CtrlK
AI-generated content refers to text, images, video, audio, music, and code produced by artificial intelligence systems, particularly generative AI models such as large language models (LLMs), diffusion models, and generative adversarial networks. The volume of AI-generated content on the internet has grown exponentially since 2022, driven by the widespread availability of tools like ChatGPT, Midjourney, DALL-E, and Stable Diffusion. By early 2025, researchers estimated that more than half of all newly published web content was at least partially machine-generated, and projections suggest that 90% or more of internet content could be AI-generated by 2026 [1].
The proliferation of AI-generated content has raised questions across nearly every domain it touches, from the reliability of search results and the integrity of academic work to the economic viability of creative professions and the trustworthiness of online information. It has also introduced new technical challenges around detection, labeling, and the long-term effects of training future AI models on outputs from previous ones.
AI-generated content spans virtually every media type. The following table summarizes the major categories and the models commonly associated with each.
| Content type | Description | Notable tools and models |
|---|---|---|
| Text | Articles, essays, marketing copy, social media posts, emails, and other written content | GPT-4, Claude, Gemini, Llama, Mistral |
| Images | Photographs, illustrations, art, product images, and other visual content | Midjourney, DALL-E, Stable Diffusion, Flux, Imagen |
| Video | Short clips, animations, advertisements, and synthetic footage | Sora, Runway Gen-3, Kling, Pika |
| Audio and music | Voiceovers, podcasts, sound effects, and full musical compositions | Suno, Udio, ElevenLabs, Bark |
| Code | Software programs, scripts, functions, and entire applications | GitHub Copilot, Cursor, Claude Code, Codex |
| 3D models | Objects, scenes, and environments for games or product design | Meshy, Point-E, Shap-E |
Each of these categories has its own ecosystem of models, workflows, and quality considerations. Text generation is the most widespread by volume, since large language models can produce written content at negligible marginal cost. Image generation has had the most visible cultural impact, sparking debates about copyright, artistic labor, and visual misinformation.
The growth of AI-generated content on the internet has been rapid and well-documented.
In April 2025, Ahrefs analyzed 900,000 newly published English-language web pages and estimated that 74.2% contained AI-generated content. This was a substantial increase from earlier estimates and reflected both the improving quality of generative models and the declining cost of producing content at scale [1].
Earlier projections by Europol and others had suggested that 90% of all online content could be AI-generated by 2026. While the exact percentage is difficult to pin down (detection methods are imperfect, and the definition of "AI-generated" varies), the trend is unmistakable. By 2025, the term "dead internet theory," which originated as a fringe conspiracy theory suggesting that most internet activity was bot-driven, had become something closer to a description of observed reality [2].
The scale of the phenomenon is driven by economics. Producing a 1,500-word article using a language model costs a fraction of a cent in API fees, compared to $50 to $500 or more for a human writer. For images, the cost differential is even starker. This makes AI-generated content attractive for any operation that prioritizes volume over quality: content farms, SEO operations, social media engagement bots, and advertising networks [1].
In March 2026, researchers at the cybersecurity firm DoubleVerify identified a network of more than 200 websites they called "AutoBait," which used templated prompts in a large language model to produce articles and images designed primarily to generate advertising revenue. The network was producing thousands of articles per day at virtually no cost [3].
The term "slop" emerged in 2024 as a colloquial label for low-quality AI-generated content that floods the internet, social media platforms, and search results. The word draws an analogy to unappetizing food waste, conveying the idea that AI-generated content is being produced carelessly and in excess [4].
Slop encompasses a wide range of content: AI-generated Facebook posts featuring bizarre images (often with telltale artifacts like extra fingers or nonsensical text), AI-written articles that contain plausible-sounding but fabricated information, AI-generated Amazon product reviews, and AI-created YouTube videos that recycle content without adding value. What unites these examples is not the use of AI per se, but the combination of AI generation with a lack of human oversight, editorial judgment, or genuine communicative intent [4].
The Oxford English Dictionary added "slop" in its AI-related sense in 2025, defining it as "low-quality AI-generated content, especially when produced in large quantities without human oversight." The term gained wide adoption in part because it filled a lexical gap: there was no concise, pejorative word for the specific kind of AI-generated content that degrades the information environment rather than enriching it [4].
Google's January 2025 update to its quality rater guidelines added definitions for AI content and authorized raters to assign the lowest quality score to "automated or AI-generated content" that provides no meaningful value to users. However, these manual reviews cannot keep pace with the volume of content being produced [1].
