You gave it the original text. It still sounds like AI. Here's what the papers say.

“I pasted whole chapters of the original into the prompt. The continuation still reads like AI. Am I prompting it wrong, or is this permanently impossible?”

We get this question a lot, and we’ve asked it ourselves. Over the past year we ran a pile of continuation experiments: a five-tier context-size ablation, twenty-round drift chains, a nine-model blind arena across two genres. What we lacked was the other half of the picture — what academia has found. So we read the 2024–2026 style-imitation literature and reconciled it against our own data. This post is the reconciliation.

The short version: “writes like the original” is not one problem but three. Word choice can be matched today. Sentence rhythm cannot be fixed by prompting, though validated paths exist. And “reads like the same person wrote it” is currently beyond every published method — worse, nobody has a reliable way to even judge it. “Permanently impossible” overstates it, but that last layer has no visible solution yet.

The counterintuitive part: continuation is the most favorable task shape

A September 2025 paper tested exactly this question. How Well Do LLMs Imitate Human Writing Style? used an authorship verifier as the referee (a classifier that decides whether two texts came from the same person) and compared four ways of asking. Describing the target style in words: mostly useless. Supplying one or two sample passages: up to 23.5 times higher style-matching accuracy. Strongest by far was completion form — give the model the first half of a human text and let it continue — where four of five models hit 99.9% style agreement and the verifier’s confidence saturated. The authors’ explanation: a human-written prefix anchors the model inside that stylistic manifold, and continuation keeps it there.

Handing the model reference text and asking it to continue is precisely that strongest form. So the task isn’t inherently doomed. Of all style-imitation setups, it has the best starting position.

Two honest footnotes. First, that paper’s corpus is English essays and its referee is a statistical verifier, not a fiction reader. Second, the same paper found that satisfying the verifier is not the same as reading human: even at 99.9% style agreement, the AI output’s perplexity — how unpredictable the text is — stayed at roughly half the human level, 15.2 against 29.5. Statistical resemblance to the author and not-sounding-like-AI are separable goals. That finding runs through everything below.

From the opposite direction, EMNLP 2025’s Catch Me If You Can? Not Yet poured cold water: across 400+ real authors and 40,000 generations per model, LLMs imitated structured registers like news and email passably, but kept sliding into a generic voice on blogs and forums. Personal style is implicit. Authors themselves can rarely articulate it, so no prompt will.

It saw your reference text. That was never the problem.

Last year we ran a five-tier ablation: original prose from an 8.9-million-character cultivation novel, packed into contexts from 4k up to 200k tokens, with no style instruction at all. At the 200k tier, simile-cliché density in the continuations came out ten times higher than the original’s. Two hundred and seventy thousand characters of the author’s actual prose sat in the window, and the model wrote its own way regardless.

Back then we only had the phenomenon. The recent literature supplies the mechanism, and the papers interlock tightly.

The CMU study in PNAS, Do LLMs write like humans?, measured GPT-4o and the Llama 3 family against Douglas Biber’s inventory of linguistic features. LLMs have a house style — noun-heavy, informationally dense, overusing participial clauses — and hold onto it even when explicitly told to match informal speech. The uncomfortable direction: instruction-tuned variants sit further from human variation than their base models. The same training that makes a model obedient nails its prose style in place.

How it gets nailed has now been taken apart. Creativity Has Left the Chat (2024) observed aligned models collapsing into a few attractor outputs, high-confidence and low-exploration. Verbalized Sampling (2025) traced the root one level deeper, into the preference data itself: human annotators systematically favor typical text — typicality bias — so even a perfect reward model trained on human preferences still collapses. Narrative Flattening (2026) measured fiction continuation checkpoint by checkpoint, base to SFT to DPO, and found each post-training stage compresses thematic, emotional, and stylistic variation a bit further. The starkest number comes from Antislop (ICLR 2026): some stock phrases appear over 1,000 times more frequently in LLM output than in human writing.

In plain terms: the model saw your reference text. Its hands were already shaped by training. Reference text supplies imitable material, but which text sets the baseline has to be said out loud. In our tests, one explicit baseline sentence in the system prompt cut cliché density roughly in half at every context size — including 200k tokens, where the instruction showed no sign of dilution.

“Sounds like the original” is three separate acceptance tests

After the reconciliation we split style fidelity into three layers, each with a different verdict.

The lexical layer — no AI stock phrases, use the author’s vocabulary. Solved: explicit baseline directive plus sample text works, numbers above.

The rhythm layer is much harder. Our source novel’s sentence-length variation coefficient is 0.84 — short and long sentences colliding violently, onomatopoeia sitting on its own line. Every experimental arm we ran landed between 0.28 and 0.48; the best window in any multi-round chain touched 0.79. No instruction wording moved it back to the original’s level, which matches the PNAS finding that grammatical-structure preferences don’t yield to prompting. This layer needs surgery, not persuasion. Next section covers the surgical options.

The third layer is overall texture — the close-reading sense that this is the same person. Our nine-model arena produced the perfect specimen: Qwen won every structural metric (average sentence length 32.1 against the original’s 32, lowest cliché density in the field) yet blind readers ranked it fourth or fifth. Their reason: it wrote smell and touch into every window, and the original contains zero smell descriptions across the entire book. Structural metrics cannot detect “writing things the author never writes.” (Full arena results here.)

The bigger problem: the judge’s seat is empty

One root cause of the third layer staying unsolved is that nobody can score it. CHI 2024’s Art or Artifice? had 10 professional writers blind-evaluate 48 stories on 14 creativity tests: LLM stories passed 3 to 10 times fewer tests than professionals’ work — and when LLMs were asked to grade the same stories, not one model’s judgments correlated positively with the experts. Our own double-blind run looked worse: LLM judges asked to pick the human-written passage agreed with human judgment only 12% of the time, while agreeing with each other 86% of the time. Consistently wrong. (Full writeup here.)

The engineering consequence is concrete: without a reliable automatic evaluator there is no objective function to optimize, and you cannot build a regression test for “same-author feel.” The only trustworthy judge today is a careful human reader. Expensive, slow, unscalable. Benchmarks are appearing — WebNovelBench (2025) ranks 24 models against 4,000+ Chinese web novels and cleanly separates canon, popular fiction, and LLM output — but it scores how well a model writes, not whether it writes like one particular author. That seat is still empty.

What’s been validated, and how far each method reaches

Line up the published methods by which layer they reach and what access they need:

TinyStyler deserves its own sentence: an 800M-parameter model conditioned on authorship embeddings outperforms GPT-4 at authorship style transfer. What style imitation lacks was never model scale. It’s a dedicated conditioning channel. The right-hand column also explains why every consumer app lands at roughly the same ceiling: the first two rows are the limit of prompt engineering, and the other four all require touching inference or weights — out of reach for products built on someone else’s API.

So — permanently impossible?

Layer by layer: the lexical layer is solved. The rhythm layer is blocked on access, not on knowledge — the methods are published, but commercial chat APIs expose no inference hooks; the day a provider ships base-model endpoints or sampling hooks, this layer moves. The genuinely open problem is the third layer, and it’s blocked evaluation-first: no judge, so no training signal.

The most honest number in our own data is that variation coefficient: the original sits at 0.84, and no chain we ever ran got back there. Until that number recovers, “writes like the original” stays an unfinished problem on our board too. What the source-annotation switch in the app delivers today is the first layer, finished: twenty rounds of continuous continuation without the slide into AI voice, cliché density held at the original’s level. The second layer waits on access. The third waits on a judge.