Despite rapid advances in math and coding, AI still struggles to produce prose that doesn’t “read” like AI. I have noticed that a giveaway of AI-generated prose is wordiness.
Having evaluated leading models (e.g. Opus and Astra) for writing, none of them have impressed me. The prose is wordy and bad. It lacks the effortless, unforced flow of quality human writing. It’s not so much that quality human writing is simple or terse, but that each word serves a function, irrespective of length. Even in a cathedral, each piece of the architecture serves a role, just as each brick does in a modest building. By comparison, AI writing adds words that offer no value. I made a post detailing how AI adds extra words to a short writing sample. It cannot just get to the point.
So my summary:
1. Pangram is scary-accurate in my experience. Even websites and services that purport to fool Pangram can fail to do so. It can flag AI content extremely well, even when it’s not that obvious what gives it away.
As Pangram and other detectors get better, students outsourcing their assignments will have a harder time avoiding detection. There seemed to be a window from 2023–2025 or so when detection was not that good, yet AI software was good enough to fool teachers. That window seems to be narrowing as detection improves relative to AI writing ability.
2. Human writing has various tics or patterns that make it “human,” and for whatever reason AI still cannot reliably reproduce them despite 3+ years of rapid AI progress elsewhere. All AI writing reads as if it is by one of two authors: “Claude” or “ChatGPT.”
With writing, the extra difficulty is that your cards are always face up. With math, the results hide behind dense notation. Or with coding, few actually read the source code. Clumsy math notation or bloated code isn’t fatal as long as the theorem is correct or the program works for the end user. With writing, bad style can quickly turn readers away before they even give the arguments a chance, however good those arguments may be.
3. Maybe there is less commercial demand for good AI writing software than for math or coding tools. That may explain why the quality is bad or has stalled. There may simply be more money in better code than in better prose. Early on, no one really know what the biggest potential for AI was, so the focus was on images, short video, and text, as those were the most obvious applications. Then in mid 2025 came the goliath of coding + math. It soon became obvious that AI being more like a smarter version of Mathematica was the future, not “the arts”.
4. I think AI writing fails because it tries too hard to sound correct. It’s like an ESL speaker who sounds forced despite being proficient, because they’re trying so hard to be correct and avoid being misinterpreted. AI similarly keeps correcting and hedging itself. This leads to bloated, wordy writing. Much of good writing is what is unsaid.
5. Regarding the debate over whether thinking is the same as writing, IMHO, they are not the same. “Idea generation” and writing are different processes, although they can overlap. Many people rely on AI because, although they have ideas, they are unable to articulate or structure them into an essay. The popularity of Twitter shows people can generate ideas, although in short form. This is not surprising: good writing is hard, requiring research, organization, and style/grammar. Research helps generate ideas, but the other two are equally important.
6. Much of the work of writing is in supplying the mortar between the bricks of ideas. Podcasts and Twitter are so popular as mediums for conveying information because the gaps can be filled with airtime (or, in the case of Twitter, left as discrete gaps between tweets), whereas with writing, you need to supply the transitions. Difficulty does not scale linearly with length: a coherent 1,000-word essay can be harder to write than ten separate 100-word tweets. The surface area for errors goes up, and more organization is required. There are more possible ways to arrange the text.