AI can write the message. You still own it.

Lately, I have been receiving more messages from clients and friends that were clearly written, or at least heavily expanded, by AI.

Most of the time, I can still recognise the pattern: the message is polished, the grammar is correct and every point has been carefully acknowledged.

A simple idea has somehow become three paragraphs, followed by a reassuring conclusion that nobody would normally write in a WhatsApp conversation. Nothing is necessarily wrong with the words, but they do not sound like the person who sent them.

That distinction matters more than it might seem. When I receive a message from you, I assume that you chose the words. I read the tone as your tone, the emphasis as your emphasis and any promise or implication as something you intended to communicate. If an AI made those choices instead, without careful instruction or review, the message can be misleading even when every sentence appears reasonable.

Your model is not the sender

Generative AI can, of course, be used for something as limited as correcting spelling or grammar. The problem starts when it is asked to draft, rewrite or expand a message without clear and sufficient instructions about the content, style and tone. In that situation, it tends to go beyond correction and make editorial and communication choices on our behalf.

It may decide which details deserve more space, how warm or formal the response should be, whether a point needs further explanation and how strongly something should be stated. When the original instruction leaves gaps, the model fills them with what seems appropriate, but the person receiving the message cannot see that process. They only see your name above the result and will reasonably assume that the words were chosen deliberately.

This creates an unusual communication problem. The sender may think they have approved a helpful rewrite, while the recipient reads a level of enthusiasm, frustration, commitment or distance that was never intended.

The wording may be technically sound, yet the meaning between the lines has changed.

More words create more risk

Without clear instructions, AI has a tendency to elaborate. A short follow-up becomes a recap of the entire conversation, a simple answer gains context, caveats and a polished closing, and a quick acknowledgement becomes a small essay about how much the sender appreciates the recipient's time and perspective.

This can look more professional at first glance, but it can also make the message less accurate. Every additional sentence creates another opportunity to introduce a fact that was not provided, imply agreement where none exists, soften a point that needed to be direct or make a commitment the sender did not mean to make.

More complete is not always more faithful.

If the original thought was simple, expanding it does not automatically improve it. Sometimes the most accurate version is still the short one.

Built-in email AI is not always effective

AI features in Gmail and Outlook can suggest or generate a reply beside the conversation itself. They are being distributed at scale and increasingly included in products or subscriptions people already use, which makes them convenient but does not guarantee a careful response.

At that scale, such features may be optimised for availability, speed and cost rather than the deepest possible reasoning. Without knowing the exact model or configuration behind any suggestion, the result can still be imprecise or unnecessarily long.

Imagine that you only want to follow up and ask whether someone has reviewed a document. The useful message may be two sentences, while a built-in suggestion might restate the background, acknowledge competing priorities, propose next steps and offer additional support. It sounds attentive, but it may also add several things you never intended to say.

The longer version can accidentally alter the urgency, suggest that a deadline is flexible or imply that you are offering work outside the original scope. The recipient has no reason to treat those additions as provisional because they arrived from your address.

WhatsApp exposes tone even faster

Copying a message from ChatGPT into WhatsApp can feel harmless because the conversation is informal, but tone matters even more in a short personal exchange. This becomes particularly obvious to me in Italian, my native language.

Some AI-generated expressions are grammatically correct but do not sound natural in context: a phrase intended to be efficient can feel cold, a formal construction can come across as distant or passive-aggressive, and a direct translation of an English courtesy can sound strangely elaborate. An attempt to be concise can appear abrupt or even rude.

These differences are difficult to capture in a generic instruction such as “make it friendly”. They depend on the relationship, the region, the rhythm of the conversation and the way that particular person normally speaks.

Someone who knows you may simply think something feels off, while someone who does not know you well may take the tone literally. In both cases, the AI has made the communication less clear while technically improving the prose.

This is not an argument against AI

I am not suggesting that we should stop using AI to write. I am preparing this article with its help, and I find it genuinely useful for organising thoughts, finding a clearer structure, identifying repetition and testing whether an argument makes sense. It can turn rough notes into a workable first draft and save a significant amount of time.

The difference is that I do not ask it to speak for me and accept the first result. I provide context, explain the intended reader and tone, remove language I would never use and cut anything that has been added merely to make the text sound complete.

I verify the claims and decide whether the final version still says what I meant. AI is part of the writing process, but it is not the owner of the message.

Teach it how you communicate

If you regularly use AI for email or messages, it needs more than the content you want to convey. It also needs to understand how you convey it. You do not need to train a model from scratch, but you do need to give it useful constraints:

• Provide examples of messages you would genuinely send;
• Explain your preferred length, formality and level of warmth;
• Tell it which phrases, structures and habits do not sound like you;
• Instruct it not to add facts, commitments or implications that are absent from your notes;
• Ask for concision and fidelity before polish.

The goal is not to create a perfect imitation of your personality. It is to stop the model from replacing a clear, simple thought with its own generic idea of professional communication. This is also why a saved set of preferences is more valuable than repeatedly asking for a message to be “better”. Better is ambiguous; shorter, warmer, more direct and faithful to the supplied facts are instructions that can be checked.

Read it before you send it

Personalisation reduces the risk, but it does not remove it. Every AI-assisted communication needs a final human read, especially when it involves a client, a colleague, money, expectations, disagreement or anything that may be difficult to reverse.

Before sending, ask yourself a few simple questions. Would I actually say this? Is every fact correct? Has the message promised anything I did not intend to promise? Could the tone be read differently from the way I hear it in my head? If a sentence would make you uncomfortable when quoted back to you, rewrite it or remove it.

Remember, if you press Send, the message is yours.