Anthropic’s Constitutional AI Is Now a Template Other Labs Are Copying

When Anthropic published its Constitutional AI paper in 2022, the reception was mixed. Three years later, every major AI lab has some version of constitutional or principle-based training in their published methodology.

Google’s Gemini documentation describes “model constitutions.” Meta’s Llama 3 training documentation references behavioural guidelines that bear the structural signature of constitutional approaches. Even OpenAI has published work on “model specification” that closely parallels Anthropic’s constitutional framing.

The spread of the technique doesn’t mean the technique is solved. Constitutional AI still struggles with value conflicts — situations where two principles pull in opposite directions — and with cultures where the principles as written don’t translate cleanly.

But the framework has won. The debate is now about which constitution, not whether to have one.