Head to head, honestly

The Eden Protocol vs Constitutional AI

Each is described by what its own text states. They differ on where values should live, and they differ enormously in maturity. This page credits the prior work first, states the difference in kind, and names what would decide between them.

The shared ground, and the prior work

Constitutional AI is Anthropic's method, published in 2022: a set of written principles guides the model to critique and revise its own outputs during training, so the values are trained into the model. Its abstract states control of behaviour as its purpose, 'to control AI behavior more precisely and with far fewer human labels'. It is prior work on this site's record, credited for what its own abstract states. The Eden Protocol's claim is over its specific dated formulation and named architecture, anchored in the 8 December 2024 manuscript and standing on its own: ethical principles at a system's foundations, from its genesis, holding in every recursive loop and every decision as the system evolves. No originality over ethics built into an architecture as a category is claimed; that is an older category, stated in the Friendly AI writings (Hibbard 2001; Yudkowsky, 15 June 2001; Bostrom 2003) and developed in the machine-ethics literature (Arkin, ACM/IEEE HRI, 12 March 2008; Wallach and Allen, Moral Machines, Oxford University Press, 2009).

The difference in kind

Constitutional AI shapes the training objective: principles, critique and revision produce a model whose behaviour reflects the constitution. The Eden Protocol proposes something structurally stronger and far less proven: that care should be entangled with capability itself, so that removing the alignment damages the system. Its four layers are a purpose kernel, graduated autonomy that is earned and revocable, a monitoring-removal test that checks whether behaviour survives the absence of oversight, and an entanglement test in which safety and capability are trained as a product.

Maturity, stated without flattery

One of these is running in production systems used by millions. The other is at pilot scale in one laboratory: the entanglement mechanism is demonstrated on toy networks, the weight-level removal test did not confirm structural entanglement at scale, and the controlled pilot found Eden matched its unaligned comparison on capability within noise, with fewer reward hacks. Those results are encouraging and small, no outside laboratory has replicated them, and the programme says so on every surface that reports them.

What would decide it

For Constitutional AI, the open question is durability under self-improvement: whether trained-in values survive when the system can modify the machinery that carries them. For Eden, the open question is existence: whether capability-entangled care can be built at scale at all. A frontier-scale system whose alignment demonstrably survives the removal of oversight would be evidence for embedding in general; which architecture gets there is an empirical race, and this programme's entry is the smaller and less proven one.

Sources: the embedded-correction priority record; the Eden Protocol page; Papers VI and VIII; the evidence spine.

All comparisons · Eden Protocol · The embedded-correction thesis