Research and references
These sources helped shape Unslop's checks and editing workflow. Each entry separates what the researchers found from the choice we made for this project.
None of these studies directly validates Unslop's current production models or its scoring thresholds. A paper's publication date can be much newer than the models it tested. Links below go to the authors' manuscripts, publishers, or official documentation.
For the current implementation, read How Unslop works. For requests and responses, use the API guide.
Rhythm and style
Sentence and paragraph variation
Desaire and colleagues, 2023. ChatGPT or academic scientist? Read the manuscript PDF.
The study compared scientific perspective articles with early ChatGPT essays. Sentence-length spread and differences between neighboring sentences helped distinguish the groups. Paragraph-length spread was also informative in the training data.
Unslop uses that as a reason to report several rhythm measurements. Its warning cutoffs are our own provisional choices. The paper's high classification accuracy belongs to its narrow scientific-writing experiment, not to Unslop or writing in general.
Grammar depends on genre
Reinhart and colleagues, PNAS, 2025. Do LLMs write like humans?
Using GPT-4o and Llama 3 variants, the researchers found differences in grammatical and rhetorical patterns even when models were asked to continue human writing. Instruction tuning did not eliminate those differences.
For Unslop, this supports considering genre and the writer's examples. It does not make a particular grammatical construction an error. Unslop does not implement the paper's full linguistic feature extractor.
More vocabulary variety is not always better
Kendro, Maloney, and Jarvis, author manuscript revised 2025. Do LLMs produce texts with "human-like" lexical diversity? Journal record.
The study compares four ChatGPT models with 240 human writers across several vocabulary measures. The groups differed; newer models in the experiment did not consistently move closer to the human samples.
Unslop reports vocabulary measurements without treating higher diversity as a goal. Its editor is told to keep useful repeated terms. This experiment does not justify replacing every repeated word with a synonym.
Polishing can change a writer's voice
Sourati and colleagues, Nature Human Behaviour, 2026. The Shrinking Landscape of Linguistic Diversity.
Across multiple datasets, the authors report that model-assisted rewriting reduced differences in writing complexity between texts. The finding concerns variation across a collection of writing, not a required amount of sentence variation inside every paragraph.
Unslop's response is to favor limited edits and expose comparisons with approved samples. Those sample ranges are descriptive, not proof that a person's voice has been preserved.
Editing and meaning
Focused edits can help
Chakrabarty, Laban, and Wu, CHI, 2025. Can AI writing be salvaged?
The LAMP corpus contains 1,057 generated paragraphs edited by professional writers. The study identifies recurring issues such as cliches and unnecessary explanation. Automatic editing showed promise, while judges generally preferred professional edits.
Unslop supplies located findings and focused editing instructions. It tells the model to use facts already present. The study's creative-writing setting and older models do not establish that every rewrite will improve a business email or technical document.
Use specific feedback after a failed attempt
Madaan and colleagues, 2023. Self-Refine: Iterative Refinement with Self-Feedback.
This work studies model-generated feedback and revision across seven tasks using GPT-3.5, ChatGPT, and GPT-4. Iteration improved results in the experiments.
Unslop uses one optional retry per invalid patch, within the generation-call budget. Code supplies mechanical failure reasons, and the retry receives the original span. Meaning review checks the assembled result; a failed review returns the original without another rewrite. This is an engineering adaptation, not a reproduction of Self-Refine or a guarantee that another attempt helps.
A similar sentence can make a different claim
Pauli, Augenstein, and Assent, Findings of EMNLP, 2025. Mind the Style Gap.
The authors show that content-preservation metrics can look reliable on test sets that do not sufficiently challenge meaning changes. Their work introduces a more demanding evaluation set.
Unslop checks protected text and numbers separately from style. Optional AI review checks meaning, but remains fallible. Neither a small edit distance nor a lower style score establishes that the original claim survived.
Readability and fairness
Hitting a reading target is difficult
Barayan, Camacho-Collados, and Alva-Manchego, COLING, 2025. Analysing Zero-Shot Readability-Controlled Sentence Simplification.
The tested models struggled to simplify sentences to the lowest target levels, and automatic measures did not always reflect the quality of the changes.
Unslop measures the candidate instead of assuming the prompt worked. This paper studies CEFR language levels, not Unslop's Flesch-Kincaid grades or age mapping. Those scales should not be treated as interchangeable.
Simple English is not evidence of AI authorship
Liang and colleagues, Patterns, 2023. GPT detectors are biased against non-native English writers.
The tested detectors frequently mislabeled writing by non-native English writers. Their results also changed with alterations to language complexity.
Unslop therefore describes style patterns rather than assigning an authorship probability. Short sentences and straightforward vocabulary may be exactly what a reader needs. This study evaluates its selected detectors, not Unslop or every current detector.
Reading formulas are only one check
Agency for Healthcare Research and Quality, guidance reviewed 2015. Use caution with readability formulas.
AHRQ explains that formulas can draw attention to difficult wording but cannot fully assess comprehension. Familiarity, organization, and the reader's context also matter.
Unslop's grade estimate is a prompt to inspect a passage. The age-to-grade mapping is a convenience, not a validated reading-age test. Test important writing with its intended readers.
What remains unproven
The default score gate of 20/100 and 50 words, patch-size limits, and provider rotation are engineering choices, not findings validated by these papers. The gate is customizable per request. Earlier whole-document benchmarks do not evaluate the current micro-edit pipeline.
The implementation has automated checks for measurements, edit budgets, protection rules, and retry behavior. Those checks establish software behavior, not writing quality or detector accuracy.
Rhythm and repetition findings remain outside the style score. We still need blinded reviews of original and edited passages, including approved human writing, technical material, second-language writing, and text for younger readers. That evaluation should count false flags and meaning changes as well as helpful edits.
The papers above motivate things to measure and test. None establishes Unslop's 0.12 variation cutoff, ten-sentence minimum, five-paragraph minimum, or age conversion. The current thresholds are disclosed so people can examine them.