InkFlow: Fighting AI Slop with an Adaptive Writing Twin

As AI-generated content floods the internet, "AI slop" is becoming a source of reader fatigue. The InkFlow project, published by developer Soumya Ghosh on Dev.to, proposes a counterintuitive approach: rather than chasing AI that writes more like a human, first learn how one specific person speaks.

What problem InkFlow aims to solve

InkFlow positions itself as an "Adaptive Writing Twin" and an "Anti-AI-Slop Engine." The naming itself reveals the design philosophy: it is not just another general-purpose writing assistant, but an attempt to replicate a specific author's expressive habits and to actively filter out the traces that make text feel mechanical.

The project comes from a week-two submission to the Hacktoberfest open-source AI challenge, categorized as a "Side Hustle Submission." Judging by the public tags—#devchallenge, #hf26challenge, #ai, #opensource—it carries a dual identity as both a competition entry and an open-source tool. That means the code can be reviewed and reused, but it also means the project is currently closer to a proof of concept than a mature product.

Why "AI slop" is hard to remove

The term "AI slop" has been circulating in English-language tech communities for some time, referring to text that is obviously generated by a large model and lacks information density. Its typical features include: overly symmetrical sentence structures, habitual "first/second/finally" framing, empty summary paragraphs, and piles of adjectives with little concrete argument.

Simply prompting a model to "write like a human" often fails, because the model's understanding of "human-like" is itself a statistical average. InkFlow's entry point is to treat individual writing habits as constraints—each author has different word preferences, sentence-length distributions, and paragraph rhythms. Being truly "human-like" should mean being "like this person."

AI slop is a derogatory label that emerged in English-language tech communities around 2024, initially used to describe the flood of low-quality AI-generated images and copy on social media, later extended to text. It differs from ordinary "AI-generated content": the latter only indicates the source, while the former emphasizes the information pollution the content imposes on readers. The popularity of the term reflects a shift in platform ecosystems—once generation costs approach zero, what becomes scarce is no longer the quantity of text, but credibility and information density. For writing-tool developers, "anti-slop" is really a quality-control problem, not merely a style problem.

Open-source approach and competition context

InkFlow chose to open-source during Hacktoberfest, and this timing has practical significance. Hacktoberfest is an open-source contribution event launched by DigitalOcean every October, in which participants must submit a certain number of pull requests. Submitting a writing tool as a project means it needs to be lightweight enough to be understood and contributed to by others, which also explains why the project emphasizes engine architecture over product packaging.

Judging by the "Side Hustle" category label, the author clearly treats it as a side project that can run independently of a main job, rather than a full-time commercial product. Such projects typically follow a path of first building a technical demo, then deciding whether to keep iterating based on community feedback.

Hacktoberfest was launched by DigitalOcean in 2014 and takes place every October. Participants earn rewards by submitting pull requests to open-source repositories (in recent years, the options changed to planting trees or receiving digital badges). The event places clear requirements on project maintainers: repositories must be tagged with the hacktoberfest topic, provide contribution guidelines, and promptly handle the influx of PRs. For a competition entry like InkFlow, participating in Hacktoberfest means the author needs to organize the code into a structure others can read and modify, which to some extent pushes the project from a "demo script" toward a "maintainable tool," but also brings the real pressure of uneven PR quality and rising maintenance costs.

The practical challenges of adaptive writing

For a writing twin to work, at least three layers of problems must be solved.

Extracting personal style

Writing style is not simply word-frequency statistics. Sentence-length variation, punctuation habits, paragraph-opening patterns, and the frequency of transition words together form a person's "textual fingerprint." Extracting them requires enough personal corpus as a training base; when the sample size is insufficient, the model easily overfits to a handful of articles.

In natural language processing, authorship attribution is a mature research direction, aiming to determine author identity based on textual features. Common features include function-word frequency, syntactic tree depth, punctuation distribution, and part-of-speech n-grams. InkFlow's "textual fingerprint" idea is in the same lineage as this field, but with the opposite goal: not to identify the author, but to replicate the author. Recent research on few-shot style transfer (such as personalized generation based on a small amount of text) suggests that dozens to hundreds of samples from the same author can capture a stable style to some extent, but with too few samples the model tends to mistake topic features for style features—for example, mistaking an author's frequent writing on technical topics for a stylistic preference for abstract nouns.

Balancing style and content

Even if the model learns personal style, it must still ensure that content quality does not decline during generation. Forcing personal sentence patterns may lead to awkward expression, and the omissions or leaps that some authors habitually make can expose logical gaps when expanded into complete paragraphs.

Criteria for "anti-AI-slop" judgment

The core difficulty lies in defining what "AI slop" is. If the rules are too mechanical—such as banning certain words or sentence patterns—they will misfire on normal expression; if it relies on another model to judge, it introduces new uncertainty and cost. The information InkFlow has made public so far does not specify which judgment strategy it uses.

Using another model to judge whether text "looks AI," technically speaking, falls under the LLM-as-a-judge paradigm. This approach has been widely used in evaluation, but has known biases: judging models tend to prefer text with neat structure and standardized wording, which is precisely at odds with the casualness of human writing. In addition, the judging model is itself a statistical product; using model A to judge model B's output easily creates a loop of "models evaluating models." A more pragmatic alternative is rule-based heuristic filtering—for example, detecting quantifiable indicators such as low sentence-length variance, high connective density, and overly symmetrical paragraph structure—but such rules require repeated tuning for specific languages and target readers.

Why this direction is worth watching

The writing-tool market is already very crowded, but most products compete in the direction of "letting AI write for you." InkFlow represents another path: helping a specific person preserve their own way of expression while removing machine traces.

For content creators, the practical value of this direction may be higher than that of general-purpose writing assistants. The distinctiveness of a person's long-term output is itself an asset. If AI can assist with expansion without damaging personal style, writing efficiency and brand consistency can improve at the same time.

Its open-source nature also lowers the barrier to trying it. Although InkFlow is currently just a competition project, the availability of the code means the development cost of similar tools is falling. In the future, more lightweight tools customized for personal style may emerge, rather than a uniform lineup of large writing platforms.

One expectation to keep in mind: at this stage it is closer to a directional experiment than a plug-and-play solution. In a release article of about a six-minute read, the author mainly introduces the concept and the competition background; specific technical implementation details and effectiveness evaluations have not yet been fully laid out. For early-stage projects like this, paying attention to the idea is more meaningful than paying attention to its completeness.

One-line Summary

InkFlow attempts to replicate a specific author's writing style and filter out AI slop—a proof of concept from an open-source competition.

Paragraph Summary

InkFlow is a writing tool submitted by developer Soumya Ghosh to the Hacktoberfest open-source challenge, positioned as an "Adaptive Writing Twin" and an "Anti-AI-Slop Engine." Its core idea is not to make AI write more like a human in a generalized sense, but to learn the word choices, sentence lengths, and paragraph rhythms of one specific author while actively removing the mechanical traces common in large-model-generated text. The article outlines the problem the project aims to solve, the difficulty of removing AI slop, the open-source and competition context, and three practical challenges: style extraction, balancing style and content, and criteria for judging AI slop. The author believes this direction has practical value for content creators, but the project is still at an early proof-of-concept stage, with technical details and effectiveness evaluations not yet fully laid out.