I’ve spent the past two years watching the evolution of generative AI models closely, from GPT-3.5 to GPT-5, and the same leapfrogging among Claude, Gemini, and Copilot. The progress has been astounding, especially in how these models handle language. Early outputs were rigid and repetitive, but the latest versions can write with a flow that almost feels natural. They pick up rhythm, tone, even bits of humor. From a linguist’s point of view, that’s an impressive shift.
Because my background is in linguistics, I’ve been less interested in the technical specs and more in the language itself, that is, how these models write, how they build meaning, and how their “voice” changes from one version to the next
Early ChatGPT text had a kind of stiffness: overly formal, flat, repetitive. With GPT-5, the sentences flow better, the rhythm feels more natural, and sometimes it even carries a trace of personality.
From a linguistic point of view, that’s a huge leap.
But here’s what I’ve noticed after using these systems daily: despite all the progress, their writing still falls short of what we’d call advanced human expression. It can mimic fluency, but not insight. It organizes information well, yet it rarely surprises you with a thought that feels original.
For everyday writing (e.g., drafting an email, polishing a paragraph, summarizing text) it’s an incredible partner. But for serious writing (e.g., a research paper, a book chapter, a conceptual essay) it simply doesn’t hold up.
I often describe it like this: AI can assist the writer, but it can’t be the writer. It edits, refines, and rephrases beautifully, but it doesn’t generate meaning in the human sense of the word.
When GPT-5 came out, I was hoping this might finally chang, that we’d see a deeper linguistic shift, something closer to how experienced writers reason through language.
But that leap hasn’t arrived.
The models seem to have reached a kind of ceiling in their expressive capacity.
That got me thinking: maybe we’ve hit the linguistic limits of this approach.
For real progress, I argue, AI systems might need access to richer, more sophisticated data; the kind of language found in academic books, peer-reviewed journals, and long-form essays where complex reasoning lives.
The problem is that most of that material is locked behind paywalls and copyright. Without it, the models are learning mainly from public internet text, which skews toward surface-level writing.
So when people ask me how far AI can go as a writer, I’d say: it depends on what kind of language it’s allowed to learn from.
Until models can legally and ethically train on high-quality academic and literary language, we may stay stuck in this in-between zone: text that sounds intelligent, but doesn’t quite think.
The post The Kind of Writing AI Still Can’t Do appeared first on Educators Technology.
