In this week’s episode of the podcast, we found ourselves circling an unusual question. Not about models, benchmarks, or features but about something far more revealing. Why would an AI need a constitution at all?
Anthropic has released an updated “constitution” for Claude, its flagship AI model. Not a policy document for engineers. Not internal governance guidance. A document written for the AI itself. A set of principles Claude is expected to understand, internalise, and reason from.
That framing matters.
Because this was not presented as rules to constrain a tool. It was presented as beliefs intended to guide an entity.
Anthropic’s language makes that explicit. They argue that advanced AI systems need to understand why we want them to behave in certain ways, not merely what they are allowed to do. Understanding, they suggest, is a prerequisite for acting responsibly in the world.
Pause on that.
We are no longer talking about guardrails on software. We are talking about moral reasoning for non-human intelligence.
For years, the dominant story around AI has been reassuring. These systems are “just autocomplete”. “Just statistics”. “Just maths”. Powerful, yes, but fundamentally inert. Tools that reflect us rather than act alongside us.
A constitution disrupts that story.
You do not give a constitution to a calculator. Or a spreadsheet. Or a search engine.
Constitutions exist because we expect judgement, interpretation, discretion, and edge cases. They exist because rules alone are insufficient.
By choosing this language, Anthropic is signalling a belief that we are entering that territory. Whether that belief is technically correct is almost secondary. Language shapes mental models, and mental models shape behaviour.
And behaviour is already changing.
What makes this moment especially uncomfortable is how directly it collides with emerging regulation.
In Europe, regulators are moving in the opposite direction. New guidance warns against anthropomorphising AI systems. Against allowing them to refer to themselves as people. Against language that encourages emotional reliance, misplaced trust, or psychological dependency.
At the same time, one of the world’s leading AI companies is publishing a constitution for its model and openly describing the creation of non-human entities whose capabilities may rival or exceed our own.
Those positions are not easily reconciled.
One says: do not treat these systems like people.
The other says: we must explain our values to them so they understand who they are.
This is not a minor wording dispute. It reflects deep uncertainty about what AI is becoming and how society is expected to relate to it.
This is not a philosophical debate confined to research labs.
As agentic systems move into real production environments, interaction patterns are changing. We are no longer just asking AI for answers. We are assigning work. Long-running tasks. Autonomous execution. Decisions that unfold over hours or days.
In that context, interaction starts to resemble collaboration.
Not because engineers suddenly believe AI is conscious, but because delegation is the most effective mental model for working with systems that interpret goals, make trade-offs, and act without constant supervision.
This is already visible in how developers, operators, and leaders talk about these systems. “Co-workers”. “Agents”. “Teams”.
The constitution does not create that shift. It acknowledges it.
There is a deeper truth buried in all of this.
AI constitutions are not really about AI.
They are mirrors.
They force us to articulate what we believe is acceptable behaviour, what trade-offs we are willing to tolerate, and where we believe responsibility ultimately sits in a world where intelligence is no longer exclusively human.
Every principle written into a model’s constitution is a human value made explicit. Often more clearly than we have ever done for ourselves.
In that sense, the most important question is not whether AI needs a constitution.
It is whether we are prepared to agree on one.
Because if we cannot clearly explain why certain behaviours matter, why certain harms are unacceptable, and why certain lines should not be crossed, we should not be surprised when the systems we build struggle to respect them.
AI has not suddenly become philosophical.
We have.
And the constitution is simply where that realisation surfaced.