Can we actually measure what makes writing good enough to teach an AI to write well?

This explores whether literary quality can be turned into a reliable training signal for AI, and where those attempts break. The corpus suggests that precision isn't really the obstacle. Parts of literary writing are very easy to measure. The hard part is that the measurements capture the wrong things, and the judges doing the measuring are easy to sway.

Start with what measures well. Style is highly detectable: even GPT-2 can identify an author from stylistic patterns with about 95% accuracy Can language models truly understand literary style?. Narrative structure is measurable too. StoryScope separated AI fiction from human fiction with 93% accuracy using only choices like how much agency characters have and whether events are told in order, with no surface style involved Can AI stories be detected without analyzing writing style?. Some researchers have even turned reading efficiency into a number, counting unique pieces of knowledge per token, and found that LLM prose scores lower because models pad Can we measure reading efficiency as a quality metric?. These are precise signals. The catch is that detecting a pattern is not the same as knowing why it matters. A model that recognizes an author's style still can't explain what those choices accomplish, which amounts to cataloguing rather than criticism.

The weaker link is the judge. When AI models rate literary passages, simply labeling a passage as human-written raised their scores by 34 percentage points. That bias is 2.5 times stronger than in human readers, who show it too Do authorship labels bias how we judge literary quality?. LLM judges can also be fooled by fake citations and attractive formatting Can LLM judges be fooled by fake credentials and formatting?. If you train a model on these judgments, you risk teaching it to perform the signals of quality rather than produce quality.

The more promising results narrow the target instead of sharpening the ruler. Sun argues that LLMs stay weak as creative writers but become capable editors once they're given one writer's explicit taste rubric. Part of the reason is that general writing quality is so hard to measure, while a specific person's preferences can be written down Can LLMs become good editors by learning a writer's taste?. Argument quality points the same way. Fine-tuning on labeled examples teaches surface patterns that don't carry over to new kinds of argument, while spelling out an explicit theory of what makes an argument good does generalize Can models learn argument quality from labeled examples alone?.

One more idea from outside literature: 'scientific taste' turned out to be learnable when it was grounded in a community's long-run judgment. Models were trained on 700,000 paper pairs ranked by how often each paper was later cited Can models learn what makes research worth doing?. Literature has a rough equivalent in which books get reread, taught and quoted over decades, though the corpus doesn't yet include anyone trying it. Overall, quality seems trainable when someone commits to whose taste counts, whether that's one writer's rubric, an explicit theory, or a community's track record. Treating it as a single universal score is where it breaks.