Making AI writing sound less formal and more like you comes down to three moves: feed the model real samples of your own writing, give it explicit tone instructions instead of hoping it guesses, and then edit out the stock phrases it defaults to.
The AI has no idea who you are unless you show it, so the fix is mostly about supplying better input and cleaning up the output, not about finding a magic prompt.
Why AI writing drifts formal in the first place
Language models are trained on a huge pile of text, and a lot of that text is corporate, academic, and journalistic. So when you give a vague instruction like "write a blog post about pricing," the model reaches for the safest, most average register it knows: complete sentences, no contractions, phrases like "it is important to note."
It is not trying to sound stiff. It is trying to sound correct, and correct in its training data often means formal. The model also has no memory of your voice between sessions unless you put that voice into the prompt.
That is the mechanism behind the problem, and it tells you where the fix lives: the input you give and the editing you do afterward.
Give the model samples of your own writing
The single highest-leverage technique is to paste two or three paragraphs you actually wrote into the prompt and say, "Match this voice." Pick samples where you sound like yourself, not the most polished thing you have ever produced. If you write casually in emails, use an email.
If you are funny in Slack, use a Slack message. Then add concrete instructions: "Use contractions. Keep sentences under 20 words most of the time.
Write at a ninth-grade reading level. Avoid the words 'moreover,' 'furthermore,' and 'delve.'" Specific bans work better than vague requests. Telling a model to "be conversational" gives it almost nothing to work with, while telling it to cut a named list of phrases gives it a clear target.
A worked example
Say you want a short intro for a newsletter about a project management tool. A vague prompt gets you something like: "In today's fast-paced business environment, effective project management is essential for teams seeking to maximize productivity." That is the default register.
Now try again with a sample of your own writing attached and instructions: "Write like this sample. Start with a short sentence. Use 'you.'
No corporate phrases." You might get: "Most teams do not have a project management problem. They have a 'where did that decision go' problem."
Same topic, completely different feel, and the second version sounds like a person because you showed the model a person. The lesson is that the sample does most of the work, and the explicit rules catch what the sample misses.
The limits, and what this does not fix
This approach will not make AI writing pass as your own in a way that fools people who know you well, and it will not fix factual errors or logic gaps. Voice and accuracy are separate problems. It also costs time: you have to gather samples and edit, which can be slower than writing the first draft yourself for short pieces.
For a two-sentence social post, the setup may not be worth it. There is also a real risk of over-correcting into fake casualness, where every sentence has a joke and the piece reads like a persona rather than a person. The honest answer is that AI can get you 70 percent of the way to your voice, and the last 30 percent is editing you have to do.
If you want to check whether a draft still sounds robotic before you publish, there is a useful walkthrough on how to stop AI from writing content that sounds like a robot wrote it. One more tip: read the draft out loud.
Your ear catches stiffness that your eye skips, and the sentences you stumble over are usually the ones to rewrite.