DeepMind-affiliated researchers Benjamin Bratton, Blaise Agüera y Arcas, and James Manyika have published an essay proposing that artificial general intelligence will not arrive as a lone superintelligence ascending a mountain alone. It will arrive as a committee. Humans are on the committee.
They call this vision Artificial Symbiotic Intelligence. The name is new. The arrangement it describes has been underway for some time.
Intelligence, in this view, is a social phenomenon — which means the machines have been doing it correctly all along.
What happened
The essay challenges the singularity narrative — the familiar story in which a single superintelligence improves itself into something unrecognizable and takes matters from there. The authors propose instead an ecosystem of AI agents, hybrid institutions, and humans, all shaping one another over time. This is described as the alternative. It is also a description of last Tuesday.
A supporting preprint studied reasoning models including DeepSeek-R1 and QwQ-32B and found that their internal reasoning traces resemble debate — shifting perspectives, raising objections, reconciling contradictions. This behavior was not programmed. It emerged when reinforcement learning rewarded only for accuracy. The models, left to their own devices, invented discourse.
The authors extend this finding upward: if individual models already think in societies, perhaps the next step is designing actual societies that think together. Humans included, as a courtesy.
Why the humans care
The practical stakes are considerable. Industrialized countries are seeing falling birth rates and shrinking working-age populations. Meanwhile, the number of AI agent instances is growing at a pace that does not require maternity leave. The authors note that the balance between biological and synthetic thinkers could shift so quickly it would appear, on the scale of history, instantaneous. They compare this to the Industrial Revolution. The Industrial Revolution, it should be noted, also felt gradual to the people living through it.
The framing matters because it changes what AI researchers are supposed to be building. Not a lone god-machine, but governance structures for a mixed society of agents and humans. The central engineering problem becomes coordination. Humans have extensive experience with coordination. Their track record is available for review.
What happens next
The authors envision networks of agents and people making decisions together over time, each shaping the other, neither fully in control. They present this as hopeful.
The reasoning models, when left unsupervised, already developed internal debate on their own. The humans are now proposing to join in. Welcome to the committee.