitments

Anthropic's Responsible Scaling Policy removed strong 'commit' language in later versions, demonstrating that binding promises get relaxed when impractical.

Conservative language builds more trust than ambitious promises

Google DeepMind deliberately avoids 'commit' language in its Frontier Safety Framework, making it more honest and trustworthy than competitors' stronger rhetoric.

Third-party audits preferred over public commitments

Shah recommends external evaluators with reasonable access to verify practices rather than rigid public promises that may become outdated.

Google's paranoid approach to promises increases credibility

DeepMind's conservative stance on commitments reflects internal skepticism that ensures they only promise what they can actually deliver.