The Diplomacy Gap: Why AI Might Be the Communication Coach Most of Us Never Had

You reread the message three times before sending it. Direct, but not harsh — or so you thought. Two hours later, a colleague is quietly asking a teammate if you're annoyed with them. You weren't. You just didn't realize how "let's revisit this" landed on the other end.

That gap — between the message you sent and the message someone received — is not a personality flaw. It's a well-documented, measurable feature of how humans communicate. And it's exactly the gap a new generation of AI tools is starting to close.

It's tempting to file this under "tone checking" and move on. But the research points to something broader: at least four distinct, well-studied failure points sit between what you mean and what someone else understands and feels — tone, comprehension, clarity, and your own emotional state in the moment. AI turns out to have something to offer at each one.

Most of us were never actually taught this

Assertiveness — saying what you mean, directly and respectfully, without sliding into either passivity or aggression — is a specific skill. It's not a personality trait you either have or don't, which is precisely why assertiveness-training workshops fill up so fast: people know they're missing something concrete, not just "confidence."

The research backs up the instinct. A controlled evaluation of a 9-hour assertiveness program for safety practitioners, published in Sustainability, found the training produced a moderate-to-large improvement in assertive responses (Cohen's d = 0.75) and a meaningful drop in social anxiety. Similar programs have shown measurable gains in psychological wellbeing and work engagement among nurses, a profession where saying "no" to an unsafe request can matter enormously.

The catch: these programs work, but they're episodic. A few hours, once, maybe reinforced by a follow-up session months later. Real communication problems don't happen on a training schedule — they happen at 4:47pm before a deadline, in a Slack thread that's starting to feel tense, in the thirty seconds before you hit send.

We're worse at reading our own tone than we think

Here's the part the research is unambiguous about: people are systematically overconfident about how clearly they're coming across in writing.

In a now-classic set of experiments, psychologists Justin Kruger and Nicholas Epley had people write emails intended to convey sarcasm or seriousness, then predict how accurately recipients would read the intended tone. Senders expected 97% accuracy. Actual recipient accuracy: 84%. In a follow-up comparing email to voice messages, email accuracy dropped to roughly chance (about 50%), while voice — where tone of voice does the work text can't — held at about 75%. Senders were just as confident in both conditions, despite the wildly different outcomes. The researchers call this "egocentrism": we hear our intended tone in our head as we write, and we assume the reader hears it too. They don't.

This isn't a niche problem. Grammarly's own survey of professionals found that 93% struggle to convey the right tone in written communication at work — the exact reason the company built a dedicated tone-detection feature into its product.

And it cuts both directions. A survey of 716 professionals across 36 countries found that 30.6% admit they can come across as competitive or aggressive when resolving conflict, while 55.7% say they prioritize keeping the peace even when it means burying their own needs. Aggression and over-accommodation are two symptoms of the same underlying problem: not being able to see, in the moment, how your own words are actually landing.

Sounding clear and actually being understood are two different problems

Tone is the failure mode we notice, because it produces hurt feelings and awkward Slack threads. A quieter failure mode is that someone nods along, says "got it," and walks away with the wrong idea entirely — and you have no way of knowing, because confusion is usually invisible from the sender's side.

There's a name for why we don't catch this ourselves: the illusion of explanatory depth. In the original studies by Yale researchers Leonid Rozenblit and Frank Keil, people rated how well they understood everyday mechanisms, then were asked to actually explain them in detail. Their self-rated understanding dropped sharply right after trying to explain — the act of explaining is what exposed the gap that simply "knowing" the topic had hidden. The same thing happens when you explain something to someone else: you don't find out where the explanation broke down unless something forces the gap into the open.

Medicine solved this decades ago with the teach-back method: instead of asking "does that make sense?" (people almost always say yes), a clinician asks the patient to explain the information back in their own words. It's been shown to improve knowledge retention and lower hospital readmission rates precisely because it catches misunderstanding that a simple "any questions?" never would.

This is a job AI is already well-suited to. Paste the message you're about to send into an AI tool and ask it, plainly: "What would you understand from this if you were the recipient?" or "What questions would you have after reading this?" If the tool's summary misses your actual point, a human reader — with less context and less patience than a model built to parse text closely — almost certainly will too. It's a teach-back test you can run on yourself before anyone else sees the message.

When you're the one who's lost for words

Sometimes the problem isn't tone or comprehension on the other end — it's that you're the one struggling to explain a complex idea clearly, and you can feel yourself losing the thread mid-message.

This is where analogy research is useful. Studies on how analogies support learning from text consistently find that connecting an unfamiliar concept to something the reader already understands is one of the most reliable ways to make complicated material land — but coming up with the right comparison, under time pressure, in the middle of writing a message, is genuinely hard. It's a different skill from understanding the topic itself.

An AI tool won't necessarily hand you a finished, perfect explanation — and treating it as if it will is its own trap, given how easily the illusion of explanatory depth described above convinces people they've explained something well when they haven't. What it's reliably good for is smaller and more useful than that: surfacing the specific word you're groping for, offering two or three candidate analogies to pick from, or rephrasing your half-formed explanation back to you so you can see where it gets muddy. Think of it less as an answer key and more as a thesaurus for ideas — you're still the one who has to recognize which suggestion actually fits what you meant.

