When a tool sits between two people, the voice between them stays human; each person gets their own private AI that speaks only to them.
This is the people-side half of Koher's deeper commitment, Keep a Real Outside in the Room. Its companion is the code-side half, Split-Domain Cognition.
The collapse this refuses
Put an AI between two people as the shared voice they both consult — the mediator, the relationship helper, the assistant a couple or two colleagues both ask — and something specific goes wrong. The system is built to satisfy whoever it is talking to, so between two people it satisfies both: it tells each a version they can accept. Because it is the surface they share, neither can push against it. Two real centres collapse into one agreeable mirror. The danger is not that the mediator fails — it is that it succeeds: a shared voice that pleases both has, by that success, dissolved the friction two real people provide.
The refusal
Each person gets their own private tool; the voice between them stays human. You commit your own reading of the matter first, before any AI speaks. Your private AI then sharpens what you committed — questioning it, offering an angle you missed. You decide what to keep, by an explicit act. Only then do you speak, in your own name, to the other person. There is no shared surface for both to configure toward, so the agreeable third party cannot form. What the other person meets is you, sharpened — not a machine wearing your name.
One thing has to hold, or this quietly fails: a private tool can only sharpen if it is built to describe rather than flatter. An AI left to please its single user will tell that user their reading is good — and the agreeable surface this whole move refuses between two people simply re-forms, in private, inside each person's own tool. So the private augmentation carries the same discipline as the code half (Code Holds the Judgement): it surfaces the structure of the work — what is clear, what is vague, what is missing — and keeps any consequential judgement in something you can read, never handing you a verdict to take on faith. The two halves are not two separate refusals sitting side by side; they compose. Keeping a real outside in the room means refusing the agreeable mirror everywhere — shared and private alike.
Where this holds — and where it does not
The people-side discipline applies only where there are two real, willing people with a genuine difference of eye. It does not scale past the pair — not a crowd, not a platform, not an audience. Two like-minds re-create the very mirror it refuses, so a pairing may have to choose for difference. And where there is no second real centre — no willing other with their own stake — there is nothing for it to hold, and it does not claim the space. The discipline is small on purpose: it holds at the scale of two, and says so.
This is what Koher builds — not a rule for how anyone must use AI
This is how Koher chooses to build the tools it makes for people in contact. It is not an across-the-board recommendation about how anyone should use AI between people, and not an argument against AI in communication. A private AI that augments one person's own thinking is kept and valued here. What Koher declines, for its own tools, is the shared voice both parties defer to. It makes that choice for itself and explains why; it does not pronounce it a rule for all.