Why AI LinkedIn posts sound like AI — and how to fix it
6 min read
Readers got good at spotting AI writing quickly, and not because of anything mysterious. The tells are mechanical and repeatable.
The usual response is to hunt for a better prompt. That treats a sourcing problem as a phrasing problem. If the tool has nothing of yours to work from, it will fill the gap with the average of everything it has read — and average is exactly what gets recognised.
The mechanical tells
These are the patterns that show up again and again in generated posts, and every one of them is countable.
- Uniform sentence length, giving the text a flat, metronomic rhythm.
- Stock scaffolding: in today's fast-paced world, it's not just X it's Y, let that sink in.
- Symmetrical structure — three points, each the same size, each with a tidy lesson.
- Specificity that goes missing at exactly the moment it would matter.
- A closing question aimed at comments rather than at anyone in particular.
- Em dashes and emoji used at a rate the author has never used them before.
Why better prompts do not fix it
A prompt can change register, and register is not voice. Ask for punchier and you get shorter uniform sentences. Ask for authentic and you get a manufactured vulnerable anecdote, which is worse than the generic version because it is fiction attached to your name.
The gap the model is filling is your material: your actual position, your actual example, your actual number. No instruction supplies that.
Change the input, not the instruction
Start from something you said. Talk for ninety seconds about the thing you already have an opinion on and use that transcript as the source. It will be messy, repetitive and full of half-finished sentences — and it will contain your reasoning, your examples and your vocabulary.
Now the job is editorial rather than generative: cut, order, sharpen. That is a job a model does well and one where it cannot smuggle in someone else's thinking.
Keep opinions honest
The most damaging failure is not a clumsy sentence. It is a confident opinion you do not hold, published under your name — a reversal of a position you have argued before, or a claim you cannot back.
Any AI in your publishing loop should be checked against what you have actually said. If it cannot support a strong statement from your own material, the right behaviours are to soften it, ask you, or label it as a suggestion — never to assert it as yours.
How Yolect handles it
Yolect never starts from a topic prompt. It starts from a capture: a voice note, a video you appeared in, or your own rough text. Deterministic code — not a model — measures rhythm, stock-phrase density, first-person presence and paragraph shape, and those numbers feed the grade every draft has to pass.
A belief consistency pass then checks the draft against the positions extracted from your published work. Contradictions get rewritten, uncertain claims get raised with you, and anything genuinely new is marked as a suggested perspective rather than presented as your established view.
Questions people ask
- Can AI detectors tell if my LinkedIn post was AI-written?
- Detectors are unreliable in both directions, so they are a poor thing to optimise for. Human readers are the real detector, and they respond to flat rhythm, missing specifics and borrowed opinions.
- Is it fine to use AI for LinkedIn at all?
- Yes, when it works on your material rather than replacing it. The line worth holding is that the thinking, the examples and the opinions are yours.
- What is the fastest way to make a generated draft sound human?
- Break the rhythm and add one specific detail only you could know — a real number, a real objection, a real name. Then delete the closing engagement question.