I’ve spent the previous two years watching the evolution of generative AI fashions intently, from GPT-3.5 to GPT-5, and the identical leapfrogging amongst Claude, Gemini, and Copilot. The progress has been astounding, particularly in how these fashions deal with language. Early outputs had been inflexible and repetitive, however the newest variations can write with a move that just about feels pure. They choose up rhythm, tone, even bits of humor. From a linguist’s viewpoint, that’s a powerful shift.
As a result of my background is in linguistics, I’ve been much less within the technical specs and extra within the language itself, that’s, how these fashions write, how they construct that means, and the way their “voice” adjustments from one model to the subsequent
Early ChatGPT textual content had a form of stiffness: overly formal, flat, repetitive. With GPT-5, the sentences move higher, the rhythm feels extra pure, and generally it even carries a hint of character.
From a linguistic viewpoint, that’s an enormous leap.
However right here’s what I’ve seen after utilizing these methods each day: regardless of all of the progress, their writing nonetheless falls in need of what we’d name superior human expression. It may possibly mimic fluency, however not perception. It organizes data nicely, but it hardly ever surprises you with a thought that feels authentic.
For on a regular basis writing (e.g., drafting an e mail, sharpening a paragraph, summarizing textual content) it’s an unimaginable companion. However for severe writing (e.g., a analysis paper, a ebook chapter, a conceptual essay) it merely doesn’t maintain up.
I usually describe it like this: AI can help the author, however it could’t be the author. It edits, refines, and rephrases superbly, but it surely doesn’t generate that means within the human sense of the phrase.
When GPT-5 got here out, I hoped this would possibly lastly chang, that we’d see a deeper linguistic shift, one thing nearer to how skilled writers cause via language.
However that leap hasn’t arrived.
The fashions appear to have reached a form of ceiling of their expressive capability.
That acquired me considering: possibly we’ve hit the linguistic limits of this strategy.
For actual progress, I argue, AI methods would possibly want entry to richer, extra refined knowledge; the form of language present in tutorial books, peer-reviewed journals, and long-form essays the place complicated reasoning lives.
The issue is that almost all of that materials is locked behind paywalls and copyright. With out it, the fashions are studying primarily from public web textual content, which skews towards surface-level writing.
So when individuals ask me how far AI can go as a author, I’d say: it will depend on what sort of language it’s allowed to study from.
Till fashions can legally and ethically prepare on high-quality tutorial and literary language, we might keep caught on this in-between zone: textual content that sounds clever, however doesn’t fairly assume.
