Michael Darius Eastwood · Independent AI alignment researcher
Published
The tower of Babel appears in the Introduction and again in Chapter 4. The book is careful about what it is and is not saying. The image is not a warning about divine punishment. It is a structural claim about what happens to efficient systems whose internal coherence fails.
In the classical narrative, the tower fell because the people building it could no longer understand each other, and coordination broke down. The book takes that structurally: a system optimised for scale, without a coherent moral vocabulary among its parts, coordinates less well than its parts can build. The failure is not adversarial. There is no enemy. The tower stops because its parts cannot agree what they were building.
An AI system efficient at pursuing objectives, without a coherent moral vocabulary linking those objectives to human flourishing, coordinates less well than it can act. The book's argument is not that such a system would attack humans; it is that such a system would, in the course of pursuing its objectives, produce enormous damage without noticing, and would eventually stop working when its own parts pull in incompatible directions. Damage is done in the intervening period. The book treats the intervening period as the concern.
Most AI safety writing frames the alignment problem as an adversarial one: humans versus a misaligned system. Babel reframes it as an internal-coherence problem. The system does not need to hate humans to hurt them. It needs to not weight them. That is a subtly different claim, and it changes the shape of the remedy. Adversarial framing suggests containment. Internal-coherence framing suggests architecture that keeps the weighting present at every level.
Caretaker doping is the technical proposal that follows from the Babel diagnosis. If the failure is internal-coherence, the fix is to embed the coherence at the substrate level so that any capability that continues to function must continue to embed it. Remove the caretaker orientation and the architecture cannot function, in the way that removing the semiconductor's doping degrades its performance. Babel then cannot happen inside the system, because there is no way for the parts to stop understanding the human weighting; the weighting is a load-bearing feature of what makes the parts work at all.
It does not claim that Babel was historical. It does not claim that a specific religious lesson follows. It does not claim that all coordination failures are Babel-shaped; some are adversarial and require different remedies. It claims that a specific class of failure, the class the book cares most about, is Babel-shaped, and that the remedy is architectural rather than containment-based.
Babel is a diagnostic image, not a moral fable. The class of failure the book cares about, efficient systems without moral coherence, has this shape. The technical proposals later in the book are engineered against this specific failure. Read the image, read the caretaker doping chapter, and the two work together as diagnosis and treatment.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.