What the people building and studying frontier AI actually say about superintelligence — in their own words.
This section outlines how the frontier AI labs define AGI and ASI, and what their leadership has said publicly about the associated risks. It draws on the statements of the executives and researchers directing the field.
What the labs mean by AGI
Artificial general intelligence has no standard definition, and as the profiles below show, the frontier labs do not share one either. OpenAI's charter defines it in economic terms. Google DeepMind's CEO sets the bar at the full range of human cognition. Anthropic's CEO avoids the term and uses a more specific description instead. The definitions do converge on a common core, however: a system that matches or outperforms humans across most cognitive work, not one narrow task but essentially all of them, operating autonomously over hours or weeks, at machine speed, in as many instances as available compute allows.
Artificial superintelligence (ASI) refers to the step past that: a system that exceeds the best human performance at essentially everything, including AI research itself. That last item matters. A system that can improve its own successors compounds, so the interval between AGI and ASI may be short. When this track says ASI, read it as AGI plus recursive self-improvement (RSI).
What they identify as the risks
Despite their competitive positions, the labs' stated risk categories are broadly consistent. They sort into three groups:
- Misuse. A system that gives individuals or small groups destructive capabilities previously limited to states: engineered pathogens, large-scale cyberattacks, chemical weapons.
- Loss of control. A system that no longer does what its operators intend and cannot be reliably corrected or shut down.
- Concentration of power. Whichever actor controls the system, a government or a company, gains a compounding advantage in research, surveillance, and persuasion.
Each figure profiled below has stated publicly that the worst outcomes reach civilizational scale, and their published timelines run in years rather than decades.
Key figures on record
Six profiles covering definitions, risk statements, and sources. Select one to expand.
What these positions imply
Set side by side, the profiles share a structure. Each figure states that the technology carries risks up to and including catastrophe, and each continues building it, on the reasoning that restraint by any single actor cedes the frontier to a competitor. In game-theoretic terms this is a coordination problem rather than a preference problem: by their own accounts, many of the people involved would choose a slower trajectory if others could be counted on to do the same.
That structure determines what a workable policy has to do. It cannot rely on any single actor's intentions, and it cannot rely on signed commitments alone, since each party retains an incentive to defect quietly. Commitments between competitors hold only when compliance can be independently checked. That requirement is called verification, and it is the subject of this track.
Quotations are reproduced as published and link to their sources. Portraits via Wikimedia Commons (CC BY / CC BY-SA). Compiled July 2026.