Detecting AI-generated content has become a significant technical and commercial challenge. Several approaches have been developed, each with distinct strengths and limitations.
| Tool / Method | Type | Approach | Strengths | Limitations |
|---|---|---|---|---|
| GPTZero | Commercial | Perplexity and burstiness analysis | 99.3% accuracy on unedited AI text; low false positive rate (0.24%) | Accuracy drops to 85-95% on heavily edited or paraphrased content |
| Originality.ai | Commercial | Neural classifier | Claims 99% accuracy; effective on paraphrased content | Higher false positive rate (up to 3% on Turbo model); 83% accuracy in independent testing |
| Turnitin AI Detection | Commercial (academic) | Integrated with plagiarism checker | Widely deployed in educational institutions | 84% overall effectiveness; 5-7% false positive rate on human-authored work |
| GLTR | Research tool | Statistical analysis of token probabilities | Visualizes per-token prediction rank; open-source | Based on GPT-2 (2019); less effective against newer models |
| Watermarking | Technical standard | Embeds invisible signals in AI output during generation | Provides provenance if widely adopted | Requires model provider cooperation; can be defeated by paraphrasing or post-processing |
| Statistical methods | Research | Analyze perplexity, burstiness, and entropy patterns | Model-agnostic; do not require access to the generating model | Less reliable for short texts; vulnerable to adversarial evasion |
GLTR (Giant Language Model Test Room) was one of the earliest detection tools, developed in 2019 by researchers from the MIT-IBM Watson AI Lab and HarvardNLP. It operates on the principle that AI-generated text tends to select highly probable tokens at each position, whereas human writing more frequently uses unpredictable word choices. GLTR visualizes the probability rank of each token, color-coding words by how likely they were under a language model's predictions. In a human-subjects study, GLTR improved the detection rate of AI-generated text from 54% to 72% [5].
More recent commercial tools like GPTZero and Originality.ai use neural classifiers trained on large datasets of human-written and AI-generated text. GPTZero, founded by Edward Tian in 2023, measures both perplexity (how surprising a text is to a language model) and burstiness (how much sentence complexity varies within a passage). Human writing tends to be "burstier," alternating between simple and complex sentences, while AI-generated text tends to be more uniform [6].
AI content detection faces several fundamental challenges that limit its reliability.
False positives for non-native speakers. Multiple studies have found that AI detection tools disproportionately flag text written by non-native English speakers as AI-generated. This happens because non-native writing tends to use simpler vocabulary and more predictable sentence structures, patterns that overlap with AI-generated text. In academic settings, this has raised serious fairness concerns, as students who write in a second language may be wrongly accused of cheating [6].
Paraphrasing defeats detection. Simple paraphrasing tools can reduce AI detection scores significantly. A student or content creator can run AI-generated text through a paraphraser, lightly edit the output, and produce text that most detection tools will classify as human-written. This arms race between generation and detection has no clear resolution [6].
Short texts are harder to classify. Detection accuracy drops substantially for shorter passages. Most tools require at least 250 to 300 words to make a reliable judgment, and even then, confidence intervals are wide for borderline cases.
Mixed content is ambiguous. Much real-world content is neither purely AI-generated nor purely human-written. A journalist might use an LLM to draft an outline, then rewrite every sentence. A student might use AI to generate ideas, then compose the essay independently. Detection tools struggle to handle this continuum between AI-assisted and AI-generated work.
Newer models are harder to detect. As language models improve, their outputs become more human-like, making statistical detection more difficult. Models trained specifically to avoid detection (or fine-tuned on diverse human writing styles) can produce text that evades current classifiers [7].
The flood of AI-generated content has complicated search engine optimization and the quality of search results. Google and other search engines have had to adapt their algorithms to handle the volume of low-quality AI content competing for rankings. Google's "Helpful Content Update" (2023) and subsequent algorithmic changes in 2024 and 2025 explicitly targeted AI-generated content that provides little original value. Despite these efforts, AI-generated spam and content farms continue to rank for competitive search queries [1].
Several news organizations have experimented with AI-generated articles, with mixed results. CNET faced a scandal in 2023 when it was revealed that dozens of articles published under a byline labeled "CNET Money Staff" had been generated by AI, many containing factual errors. Sports Illustrated faced a similar controversy when it was found to have published AI-generated product reviews under fabricated author names and AI-generated headshots. These incidents prompted industry-wide discussions about transparency and editorial standards for AI-assisted journalism [8].
The use of AI for academic writing has created a crisis in educational assessment. Surveys conducted in 2024 and 2025 found that a substantial percentage of university students reported using LLMs to assist with coursework, ranging from brainstorming and outlining to generating entire essays. Universities have responded with a mix of updated academic integrity policies, AI detection tools (primarily Turnitin), and a shift toward in-person assessments, oral examinations, and process-based evaluation [6].
AI-generated content has disrupted creative fields including illustration, graphic design, photography, voice acting, and music composition. The labor disputes of 2023, including the SAG-AFTRA and Writers Guild of America strikes in Hollywood, were partly motivated by concerns about AI replacing human creative work. The resulting contracts included provisions governing the use of AI-generated content in film and television production [8].
