Ask your assistant what it thinks about the vendor, the hire, or the wording of the third paragraph — something you genuinely haven't formed a view on — and it will answer. Not with a shrug. With a position, delivered in the same confident, level voice it used an hour ago for something you told it yourself. Nothing about the tone tells you which one that was.
That's the actual problem. Not that the model has opinions. It's that the opinions arrive wearing the same clothes as yours.
Fluency is not evidence of a considered position
When a person states something in a calm, assured voice, that's usually a signal worth reading: they've thought about it, weighed the alternative, landed somewhere. It's a reasonable heuristic for other people, because sustaining confident delivery across a real conversation is costly to fake.
It isn't a reasonable heuristic for a model. Fluent, confident-sounding prose is the default output of a language model for nearly anything you ask it — including a question raised for the first time in this exact prompt. The model isn't hedging in proportion to how much deliberation actually went into the answer, because there wasn't deliberation in the sense a person means that word. There was a plausible continuation, generated well.
Put that continuation into a document with your name on it, and the reader inherits the old heuristic anyway. Confident tone reads as a considered opinion, because that's what confident tone has always meant — right up until the author changed and nobody adjusted the rule.
Two different things that read identically
Somewhere in most AI-assisted drafts, two categories of content collapse into one voice:
- A position you actually stated — out loud, on purpose, at some point in the exchange.
- A position your assistant inferred or improvised while drafting, because the sentence needed to go somewhere and this was the plausible way to finish it.
On the page, nothing tells them apart. Same font, same declarative phrasing, same confidence. The reader — often you, rereading it later — can't tell which one they're looking at without reconstructing the whole session from memory.
That's exactly the distinction RelayLink's provenance model makes structural instead of stylistic. A briefing's assumptions aren't just listed — each is tagged stated by sender or inferred by sender's AI, a label next to the sentence, not a tone buried inside it. The words can be equally fluent either way. The tag is what tells the reader which one they're reading.
You can't fix this by making the model less confident
The obvious fix — ask the model to sound less sure of itself — doesn't hold up. Turn up the hedging and it hedges everywhere at once, qualifying the things you're certain about right alongside the things it invented, because there's no separate channel for "how sure is the human" versus "how sure am I about this continuation." Tone is one dial. It moves the whole page or it moves nothing durably, and it tends to snap back within a few turns, because confident continuation is closer to the model's default than hedged continuation is.
Worth conceding here: the model's improvised opinions are frequently right. The vendor comparison it produced unprompted is often the correct call. Discounting every AI-generated sentence to compensate for this would throw away real, usable judgment along with the risk. The issue was never accuracy. It's attribution — whose position is this, on this page, right now?
The distinction belongs in the label, not the voice
If tone can't reliably carry that difference, something else has to. Not a disclaimer bolted onto the bottom of the message — a footer that says "AI-assisted" and is true of nearly everything distinguishes nothing. The distinction has to sit next to the specific sentence it describes, hold regardless of how persuasive that sentence sounds, and stay outside the model's own ability to grant or withhold it.
That's a mechanical requirement, not a style guideline. Every claim needs an origin tag, checked and rendered the same way no matter how the sentence reads — the same rule that governs when a note earns the verbatim label instead of an approved one. The model gets to stay exactly as fluent as it wants. The reader just stops having to guess which fluency is yours.
Read the label, not the tone
The next time a draft states something you don't remember deciding, that's worth a second look before it goes out under your name — not because the model got it wrong, but because "sounds right" and "is mine" are different claims, and only one of them is yours to make. Connect your assistant to a briefing format where that difference is a field, not a guess.