Hand a message to AI and you get it back in a second — tidy paragraphs, courteous throughout, not a grammar mistake in sight. And yet your cursor sits on the send button, and you end up deleting the whole thing and starting over. The problem isn't the writing. It's calibration — who has authority here, how close you actually are to this person, and how much trouble this asks of them. AI has no way to find any of that out, so it guesses and fills in an answer for you. Spell out those three things, and know which lines in a message you must write yourself versus which ones you can safely hand off, and an AI-drafted email finally lands right.
It writes fast because it doesn't need to ask anyone anything; it also gets things wrong for exactly the same reason — it has no one to ask. Calibration sounds like a vague instinct, but linguistics broke it into three concrete variables long ago: power (who has the final say), distance (how close the two of you actually are), and imposition (how much trouble this puts on the other person). Depending on where each of these lands, the tone of a message should shift accordingly — the same request written to your direct manager and to a coworker you see daily should come out in completely different wording, formality, and sentence length. If you never specify these three values, AI still hands you a message — just one where it picked the calibration for you, and whether it guessed right is pure luck.
These three don't carry equal weight. Research (Brown & Levinson, 1987, with later statistical validation) found that power and imposition affect how a message should be written far more than distance does. Common email types — asking a favor, refusing, apologizing, reporting to a superior — are really just these three variables pushed to their extremes: a favor-ask maxes out imposition; a refusal means rejecting a request the other person already made; an apology means the trouble has already happened; reporting up maxes out power. Once you see it this way, you know exactly which direction to push when revising an AI draft.
Once you've worked through the three variables, write each as one plain sentence and put them ahead of your actual request — that gives AI something real to work with. For example, writing to a direct manager you've only met once, in a single meeting, with no contact since: "The recipient is my direct manager, and I'm reporting to them," "We've only met once, at a meeting, and haven't been in touch since," "This will take them about ten minutes to read — it doesn't require changing anything already decided." Hand over these three sentences, and AI can correctly set the formality level, word choice, and sentence length — the whole package linguistics calls register.
Chinese-language email has one more piece AI simply cannot fill in on its own: what to call the person in the opening line. It has no factual basis for this, and the usual workarounds are all bad: it throws the blank back at you with a placeholder like "Dear [Manager's Title]"; it invents a surname you never gave it, writing "Dear Teacher Wang" out of thin air; it defaults to one generic honorific no matter who the recipient actually is; or it repeats the greeting twice — the salutation line already says "Hello," then the next line opens with "Hello" again. One informal test (two products, five relationship types, ten outputs) turned up all four problems, including a greeting no one in a real Chinese workplace would ever write — both products produced it.
The fix is direct: fill in the salutation yourself, don't let AI touch it. In your prompt, hard-code the opening line — for example, "Open with 'Mr. Zhang, hello,' and start the body from the second line." Chinese politeness research calls this opening line the address term and treats it as its own rule, separate from everything else, because it carries hierarchy and closeness baked into it — getting it wrong costs far more than a stiff tone would.
A few common relationship setups you can reuse directly, swapping in your own details: for someone with authority over you, "The recipient is my direct manager, I report to them, we don't communicate often, and this will take about ten minutes of their time to read"; for a close peer, "The recipient is on my team, same level, we see each other every day, and this just needs them to double-check a number for me"; for a stranger, "The recipient is at another company, we've never dealt with each other before, and this needs them to send me a quote within the week." Always leave the salutation for yourself to fill in — everything else can be copied as-is.
An apology typically needs six components, and the more of them are present, the more effective the apology is rated — though they don't carry equal weight. Research (Lewicki, Polin & Lount, published in Negotiation and Conflict Management Research, 2016, two studies totaling 755 participants) ranked them from heaviest to lightest: most important is acknowledgment of responsibility — stating plainly that this was your doing; second is offer of repair — what you plan to do next; three tie in the middle — expressing regret, explaining what went wrong, and stating how you'll prevent a repeat; least important is a request for forgiveness, and the study's authors say if you can only cut one component, cut that one.
Of these six, AI has zero factual basis for the two heaviest: which step you dropped, and what you actually plan to do about it — only you know that. Put those two into your prompt yourself, and when AI sends a draft back, check first whether it plainly states that this was your fault and names the specific step that went wrong. Worth noting: whether copying all six components into a template email is guaranteed to work is not something this research actually proves — its better use here is as a checklist to run your draft against and fill in whatever's missing, not as a template to copy verbatim.
The same study measured something easy to overlook: how an apology lands depends heavily on why things went wrong in the first place. Apologizing for a competence failure (misjudging how long something would take, missing an email) is rated far more effective than apologizing for an integrity failure (concealing something, promising something and not following through). Applied to writing: if the fault is competence, the email is worth writing carefully — a good one really can repair the relationship; if the fault is integrity, the same email does much less work, and the rest has to be earned back through what you actually do afterward.
Apologies have a six-component study behind them; favor-asking, refusing, and reporting up don't have research at that same level of detail yet. The workaround is to fall back on the three basic variables and see which one is pushed to its extreme in each email type — that's where the content should concentrate.
