LLM-Dominant Content: Detection & Implications

• LLM-dominant content is defined as material predominantly generated by large language models with minimal human input, distinguishing it from mere AI-assisted writing.
• Detection methodologies aggregate page-level outputs using scalable pipelines and classifiers, achieving near-perfect accuracy on curated datasets.
• The rise of LLM-dominant content prompts critical debates on reliability, ethical standards, and moderation challenges across various digital domains.

LLM-dominant content denotes content that is automatically generated by LLMs with little human input (He et al., 18 Jul 2025). In the web-specific formulation, LLM-dominant websites are sites “where most, if not all, of the site’s textual content is generated using LLMs” (He et al., 30 Apr 2026). The category is narrower than AI-assisted writing: a human may polish text, but if the site is substantially authored by humans and only edited with AI, it is not the target class (He et al., 30 Apr 2026). The topic has become technically and institutionally important because websites rarely disclose such content, human readers struggle to distinguish it, and LLM-generated material can be unreliable or unethical because LLMs plagiarize and hallucinate (He et al., 18 Jul 2025).

1. Definition and conceptual boundaries

The defining property of LLM-dominant content is not merely the presence of AI assistance, but the dominance of the LLM in the authorship process. The web literature makes this distinction explicit. The 2025 preprint on website detection defines “LLM-dominant” content as web content automatically generated by LLMs with little human input, and frames the problem at website scale rather than at the level of isolated snippets (He et al., 18 Jul 2025). DeGenTWeb sharpens the definition further by excluding sites that are only AI-assisted or contain a mixture of human-written and AI-written pages unless the aggregate pattern across the site indicates that the content is overwhelmingly LLM-generated (He et al., 30 Apr 2026).

This boundary matters because many real corpora are hybrid. GigaCheck therefore separates two tasks: full-text classification, which asks whether an entire text is human-written or LLM-generated, and interval localization, which identifies LLM-generated spans within human-machine collaborative texts (Tolstykh et al., 2024). The parliamentary-text study adopts a similarly cautious operationalization: its classifier flags whether a text or paragraph appears to exhibit LLM-like characteristics, but does not quantify the precise amount of LLM use and does not claim that every flagged item is 100% machine-authored (Suvanto et al., 12 Jun 2026). A plausible implication is that “LLM-dominant” is best understood as a spectrum-sensitive category that becomes operationally stable only when paired with aggregation, disclosure, or localization procedures.

The concept also extends beyond open-web prose. In legal-video analysis, the LLM is the primary engine that turns multimodal evidence into summaries and complaint letters (Hoeben-Kuil et al., 15 Nov 2025). In legal-domain adaptation, Lawyer LLaMA places the LLM at the center of legal synthesis and reasoning, while retrieval serves as supporting evidence rather than replacing generation (Huang et al., 2023). In control engineering, the PyIRK pipeline uses Google Gemini as a semi-automated structured text processor, but retains manual review, making it closer to LLM-supported formalization than to fully autonomous content generation (Fiedler et al., 4 Nov 2025).

2. Detection methodologies

Detection research has converged on the view that naive page-level or sentence-level methods are inadequate for LLM-dominant content on the web. The 2025 website paper states that state-of-the-art LLM detectors perform well mainly on clean, prose-like text, while web content has complex markup and diverse genres (He et al., 18 Jul 2025). Its proposed remedy is a scalable pipeline that classifies entire websites by aggregating an LLM text detector’s outputs across multiple prose-like pages, rather than classifying text extracted from each page in isolation. Using two distinct ground-truth datasets totaling 120 sites, the paper reports 100% accuracies testing across them (He et al., 18 Jul 2025).

DeGenTWeb provides the most detailed website-level methodology. Its six-step pipeline samples pages, downloads HTML, extracts content, filters pages, scores them with Binoculars, and classifies a site using score deciles (He et al., 30 Apr 2026). Pages are loaded in a Chromium-based headless browser; Trafilatura extracts the main textual content; only English pages with at least 200 tokens are considered; duplicate-heavy pages are removed using Rabin fingerprinting; and a relaxed Dolma quality filter excludes non-prose, low-quality, or repetitive material (He et al., 30 Apr 2026). After scoring each page with Binoculars, DeGenTWeb represents a site by the 10th through 90th percentiles of its page scores and feeds that 9-dimensional vector to a linear SVM (He et al., 30 Apr 2026).

The methodological payoff is substantial. On a ground-truth dataset of 144 sites totaling 4,475 pages, DeGenTWeb reaches 98.7% accuracy with 15 pages per site in out-of-distribution cross-validation (He et al., 30 Apr 2026). Site aggregation also materially improves detector performance: for Binoculars, page-level best accuracy is 92.8%, while site-level accuracy is 100.0%; for Fast-DetectGPT, the corresponding numbers are 92.7% and 100.0% (He et al., 30 Apr 2026). Filtering is equally consequential: without filtering, the false positive rate on 2,354 pre-ChatGPT Common Crawl sites is 12.9%, whereas with filtering it drops to 0.30% (He et al., 30 Apr 2026).

Mixed-authorship detection requires different machinery. GigaCheck uses Mistral-7B-v0.3 with LoRA for text-level classification, and a DETR-like detector for character-level localization of LLM-generated intervals inside collaborative documents (Tolstykh et al., 2024). The interval detector predicts one or more character-level spans rather than a single global label, which is specifically designed for texts where human and machine authorship are interleaved (Tolstykh et al., 2024). This suggests that detection methodology is now bifurcating into at least two regimes: conservative site-level attribution for open-web content, and span-level localization for collaborative writing.

3. Web-scale prevalence and web ecology

The empirical literature treats LLM-dominant content as a measurable web phenomenon rather than an anecdotal trend. The 2025 preprint reports detecting a sizable portion of sites as LLM-dominant among 10k sites in search engine results and 10k in Common Crawl archives, and finds that such sites are growing in prevalence and rank highly in search results (He et al., 18 Jul 2025). DeGenTWeb extends that result with larger-scale measurements and stronger false-positive controls (He et al., 30 Apr 2026).

In Common Crawl, DeGenTWeb randomly samples 409,805 sites from January 2020 to May 2025 and retains 94,908 sites with at least 15 analyzable pages; 6.0% of these are classified as LLM-dominant (He et al., 30 Apr 2026). The share rises from 2.1% in the second half of 2022 to 29.4% in the first half of 2025 (He et al., 30 Apr 2026). For transitions within sites, 3,525 of 26,749 sites with both pre- and post-ChatGPT pages show a strong shift from more human-like to more LLM-like content, while random date shuffling gives only about 256 such sites on average (He et al., 30 Apr 2026).

Search results appear even more saturated. For 10,000 how-to-style queries based on WikiHow titles, the crawl yields 59,046 sites, of which 18,169 have at least 15 filtered pages; 15.4% of those are labeled LLM-dominant (He et al., 30 Apr 2026). At the query level, 46.6% of queries have at least one LLM-dominant site in the top 10 results, and 65.7% have one in the top 20 (He et al., 30 Apr 2026). Bing shows no strong evidence of down-ranking: median rank is 11.0 for LLM-dominant sites versus 10.0 for non-LLM sites, and average rank is 10.7 versus 10.2 (He et al., 30 Apr 2026).

The web ecology of LLM-dominant sites is not random. In Common Crawl, 78.8% of LLM-dominant sites have a clear financial incentive, versus 55.8% of non-LLM-dominant sites (He et al., 30 Apr 2026). The paper highlights service businesses and SaaS, including HVAC, plumbing, restoration, behavioral health, and businesses in emerging markets such as .ae, .pk, and .ph domains (He et al., 30 Apr 2026). LLM-dominant sites also favor cheaper ad stacks such as AdSense and Ezoic, while non-LLM-dominant sites more often use premium networks like Raptive/AdThrive, PubMatic, Sovrn, and Mediavine (He et al., 30 Apr 2026). Cluster analysis further suggests shared operators: one cluster contains 18 publisher/editorial sites under one AdSense ID, and another contains 31 identically looking publisher/editorial sites sharing a common Gatsby-based stack (He et al., 30 Apr 2026). This suggests that LLM-dominant content is frequently industrialized rather than merely casual or individual.

4. Mixed authorship, institutional writing, and domain-specific generation

Although websites are the most visible locus of LLM-dominant content, institutional and professional corpora show related dynamics. In parliamentary writing, an interpretable glass-box classifier trained on pre-LLM parliamentary texts and LLM-generated counterparts finds a steady increase in undisclosed LLM use in both the UK Parliament and the Swedish Riksdag from 2022 onward (Suvanto et al., 12 Jun 2026). On full-text classification, UK detection rates reach 0.0212 in 2026 and Swedish detection rates reach 0.0647 in 2025–2026 (Suvanto et al., 12 Jun 2026). At paragraph level, the rates are higher: 0.1545 for the UK in 2026 and 0.0943 for Sweden in 2025–2026 (Suvanto et al., 12 Jun 2026). The paper treats this primarily as a transparency and governance issue rather than as proof of wholly machine-authored parliamentary texts.

Collaborative writing creates a different profile. GigaCheck reports 0.943 accuracy on TweepFake, 100% F1 in all in-domain Ghostbusters settings, and 0.646 F1@3 on TriBERT for authorship-boundary detection (Tolstykh et al., 2024). Those results indicate that full-text detection and boundary localization are both technically feasible, but they also reinforce the distinction between LLM-dominant and mixed-authorship content. The paper explicitly cautions against using the model as a final arbiter in high-stakes decisions (Tolstykh et al., 2024).

Professional workflows show how LLM-dominant content can be useful while remaining error-prone. In the legal-video study, Gemini 2.5 Flash is the main system that watches video, processes audio, and produces either a 2–3 sentence summary or a longer complaint letter (Hoeben-Kuil et al., 15 Nov 2025). Across 120 YouTube videos, 37.5% of summaries are rated high quality, 34.2% medium, and 28.3% low; 64.2% are complete and 55.0% factual (Hoeben-Kuil et al., 15 Nov 2025). The legal-letter outputs are described as drafting assistance rather than autonomous legal analysis, because factual errors in the underlying summaries can bleed into the legal reasoning (Hoeben-Kuil et al., 15 Nov 2025).

In legal-domain adaptation, Lawyer LLaMA uses continual pre-training on legal corpora, supervised fine-tuning, and retrieval grounding to make the LLM itself the core legal reasoning engine (Huang et al., 2023). Retrieval substantially reduces hallucination: for responses mentioning legal articles, H1 falls from 64.8% to 25.9% and H2 from 60.2% to 14.8% (Huang et al., 2023). In control engineering, the PyIRK pipeline introduces Formal Natural Language as an intermediate controlled language, uses Google Gemini to convert LaTeX snippets into FNL, and then applies manual review and algorithmic conversion into PyIRK and interactive HTML; about 10% to 20% of the statements require manual intervention, and an eight-page test case produced around 700 tooltip elements (Fiedler et al., 4 Nov 2025). These cases show that LLM-dominant content often emerges inside structured human workflows rather than only as mass-generated public prose.

5. Moderation, calibration, and legitimacy

The growth of LLM-dominant content has made moderation itself a major research area. AEGIS treats moderation as moderation of human–LLM interactions, introduces a taxonomy with 13 critical risk categories and 9 sparse risk categories, and builds AegisSafetyDataset from approximately 26,000 human-LLM interaction instances (Ghosh et al., 2024). BingoGuard argues that binary moderation is insufficient, introduces per-topic severity rubrics for 11 harmful topics, and trains on 54,897 examples to predict binary safety labels and severity levels (Yin et al., 9 Mar 2025). In large-scale advertising moderation, Google Ads adopts a hybrid system in which heuristics, deduplication, clustering, and label propagation reduce 400 million ad images over 30 days to less than 0.1% for LLM review, cutting reviews by more than 3 orders of magnitude while achieving 2x recall compared to a baseline non-LLM model (Qiao et al., 2024). A plausible implication is that moderation at LLM-content scale is becoming an orchestration problem rather than a single-model classification problem.

Reliability is a separate issue. The calibration study on guard models evaluates 9 open-source LLM-based guard models on 12 benchmarks and finds that current systems tend to produce overconfident predictions, are significantly miscalibrated under jailbreak attacks, and show limited robustness to the outputs of different response models (Liu et al., 2024). Temperature scaling improves many response-classification settings, while contextual calibration is particularly useful when no validation set is available (Liu et al., 2024). Toxicity-classifier analysis reaches a related conclusion from a different angle: classifiers trained on human-authored moderation corpora are brittle on LLM-generated or adversarial toxic text, and suppressing vulnerable attention heads can improve robustness while revealing demographic-level fairness gaps (Furniturewala et al., 16 Sep 2025).

Normative work argues that moderation of LLM-dominant content cannot be judged by accuracy alone. The legitimacy framework distinguishes easy cases, where accuracy, speed, and transparency matter, from hard cases, where justification and participation are the core legitimacy criteria (Huang, 2024). AI Watchman extends the governance perspective to refusal behavior, treating refusal as a form of content governance and auditing more than 400 social issues over time across OpenAI’s moderation endpoint, GPT-4.1, GPT-5, and DeepSeek in English and Chinese (Dai et al., 24 Sep 2025). Reported refusal rates vary from 1.2% to 3.9%, and the system detects company- and model-specific changes, including changes that were not publicly announced (Dai et al., 24 Sep 2025). This suggests that LLM-dominant content is governed not only by what models generate, but also by what they decline to repeat, summarize, or surface.

6. Limitations, controversies, and research directions

The main controversy is epistemic rather than purely technical: as models improve, detection becomes harder even while the social cost of false attribution remains high. DeGenTWeb argues that when the goal is to minimize the chances of falsely attributing human-authored content to LLMs, detectors of LLM-generated text perform much worse than advertised (He et al., 30 Apr 2026). Its false-positive analysis shows three recurring error sources: filtering gaps, template-heavy pages, and formulaic domain-specific prose such as scripture, liturgical text, software documentation, and review-style publisher pages (He et al., 30 Apr 2026). The pre-ChatGPT Common Crawl evaluation yields only 0.29% classified as LLM-dominant across 35,856 sites, which the authors treat as an approximate upper bound on false positives (He et al., 30 Apr 2026).

The frontier is also moving quickly. On 80 synthetic sites generated with frontier models including GPT-OSS-120B, Claude Haiku 4.5, Claude Sonnet 4, and Claude Sonnet 4.6, DeGenTWeb accuracy degrades as the generating model becomes more capable, with a strong negative correlation of $r = -0.88$ between DeGenTWeb accuracy and the model’s Artificial Analysis Intelligence Index score (He et al., 30 Apr 2026). The average Binoculars score rises from 0.89 with Sonnet 4 to 0.96 with Sonnet 4.6, making the text look more human-like to the detector (He et al., 30 Apr 2026). The earlier website paper reaches the same strategic conclusion in shorter form: current detectors are insufficient for real web content, and reliable classification requires site-level aggregation over multiple prose-like pages (He et al., 18 Jul 2025).

Other limitations concern granularity, provenance, and human oversight. GigaCheck notes context-window limits, no multilingual benchmark evaluation, and imperfect robustness under paraphrasing and challenging domains (Tolstykh et al., 2024). The parliamentary study cannot localize beyond paragraphs and cannot measure mixed authorship directly (Suvanto et al., 12 Jun 2026). The legal-video study states that outputs are not reliable enough to be used without human review (Hoeben-Kuil et al., 15 Nov 2025). The control-engineering semantic-layer paper identifies manual correction as the current bottleneck and still relies on LaTeX source code rather than PDF-only ingestion (Fiedler et al., 4 Nov 2025). The legal-interpretation chapter generalizes the same caution: LLMs may dominate the workflow interface, but they should not replace human decision-making, especially in high-risk settings (Corbo, 10 Dec 2025).

The emerging research program therefore combines conservative attribution, workflow-aware disclosure, calibration, and governance analysis. This suggests that the most durable treatment of LLM-dominant content will not be a single detector or benchmark, but a layered regime that distinguishes AI-assisted from overwhelmingly machine-generated material, separates full-text dominance from mixed authorship, and treats disclosure, explanation, and institutional legitimacy as first-class technical requirements rather than afterthoughts.