Every time an assistant touches a message between two people — tightening a sentence, drafting a paragraph, composing and sending the whole thing — that exchange is AI-mediated communication. The term covers an enormous range of behavior under one name, which is exactly the problem: two exchanges that both qualify as "AI-mediated" can differ almost completely in how much of what the reader sees actually came from the person whose name is on it.
What the term covers
AI-mediated communication is any exchange between people where an AI assists on one or both ends — drafting, summarizing, translating, or composing the message outright. It does not require the reader to know an assistant was involved. The definition describes a mechanism, not a disclosure: what happened on the way to the recipient, separate from whether anyone said so.
That is broader than most people picture when they hear "AI writing your messages." It includes a spell-checker nudging your phrasing, a translation tool putting your sentence into a language you do not speak, an assistant drafting a first pass that you rewrite completely, and one that drafts and sends with nobody reading it first. All of it is the same category. What differs is where each sits between those extremes.
The spectrum, lightest to heaviest
Laid out in order, roughly by how much of the final message originates with the model rather than the person:
- Grammar and phrasing help. You write every word; the assistant tightens sentences or fixes typos.
- Translation. Your meaning, rendered in language you did not write yourself. The ideas are yours; the words on the page are not.
- Summarizing. The assistant compresses something longer — yours or someone else's. Selection becomes a machine decision: what gets kept and what gets cut.
- Drafting from a prompt. You describe what you want; the assistant produces the actual sentences. You read, edit, and decide what goes out.
- Approving without close review. The assistant drafts, you skim, you send. The words are the model's; the sign-off is increasingly a formality.
- Fully autonomous. The assistant drafts and sends with no per-message review. Nothing about that specific message ever passed in front of the person it is credited to.
Moving down that list, two things happen together: more of the wording comes from the model instead of the person, and less of it gets looked at by a human before the recipient sees it. Those are related but not identical — a system can be heavy on drafting and still keep a real review step, or light on drafting and skip review anyway.
Two questions that matter anywhere on the range
Wherever a given exchange sits, the same two questions decide whether it is honest.
Who does the reader think they are talking to? If the recipient assumes they are reading the sender's own unprompted phrasing, and most of the message actually came from a model, the reader's mental model of the exchange is simply wrong — not because anyone lied, but because nobody corrected the default assumption.
Can they tell which words are whose? Even a reader who knows an assistant was involved somewhere usually cannot tell where. Was this sentence typed by the person, or generated and left unedited? Without an answer, the reader cannot tell which parts to hold the sender to and which are the software's best guess at what they meant.
A heavy, mostly autonomous system can still answer both questions honestly, if it is transparent about the process and labels its output. A light one — a single grammar suggestion — can still get both wrong, if the polish quietly changes what a sentence claims and nobody flags the change.
What keeps it honest
Two mechanisms do the actual work.
Provenance is the record: which specific words came from the person and which from the model, attached to the message itself rather than asserted in a disclaimer. It answers "can they tell which words are whose" directly, part by part, instead of leaving the reader to guess.
Human-in-the-loop design is the gate: a point before anything irreversible where a person actually sees what is about to go out and can stop it. It keeps the heavy end of the spectrum from sliding into "nobody ever looked at this," and it is what makes a provenance label mean something — a label is only trustworthy if a human process, not the model itself, attaches it.
Neither mechanism does much alone. Provenance without a review step tells you who wrote something after it has already been sent. A review step without provenance tells you a human looked at the message, not which parts were theirs to begin with.
The hardest case: two different assistants
Most discussion of AI-mediated communication pictures one person using an assistant to write to another human directly. The harder case is both ends having their own assistant, representing two different people who never share a session. Can two AI assistants talk to each other? covers what changes once there are two separate owners instead of one. That version of the category is also, almost always, asynchronous by construction — a shared session and a correspondence between two private ones are not the same shape of exchange at all.
For the full vocabulary this category runs on — briefing, provenance label, contact pair — the AI correspondence glossary collects it in one place.
Where this leaves you
"AI was involved" is not a useful sentence; it is true of too much to mean anything on its own. The useful questions are where on the range a message sits, whether the reader's assumption about who they are talking to is correct, and whether anyone could check which words are whose if it mattered. Everything else is detail.
If you want those questions answered structurally rather than left to a norm, connect your assistant and look at what a fully labeled exchange actually contains.