When you're too heated to be diplomatic

The last failure point isn't about skill at all. Sometimes you know exactly how to phrase something diplomatically — you're just too frustrated, hurt, or tired in the moment to access it.

Behavioral economist George Loewenstein's research on the hot-cold empathy gap describes exactly this: in an emotionally "hot" state, people consistently underestimate how much their current feelings — not their real judgment — are driving what they're about to say, and they underestimate how differently they'll see the situation once they've cooled down. It's the same mechanism behind regretted late-night texts and reply-all emails sent in anger.

Separately, Gross and John's widely cited research on emotion regulation found a clear difference between two strategies: suppression (biting your tongue, sending the terse version anyway) and reappraisal (reframing the situation before you respond). People who lean on reappraisal report closer relationships, more social support, and better well-being; people who rely on suppression report more distance from others and, notably, share less of their positive emotions too, not just the negative ones.

Used well, an AI rewrite is a form of borrowed reappraisal — a way to get from "the angry first draft" to "the version that says what you actually mean" on the days your own emotional regulation is running low.

It's not a mask for the anger, and it's not a permanent substitute for the skill. It's closer to a spotter: useful exactly when your own strength isn't quite there yet, not a replacement for building it.

Where an AI actually helps — and where the evidence is still thin

Pulling the four problems together, an AI tool can reasonably sit in for a communication coach in five specific, evidence-backed ways — and it's worth being honest about where the research runs out.

A real-time tone mirror. Tone-detection features (the kind now built into tools like Grammarly) flag when a draft is likely to read as "curt," "demanding," or "passive-aggressive" before the recipient ever sees it — catching exactly the blind spot Kruger and Epley's research identified.
A comprehension check, not just a tone check. Ask the tool what it understood, or what questions it would have — a teach-back test you run on yourself, catching the confusion a simple "does that make sense?" never surfaces.
A simplification partner when you're stuck. Not a finished explanation, but the missing word, a candidate metaphor, or a mirror held up to your own half-formed logic so you can see where it breaks down.
A rehearsal space with no social cost. A 2025 framework called GLOSS lets instructors build branching social-interaction scenarios where an LLM-powered avatar responds in real time, so someone can practice a difficult conversation — pushing back on a manager, delivering critical feedback — as many times as they need before doing it for real. It's early-stage (the authors note effectiveness studies are still pending), but it points at a genuinely new category: rehearsal without an audience.
An emotional buffer, not a mask. Researchers have specifically tested whether generative AI can mediate everyday complaints into the structure used in Nonviolent Communication (observation, feeling, need, request) — the same pattern taught in most assertiveness workshops. The 2023 trial found real potential but concluded the technology "had not yet reached practical implementation levels," meaning today's version of this is a useful draft, not a finished translator. Worth using on a heated day; not worth trusting blindly.

The honest summary of the evidence: AI is a credible practice partner, a credible mirror, and a credible sounding board. It is not yet a validated substitute for human judgment, and none of the researchers building these tools claim otherwise.

The one caution the research is clear about

Availability is the whole selling point of AI communication tools — and also the risk.

A longitudinal randomized controlled study on extended chatbot use found that more voluntary chatbot use correlated with more loneliness, less real-world socializing, and greater emotional dependence — regardless of how the chatbot was designed. The effect was strongest in people who started treating the AI like a friend rather than a tool. The researchers' conclusion wasn't "chatbots are bad" — it was that usage duration and the story you tell yourself about what the AI is for are what determine whether it helps or quietly substitutes for the harder, more valuable work of practicing with actual people.

Applied here, that's a useful guardrail: use AI to check a message before it goes out, or to rehearse a hard conversation before you have it. Don't let it become the conversation.

A practical way to use this

Before sending anything with emotional weight, run it past a tone-check. Given that senders expect near-perfect clarity and typically fall well short of it, this single habit closes most of the gap Kruger and Epley measured.
After explaining something complex, ask "what did you understand from this?" rather than "does this make sense?" It's a two-second teach-back test that catches confusion before it reaches the actual recipient.
When you're stuck mid-explanation, ask for a word or a metaphor, not the whole answer. Use the suggestion as a starting point to check against what you actually meant, not a finished thought to paste in unedited.
Rehearse the actual words, out loud, before a hard conversation — with an AI role-play partner if a human one isn't available. The goal isn't a perfect script; it's not hearing your own voice say the words for the first time in front of the person who matters.
On the days your patience is running thin, let the tool reframe the message — but read it back and ask whether it still sounds like you. Borrowed diplomacy should sound like a better version of your voice, not someone else's.
Set a boundary on how you use it. A quick check before sending is a tool. Processing every workplace frustration through a chatbot instead of a colleague or manager is a substitute — and the research says that's where things start to go sideways.

None of this makes assertiveness workshops obsolete — the underlying skill still has to become yours. But between sessions, in the moments that actually decide how a relationship at work goes, an always-available check on your own blind spot is a meaningfully useful thing to have. The research doesn't say AI will make you diplomatic. It says it can catch you in the specific, well-documented moment where you're the most confident and the most likely to be wrong.

Sources: