Filler Pattern Detection in AI-Generated Drafts

Filler pattern detection in AI-generated drafts is most useful when it examines architecture. A predictable arrangement of claims, announcements and recaps survives sentence-level editing. Detecting that arrangement gives editors something a word-level verdict cannot provide, namely a diagnosis of what to remove and what to change upstream.

Why AI drafts feel similar even when they don't share words

An author reads back an AI draft and feels it dragging, although no sentence is obviously wrong. The language is clear, the headings are orderly, and the claims appear relevant. Yet the introduction explains what the article will explain, each section comes as promised, and the ending goes back to where it began. The reader has moved through the page without gaining as much as its length promised.

That drag comes from structural convergence, even when the words themselves do not repeat. SlopShape identifies a common shape in AI-generated commercial posts across models, including a payoff promised in the title, an early thesis, an editorial-explainer voice and a closing restatement. Different words can occupy the same positions and perform the same redundant work. The filler includes tired expressions and a scaffold of repeated announcements and summaries that stalls the argument.

What the structural shape of an AI draft actually looks like

The first visible feature is thesis placement. Before the first substantive section, the draft states its answer and announces how the article will deliver it. An early thesis becomes suspect when the surrounding sentences describe the article’s intentions without establishing its case. SlopShape names this feature Thesis Before First Unit and finds that human posts more often withhold or embed the thesis.

The ending supplies another independent check. Under Conclusion Behavior, a draft may restate its opening claim or reframe it through what the argument has established. Those moves are not equivalent, because a restatement repeats the destination while a genuine reframe changes what the reader understands about it. Functional Stage Inventory checks for a similar problem in other parts of the draft. A section set aside for summary or synthesis can add another stage without adding another thought.

External Participation Pathway looks at the reader’s role in that structure. A draft can explain a subject to an imagined audience without giving that audience a question, decision point or objection to inhabit. Combined with announcements and recaps, this produces the self-announcing sequence of telling the reader what the article will do, doing it, and saying it has done it.

Idea convergence adds a second dimension. Sem-Detect, by Duarte et al., ICML 2026, found that different AI models converge on similar points in peer reviews while human reviewers introduce more unique and diverse ones. The observation applies directly to peer review and serves as a parallel for marketing posts. It nevertheless gives commercial editors a useful distinction between varied wording and genuinely different claims.

Why word-level detection misses what structural detection catches

Word-level detection can identify unedited AI output effectively, but that strength does not settle its value in an editorial workflow. Editors usually work on drafts that have already been revised. If a detector relies on wording that gets edited out, its signal weakens even when the same problems stay in the draft. Structural detection checks whether the same stages still do the same work.

SlopShape’s 2026 analysis tested this difference by having each model reword the posts it had written while keeping the same claims, facts and links. An average of nearly three-quarters (73%) of each post’s 13-word sequences no longer appeared verbatim. The word-level classifier performed better on unedited text but degraded sharply after rewording. Structural detection scored nearly identically in the edited and unedited conditions. The practical distinction is that a new verbal surface did not produce a new architecture.

Commercial humanizer tools have made this kind of rewording a routine step that takes little effort to carry out. Masrour et al. (2025) documents this commercial ecosystem. A workflow built around lexical signals therefore depends on features that another routine step is designed to erase. Even when those signals remain accurate, an authorship verdict does not identify an unnecessary summary or a conclusion that merely repeats the introduction.

TELL, by Creo and Ranganath at UC San Diego, makes explicit that a numeric score without an explanation gives a professor or editor no basis for judging the finding in context. Word-level tools may be fast and free, while structural analysis takes more computing power and more judgment. But speed does not turn a verdict into a diagnosis, and diagnosis is what an editor needs to decide where to cut.

How to detect filler patterns structurally — a working checklist

The first check, thesis placement, examines whether the introduction both states the answer and announces the structure before substantive work begins. Look for sentences that describe what the article will cover without contributing evidence, a distinction or an implication. The signal is the arrangement. An opening can state its message plainly without also rehearsing every stage that follows.

The second check is conclusion behavior. Compare the ending’s contribution with the opening’s claim. Look for a restatement disguised by new wording, and distinguish it from a genuine reframe or new implication. This comparison tests whether the argument has changed what the ending can say. If the conclusion could have been written before the body, it warrants scrutiny.

The third check is functional stage inventory, which looks at the job each section does. Look for a dedicated summary or synthesis stage that repeats work already completed. Human posts more often close on implication.

