AI writing sounds stiff because the model is doing exactly what its training and its app-level instructions reward: producing safe, neutral, formal text that is unlikely to offend anyone or be wrong — and that default is strong enough to override a casual tone request in your prompt.
You are not imagining it, and you are not doing anything wrong. The stiffness comes from three stacked layers: what the model learned from, how it was tuned to behave, and the hidden instructions the app wraps around your message before the model ever sees it. Understanding those layers tells you which lever to pull — and, just as usefully, when pulling it will not work.
The first layer is the training data. Models learn by predicting text, and a huge share of the text they learn from is written material: documentation, articles, reports, forum answers, help pages. Written English skews more formal than spoken English.
Nobody writes "hey can't make it tonight, rain check?" in a product manual. So the statistical centre of gravity for any given request sits closer to "I am unable to attend this evening" than to how you actually text a friend.
The second layer is tuning. After base training, models are adjusted to be helpful, harmless and honest, and that adjustment rewards a cautious, hedged, neutral register — the kind of voice that will not accidentally insult a user or commit to a claim. Politeness and vagueness get reinforced; slang and strong opinion get dampened.
The third layer is the app. Most chat products wrap your message in a hidden instruction that says something like "you are a helpful assistant" and may add formatting rules, safety reminders, and length guidance. That wrapper sits above your request in priority, and it is often the thing actually fighting you.
Here is what that looks like in practice. Suppose you type: "write a text to my friend cancelling tonight, keep it casual." A stiff output reads something like: "Hi, I wanted to let you know that I won't be able to make it this evening. I apologize for any inconvenience and hope we can reschedule soon."
That is a formal email wearing a text message's clothes. Now compare a genuinely casual version: "hey, something came up — can't make tonight. rain check?"
The difference is not vocabulary alone. It is sentence fragments, a missing subject, a lowercase opener, and no apology paragraph. The instruction that reliably shifts the output is not the word "casual" — it is a style sample plus an explicit register.
Try: *"Match this voice exactly: 'hey, something came up — can't make tonight. rain check?' Write my cancellation text in that voice.
No greetings, no apologies, max 20 words."* Naming a persona helps too: "write as a 28-year-old texting a close friend" gives the model a concrete target instead of an abstract adjective. Constraints do more work than adjectives, because "casual" is a vibe and "max 20 words, no greeting" is a rule.
Now the honest part, because this is where most advice oversells. Prompting for casual tone frequently produces the wrong kind of casual — forced slang, random emoji, "Hey buddy!!" energy that no human would send.
The model is reaching for a stereotype of informal speech rather than your actual voice, and it overshoots. Some apps also hard-code a formal system prompt you cannot see or override, in which case no amount of clever phrasing in your message will win; the wrapper outranks you. A practical decision rule: if the output is stiff, first paste a one-sentence sample of your own writing and state the register explicitly.
If it is still stiff, the app's hidden instructions are overriding your request — switch apps, or use a custom instruction or persona setting if the product offers one. And if the casual version comes back with emoji you did not ask for, tighten the constraint rather than loosening the tone request.
One more thing worth knowing: the same model behind two different apps can produce noticeably different tone, because the wrapper differs even when the underlying model does not. Our internal AI tool database, which holds pricing and capability snapshots for 360 AI tools recorded at verification time, is a reminder that these products are not interchangeable — the packaging matters as much as the engine.
If you want to understand why a tool behaves the way it does, the model is only half the story.