cs.CYApr 14, 2026

The Impact of AI-Generated Text on the Internet

Authors: Jonas Dolezal, Sawood Alam, Mark Graham, Maty Bohacek

Organizations: Imperial College London · Internet Archive · Stanford University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

The proliferation of AI-generated and AI-assisted text on the internet is feared to contribute to a degradation in semantic and stylistic diversity, factual accuracy, and other negative developments (sometimes subsumed under the Dead Internet Theory). What has hindered answering these questions is that it has not been understood just how much of the internet is actually AI-generated or AI-edited. To this end, we construct a representative sample of websites published on the internet between 2022 and 2025 using the Internet Archive, and apply a state-of-the-art AI text detector on them. We find that by mid-2025, roughly 35% of newly published websites were classified as AI-generated or AI-assisted, up from zero before ChatGPT's launch in late 2022. We also find statistically significant evidence for some of the identified hypotheses; for example, that increases in AI-generated text on the internet correlate negatively with semantic diversity and positively with the prevalence of positive sentiment. We do not, however, find statistically significant evidence supporting the hypothesis that an increased rate of AI-generated text on the internet decreases factual accuracy or stylistic diversity. Notably, this diverges from public perception, which we measure in a user study, where the majority of US adults turned out to believe in all four of the above-mentioned hypotheses. Individuals who do not use AI or use it infrequently tend to believe in these negative impacts more than those who use it frequently; similarly, individuals who hold negative views of AI tend to believe in these hypotheses more than those with favorable views of the technology.

