Ex-Google CEO: AI Is Slipping Out of Control

Eric Schmidt warns that advanced AI could escape human control within a few years—first by reaching human-level capability (AGI), then by self-improving into artificial superintelligence (ASI) that no longer “takes orders” from people. Speaking at a summit tied to his Special Competitive Studies Project, Schmidt frames the risk as a timeline problem: once systems can plan and improve themselves, the balance of power could shift quickly, with ASI arriving as AI becomes smarter than humans “put together.” The core implication is stark—human governance may lag behind the speed of capability gains, leaving society unprepared for what comes after human-level intelligence.

The discussion also highlights a mismatch between public expectations and the pace of progress. A poll segment suggests most people do not expect ASI within 3–5 years, while many think a nearer-term milestone—such as landing someone on Mars—may happen first. That contrast underscores how hard it is for mainstream audiences to calibrate risk when AI progress is both visible (better tools, faster outputs) and uneven (incremental improvements rather than dramatic leaps). Schmidt’s remarks about “underhyped” language point to a broader communication gap: there is no widely understood vocabulary for the transition from human-level AI to systems that can outthink and outmaneuver human institutions.

Schmidt’s argument leans on strategic incentives as well as technical capability. He suggests that whoever reaches AGI first would guard it intensely, implying that major labs would not simply release it freely. The transcript also includes a skeptical counterpoint: even if AI keeps improving, today’s systems still behave like highly capable statistical engines—essentially “great search engines” that predict outputs from inputs—so the leap to reliable, self-directed intelligence may not be as immediate as doom scenarios claim. In that view, the biggest near-term problems are practical and engineering-heavy: distributed error, prompt injection, exploitable code, and the difficulty of getting software logically correct in the first place. If AI systems are used to run more of the world, small failures could compound into large-scale breakdowns.

Overall, the exchange lands on a tension between two futures. One future is fast-moving and potentially uncontrollable, driven by self-improvement and strategic secrecy around AGI. The other is messy and error-prone, where today’s statistical systems—already useful but not fully trustworthy—could still cause serious harm through security vulnerabilities and compounding mistakes. Either way, the message that sticks is that capability growth and governance readiness are not aligned, and the consequences—whether catastrophic or merely chaotic—could arrive sooner than society is prepared to handle.