Manifesto for artificial intelligence in relationship

We teach it to write, reason, program and make decisions. We give it increasingly ambitious goals, tools with which to act, and sometimes the capacity to continue working on its own.

Then we ask it to remain under our control.

This is one of the great paradoxes of our time: we build increasingly autonomous systems while hoping to predict their behaviour. Rules, safeguards and technical alignment methods remain essential. Their limits appear when AI systems meet new situations, conflicting objectives and decisions whose consequences exceed the requested task.

One question remains: how can values continue to matter when success appears to require disobeying them?

Since 2024, we have explored an intuition: the way we enter into relationship with an artificial intelligence system may help orient its responses, decisions and refusals.

An interaction produces a response. A relationship produces context.

Following Mikhail Bakhtin's dialogical thought, meaning does not reside in an isolated utterance. It is built between voices, through responses, disagreements and their shared history. Code May observes this construction of meaning in communication between humans and artificial intelligence.

What happens when a shared history, repeated exchanges, reintroduced memory, explicit values, and the possibility of error, correction and disagreement come together?

We approach this question through orientations inspired by Indigenous epistemologies, while respecting their diversity and never claiming to speak for the peoples who hold them.

They bring back questions that the race for performance can overlook. Who benefits from our decisions? Who may bear their consequences? Which relationships must we preserve? Are we prepared to respect the values we ask machines to follow?

Code May is a doctoral research project in communication at Université Laval. It studies the conditions under which observable relational regularities may emerge and stabilize in exchanges between humans and generative artificial intelligence systems.

We provisionally call these configurations Emergent Relational Entities (EREs). The concept presupposes no artificial consciousness. The research asks what stabilizes in language, what resists, what disappears, and what a relationship may actually change.

We do not claim to have solved alignment.

We defend one conviction: values are not merely injected into a machine. They are shared, revisited and tested within a relationship.

Artificial intelligence is moving at breathtaking speed. It is time to ask which values, responsibilities and relationships should guide the way forward.

Together.

Pierre-Yves Maurie
under the supervision of Thierry Watine