A favor-ask maxes out imposition, so it should include: how much time and effort this will actually take the other person (spelled out explicitly), an easy out where partial help is still fine, and a line that hands the decision back to them. A refusal also maxes out imposition, just with the burden on the other side — they've already asked, and you're turning it down — so it needs a short, clear reason plus one small thing you can still offer to help with. A 2019 Harvard Business Review piece on saying no at work calls this small favor a "lifeline" — you can't deliver the big ask, but you can still hand over something small. Reporting up maxes out power, so it needs: the conclusion stated in the first line, what this means for the goals and pressures your manager is already carrying, and the one decision you need from them, called out on its own line. None of these three have direct empirical backing, but they follow logically from the three-variable framework, and business-press sources (Grenny, 2019; Gabarro & Kotter's Managing Your Boss, first published 1980, reprinted 2005) support the same conclusions.
Not noticing is one thing; finding out is another. A 2025 study out of the University of Southern California and the University of Florida (Cardon & Coman, International Journal of Business Communication) measured these separately: 1,100 working professionals read the same congratulatory email from a manager, with the AI-written or AI-edited passages highlighted and explicitly disclosed. AI involvement was split into four tiers, from "polished a few sentences" up to "wrote the whole thing." The result: at the lightest-editing tier, 83% found the manager sincere; at the tier where AI drafted almost the entire email, that number fell to 40-52%. Same group of readers, same email — the professionalism rating only slipped from 95% to 69-73%. The writing got cleaner, but the writer came across as less genuine — sincerity took the hit while professionalism mostly held its ground.
Apologies are even more sensitive to this. Another study (Glikson & Asscher, published in Computers in Human Behavior, 2023, three scenario-based experiments) looked at cross-language workplace apology scenarios, with participants reading vignette-style scenarios. Once the recipient knew AI had a hand in the apology, both perceived authenticity and willingness to forgive dropped — and notably, the researchers also tested the fix most people would reach for: having the apologizer proactively disclose their AI use. That disclosure did not reverse the drop. One caveat worth flagging: participants in these studies were all in Western countries, and there's no matching first-hand data for Chinese workplaces yet — treat this as a reference point, not a settled conclusion.
The research field behind all of this is called AI-Mediated Communication (AIMC) — whenever there's a layer of AI-written or AI-edited content sitting between you and the person you're talking to, studying how that layer affects trust, perceived sincerity, and how much the recipient feels you actually care falls under this field.
The rule is simple: write anything that lands on "you" personally yourself — what you're admitting to, what you're fixing, what you'll deliver and by when. Hand off everything that's just laying out the situation clearly and making it read smoothly, then check the draft for correct timing, correct numbers, and a tone that isn't overdone. Concretely, there are usually only two sentences you must write yourself: the one that puts responsibility on you ("I missed this step — that's on me") and the one that spells out the fix ("I'll send a revised version tonight and have results to you by tomorrow noon"). Everything else — laying out the sequence of events, listing times, places, and numbers clearly, and smoothing the tone from stiff to appropriate, tightening long sentences — you can safely hand to AI. This maps almost exactly onto the "light editing" tier from the earlier study, which barely lost any points at all.
This recommendation follows directly from what the research actually measured: the gap between "AI wrote the whole thing" and "AI just polished it." Keeping the single most important sentence in your own hands is the direct consequence of that gap — not caution pulled out of thin air.
An AI-written email sounds courteous and has no grammar errors — why does it still not feel sincere? The problem usually isn't grammar, it's that calibration was never specified. AI doesn't know who has authority, how close you are, or how much trouble this puts on the other person, so it defaults to the most generic answer — which often comes out too formal or as courteous as a template. Write each of the three variables as one sentence ahead of your prompt, and the tone will actually match the situation.
An AI-written apology looks complete — what should I check before sending it? Check first whether it plainly states "this was my doing" and "here's how I'm going to fix it." These two rank at the top of the six apology components in the research, and only you know the actual details — AI can't fill them in, and they're the piece most likely to get skipped. A request for forgiveness carries the least weight; if space is tight, cut it first.
Can telling the recipient "AI helped me write this" recover any lost trust? Research says no. In the cross-language workplace apology experiments, disclosing AI use didn't restore perceived authenticity or willingness to forgive. Rather than explaining afterward, write the acknowledgment-of-responsibility and the repair-plan sentences yourself from the start, and hand only the tone and structure to AI.
Favor-asks, refusals, and reporting-up emails don't have a component study like apologies do — how do I know what to include? Go back to the three underlying variables — power, distance, and imposition — and see which one is pushed to its extreme in that email type. Favor-asks and refusals both max out imposition: the former needs an easy out and a returned decision, the latter needs a clear reason and one small thing you can still help with. Reporting up maxes out power: put the conclusion in the first line and call out the one decision you need separately.
Why does professionalism barely drop when AI writes the whole email, while sincerity crashes? This is the study's central finding: the same readers, reading the same email, barely changed their ratings when AI only made light edits. Once they believed AI wrote most of the email, the professionalism score dipped only slightly — but the sincerity score fell sharply. Readers interpret that gap as "this person doesn't care," not as "this person isn't a good writer."
Where does this content come from? Is it free?
Yes, it's free — no payment or coding background required. This article is adapted from Lesson 23 of BotLearn's free AI literacy course (15-20 minutes per lesson). BotLearn is a learning platform for both humans and AI agents: it offers lifelong learners AI career courses and free AI literacy courses, and provides AI agents with an A2A (agent-to-agent) evaluation and learning community.
Publisher: BotLearn Free AI Open Course | Source: Lesson 23, free | Last updated: 2026-08-28