The fourth check is external participation, which looks at whether the reader has something to do besides take in an explanation. Look for a real question, decision point or objection that the reader can inhabit. The fifth check is idea convergence. Compare the draft with a competing AI-generated post on the same topic. Look for the same claims in the same order even when the words differ.

SlopShape validated its feature instrument against human annotation, using a 203-feature instrument with human-human and human-model agreement of kappa 0.939 and 0.951 respectively. An LLM-based linter can run the first four checks using a defined set of features, while the fifth also needs a draft to compare against. TELL offers the appropriate output principle for this process. The result should be a flagged feature with a sentence-level pointer.

Why the structural shape persists even after editing — and what actually changes it

Sentence-level editing leaves structural filler intact when it preserves the order of claims, the stages of explanation and the conclusion’s logic. A smoother introduction can still announce the entire article unnecessarily. A shorter summary can still repeat things that need no summary.

SlopShape’s rewording test supports that distinction, but it also leaves an important ambiguity. The way the test was set up leaves it unclear how much of the structure came from the model and how much came from the brief. If the brief calls for a thesis-first introduction and a concluding recap, some of the detected shape may be an artifact of those instructions. For practitioners, however, either explanation points to the same lever, which is the generation setup.

Changing the brief means changing what the draft is required to do. Those choices shape how the argument is built, which evidence comes first and what the ending does. When a brief still calls for stages that repeat earlier work, an editor working on sentences has little room to cut them. Structural revision must be allowed to alter those decisions.

The model’s defaults also matter because base models tend toward the generic web-English shape dominant in their training data. Fine-tuning on a brand corpus can change both the structures treated as normal and the vocabulary used to express them. RAG over on-brand examples can supply alternatives at generation time, including posts that embed their thesis, close on implication or involve the reader in a decision.

How adversarial multi-model review catches structural filler that a single model misses

Once the problem is located upstream, review must be designed to challenge structure. A model that writes a draft is poorly placed to question the defaults that shaped its own answer. A second model given an adversarial role has a different assignment, which is to identify unnecessary scaffolding, redundant stages and conventional choices that have not earned their place. This is the same architectural principle that makes code review useful for logic errors a self-review misses.

TriAdReview, by Zhou, Dai, and Ling, June 2026, provides controlled evidence for adversarial review across five benchmark tasks. Its triple adversarial model achieved a 10.1% overall improvement over the single-model baseline. That result offers indirect evidence for filler removal in commercial articles through its finding on tasks in which generators over-produce by adding unnecessary components or defaulting to popular choices. A challenger can identify what to remove or simplify, whereas a reviewer asked only to improve an output may add more.

The boundary matters as much as the gain. TriAdReview showed degradation on requirements-analysis tasks, where completeness mattered and simplification was the wrong instinct. Most marketing blog posts fit the former category, but adversarial review is not a universal instruction to shorten everything.

One model drafts, a cross-vendor critic attacks the draft, and a human editor makes the final call. The cross-vendor distinction matters because models from the same family may share structural defaults. This separation makes challenge an assigned function rather than an optional second pass by the same generator.

What structural filler detection requires from the generation pipeline, not just the editor

Reliable structural detection requires the writing system to carry editorial expectations in a form it can use. For generation-time control, the requirements in a PDF voice guide need to enter the process through a structured prompt, a tagged retrieval corpus or a fine-tuned adapter. Those requirements must describe argument patterns as well as word choices, or the system will keep producing the same architecture in a different voice.

A corpus of the best samples shows what those patterns look like in practice. Post-training on a brand’s own on-voice posts changes the model’s defaults toward the brand’s actual ways of building arguments. PEFT, or parameter-efficient fine-tuning, makes this possible by changing a small set of added parameters while leaving most of the base model as it is. Retrieval gives the model examples of how to structure the piece while it writes. Both tasks need to be shared beyond the editor who gets the finished draft.

Detection signals also need distinct roles. Provider-level watermarks can contribute to provenance assessment, but they do not diagnose whether a section advances an argument. Watermarks remain vulnerable to black-box querying and scrubbing. Structural detection does not depend on watermark integrity, which makes it more durable as an editorial quality gate. A pipeline can therefore account for provenance without confusing it with writing quality.

Automated structural linting belongs before human review, at the stage where tests belong in a CI pipeline. Its output must expose defined features and flagged locations that an agent or editor can inspect. A structural linter, a voice-corpus retriever and an adversarial critic can operate as separate tools connected through API. Their separation allows generation, detection and challenge to retain distinct responsibilities. Serious AI writing infrastructure makes structural quality a property of the pipeline.