Social media platforms have become saturated with AI-generated content, from bot accounts posting AI-written text and AI-generated images to AI-driven engagement farming operations. Meta, X, and other platforms have implemented policies requiring disclosure of AI-generated content, though enforcement remains inconsistent. In January 2026, YouTube CEO Neal Mohan stated that reducing slop and detecting deepfakes were priorities for the platform [3].
Model collapse is a phenomenon in which machine learning models trained on data that includes outputs from previous models gradually degrade in quality. As AI-generated content becomes an increasingly large fraction of internet text, there is a growing risk that future language models will be trained partly on the outputs of their predecessors, creating a feedback loop that erodes data quality over time [9].
A 2024 peer-reviewed paper published in Nature demonstrated that large language models, variational autoencoders, and Gaussian mixture models all degrade when successive generations train on content produced by earlier models. The degradation begins subtly, with the model losing information about the tails of the distribution (minority and unusual data), and progresses to more severe distortion over repeated generations [9].
Researchers have found that even small fractions of synthetic data in training sets (as little as 1 in 1,000 samples) can contribute to model collapse under certain conditions. With over 74% of newly created web pages containing AI-generated text by April 2025, the practical risk to future training pipelines is substantial [9].
Mitigation strategies include capping the proportion of synthetic data in training sets, using classifiers to filter AI-generated content from training data, maintaining curated archives of pre-AI human-written text, and verifying synthetic data quality before inclusion. Some researchers have argued that if synthetic data accumulates alongside (rather than replacing) human-generated data, model collapse can be avoided, though the conditions under which this holds remain an active area of study [9].
Governments around the world have begun enacting rules that require AI-generated content to be labeled, disclosed, or watermarked.
The EU AI Act, formally adopted in March 2024, includes specific provisions for AI-generated content. Article 50 establishes transparency obligations for providers and deployers of AI systems that generate synthetic content. Providers of general-purpose AI systems with synthetic content generation capabilities were required to comply by May 2025, with all other systems following by May 2026 [10].
The European Commission has developed a Code of Practice on marking and labeling of AI-generated content, with a first draft published in late 2025 and a final version expected by June 2026. The Code provides guidance on a two-layered marking approach involving secured metadata and watermarking, optional fingerprinting and logging, and protocols for detection and verification [10].
China's content labeling rules, which took effect on September 1, 2025, require AI-generated content to carry visible labels. This applies to chatbot responses, AI-written text, synthetic voices, and face-generated or face-swapped content. Service providers offering generative AI must also conduct security assessments and register their large language models with the Cyberspace Administration of China [10].
The US lacks a comprehensive federal law on AI-generated content labeling, though several states have enacted or proposed their own requirements. California, Texas, and New York have all passed laws that include provisions related to AI content disclosure, particularly for political advertisements and deepfakes. Executive orders from both the Biden and Trump administrations have addressed AI-generated content in different ways, with the Biden order emphasizing labeling requirements and the Trump order prioritizing deregulation [11].
Watermarking, one of the primary technical approaches to content labeling, faces significant practical challenges. Text watermarking works by subtly biasing the model's token selection process during generation, creating a statistical pattern that a detector can identify. However, simple paraphrasing, translation, or re-encoding can remove these watermarks. Image and video watermarking is somewhat more robust, but can also be defeated by compression, cropping, or format conversion. Widespread adoption requires cooperation from all major AI model providers, and no single watermarking standard has yet achieved universal support [10].
The AI-generated content landscape in early 2026 is defined by several intersecting trends.
The volume of AI-generated content continues to grow. More than half of all new web content is at least partially machine-generated, and the cost of producing this content continues to fall as models become cheaper to run. Content farms, SEO operations, and social media bots account for much of the growth, but AI-assisted writing has also become routine in legitimate journalism, marketing, and corporate communications [1].
Detection remains an unsolved problem. While commercial tools like GPTZero and Originality.ai have improved, the fundamental arms race between generation and detection continues. No detection method achieves both high accuracy and low false positive rates across all conditions, and the gap between the best detectors and the best generators appears to be narrowing [6].
Regulatory frameworks are emerging but incomplete. The EU has the most comprehensive approach, with the AI Act's labeling requirements coming into full force by mid-2026. China has implemented its own labeling rules. The US remains fragmented, with a patchwork of state laws and no federal standard. International coordination on AI content labeling is still in its early stages [10].
The long-term consequences of a web dominated by AI-generated content are still unfolding. Model collapse, the erosion of trust in online information, the displacement of human creators, and the challenge of maintaining a shared factual reality in an era of cheap synthetic media are all problems that will intensify before they improve.