Figures & tables

Figure 1: AI-generated Text on the Internet from Mid-2022 to Mid-2025. The figure shows the proportion of websites classified as fully AI-generated (red) and AI-generated or AI-assisted (purple) based on Pangram v3 detection applied to representative samples obtained from the Internet Archive. Curves represent smoothed estimates.
Figure 2: Results for Hyp. 1: Semantic Contraction. The figure shows results for the Semantic Contraction Hypothesis from the participant study (RQ1) and quantitative analysis of randomly sampled websites from the Internet Archive (RQ3). In (a), the average pairwise cosine similarity of semantic embeddings is plotted against AI Likelihood score, which combines the rate of AI-generated and AI-assisted samples, as detected by Pangram v3 ( ρ=0.47\rho=0.47 , p=0.004p=0.004 ). The overall results of the participant study are shown in (b), with responses ranging from Strongly Disagree (SD) to Strongly Agree (SA). These are broken down by AI usage frequency in (c) and general view of AI impact in (d).
Figure 3: Results for Hyp. 3: Positivity Shift. The figure shows results for the Positivity Shift Hypothesis from the participant study (RQ1) and quantitative analysis of randomly sampled websites from the Internet Archive (RQ3). In (a), the rate of positive documents (classified by sentiment analysis) is plotted against AI Likelihood score, which combines the rate of AI-generated and AI-assisted samples, as detected by Pangram v3 ( ρ=0.56\rho=0.56 , p=0.0003p=0.0003 ). The overall results of the participant study are shown in (b), with responses ranging from Strongly Disagree (SD) to Strongly Agree (SA). These are broken down by AI usage frequency in (c) and general view of AI impact in (d).
Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Length Sensitivity Analysis. Detection scores for (a) Binoculars, (b) Desklib, (c) DivEye, and (d) Pangram v3 across varying text lengths. Except for DivEye, these detectors generally achieve strong performance for texts exceeding 100 words.
Figure 5: HTML vs. Plain Text Robustness. Detection score distributions for AI-generated text presented in plain text versus HTML-embedded format, for (a) Binoculars, (b) Desklib, (c) DivEye, and (d) Pangram v3. Binoculars and DivEye exhibit a substantial drop in detection rate when text is embedded in HTML, Desklib exhibits some drop, and Pangram v3 exhibits minimal drop in scores.
Figure 6: Model Family Robustness. Detection score distributions for text generated by GPT-4o, Claude, and Gemini, for (a) Binoculars, (b) Desklib, (c) DivEye, and (d) Pangram v3. Binoculars performs differenly on Claude, GPT, and Gemini model families; Desklib and Pangram v3 have a similar performance across the board.
Figure 7: Model Version Robustness. Detection scores across OpenAI model generations for (a) Binoculars, (b) Desklib, (c) DivEye, and (d) Pangram v3. Binoculars and Desklib perform consistently across the board; DivEye and Pangram v3 perform differently on completion models ( davinci-002 and babbage-002 ) compared to instruction-trained models (GPT-3.5, GPT-4, and GPT-4o).
Figure 8: Language Robustness. Comparison of detector score distributions across languages for Binoculars, Desklib, DivEye, and Pangram v3 (1/3).
Figure 9: Language Robustness. Comparison of detector score distributions across languages for Binoculars, Desklib, DivEye, and Pangram v3 (2/3).
Figure 10: Language Robustness. Comparison of detector score distributions across languages for Binoculars, Desklib, DivEye, and Pangram v3 (3/3).
Figure 11: Welcome Screen. The fact-checkers are presented with a list of expectations and asked to provide their Prolific IDs.
Figure 12: Instruction Recap Screens. The fact-checkers are required to read through detailed instructions. The bottom of the screen is pictured.
Figure 13: Item Annotation View. Each claim for an article asks the fact-checkers to provide the verdict (Supported, Refuted, Not Enough Evidence, or Conflicting Evidence), evidence URLs, optional explanation, and confidence (range between 1, least confident, and 5, most confident). The top of the view indicates the progress within the assigned batch of articles.
Figure 14: Completion Screen. Upon completion, the fact-checkers are presented with the confirmation code for Prolific and instructions asking them to return to Prolific.
Figure 15: Results for Hyp. 2: Truth Decay. The figure shows results for the Truth Decay Hypothesis from the participant study (RQ1) and quantitative analysis of randomly sampled websites from the Internet Archive (RQ3). In (a), the average factual error rate is plotted against AI Likelihood score, as detected by Pangram v3 ( ρ=−0.19\rho=-0.19 , p=0.27p=0.27 ). The overall results of the participant study are shown in (b), with responses ranging from Strongly Disagree (SD) to Strongly Agree (SA). These are broken down by AI usage frequency in (c) and general view of AI impact in (d).
Figure 16: Results for Hyp. 4: Epistemic Islands. The figure shows results for the Epistemic Island Hypothesis from the participant study (RQ1) and quantitative analysis of randomly sampled websites from the Internet Archive (RQ3). In (a), the outbound link density (links per 1,000 words) is plotted against AI Likelihood score, as detected by Pangram v3 ( ρ=−0.12\rho=-0.12 , p=0.48p=0.48 ). The overall results of the participant study are shown in (b), with responses ranging from Strongly Disagree (SD) to Strongly Agree (SA). These are broken down by AI usage frequency in (c) and general view of AI impact in (d).
Figure 17: Results for Hyp. 5: Entropy Dilution. The figure shows results for the Entropy Dilution Hypothesis from the participant study (RQ1) and quantitative analysis of randomly sampled websites from the Internet Archive (RQ3). In (a), the Gzip compression ratio is plotted against AI Likelihood score, as detected by Pangram v3 ( ρ=−0.02\rho=-0.02 , p=0.89p=0.89 ). The overall results of the participant study are shown in (b), with responses ranging from Strongly Disagree (SD) to Strongly Agree (SA). These are broken down by AI usage frequency in (c) and general view of AI impact in (d).
Figure 18: Results for Hyp. 6: Stylistic Monoculture. The figure shows results for the Stylistic Monoculture Hypothesis from the participant study (RQ1) and quantitative analysis of randomly sampled websites from the Internet Archive (RQ3). In (a), the average pairwise character 3-gram Jaccard similarity is plotted against AI Likelihood score, as detected by Pangram v3 ( ρ=0.24\rho=0.24 , p=0.17p=0.17 ). The overall results of the participant study are shown in (b), with responses ranging from Strongly Disagree (SD) to Strongly Agree (SA). These are broken down by AI usage frequency in (c) and general view of AI impact in (d).

Explore similar work

CardsList

  1. The uneven rise of AI-generated web content

    Oct 27, 2026Adam Visokay, Soumyakanti Pan, Molly Griston +10Generative AI Adoption

    School of Information, University of California, Berkeley · Department of Statistics, University of Washington · Department of Physics, University of Colorado, Boulder +7

  2. "That's AI Slop, You Bot!" Studying Accusations, Evidence, and Credibility in Online Discourse Towards LLM-Generated Comments

    Jun 10, 2026Jason Miklian, John E. Katsos

    University of Oslo · American University of Sharjah

  3. 'Your AI Text is not Mine': Redefining and Evaluating AI-generated Text Detection under Realistic Assumptions

    Jun 3, 2026Nils Dycke, Marina Sakharova, Nico Daheim +1AI-Generated Text DetectionGenerative AI Evaluation

    Ubiquitous Knowledge Processing Lab (UKP Lab), Department of Computer Science, Technical University of Darmstadt · National Research Center for Applied Cybersecurity ATHENE, Germany · Zuse School ELIZA