31 July 2025
7 min read

The internet is filling up with AI content and nobody can tell anymore

By mid-2025, a practical problem had emerged that the AI industry had discussed theoretically for years: the web was filling up with AI-generated content at a rate that was starting to affect the quality of information retrieval. Not just spam or low-effort articles, but genuine, detailed, well-formatted text that was plausible enough to pass casual inspection but that cited sources that did not exist, contained errors presented confidently, and was contributing to what researchers were calling model collapse in downstream training data.

The volume problem

Content farms have existed for decades. The difference in 2025 was the combination of quality and scale. A single person with access to a mid-range language model could generate hundreds of articles per day at a quality level that would have required a team of writers a year earlier. The marginal cost of content production approached zero.

Research published in mid-2025 estimated that a significant portion of new web content published in early 2025 was AI-generated or AI-assisted to a degree that made it effectively synthetic. Precise figures are hard to establish because detection is imperfect, but the direction was unambiguous. More content was being published than ever before, and less of it was grounded in direct human experience or original research.

~200M
New web pages indexed weekly (est.)
$0
Marginal cost per AI article
92%
AI detector false-positive rate on human writing
3x
Growth in content volume 2023-2025
Fig. 1. AI content flood scale — volume, cost, detection failure rate, and growth.

Why detection does not work well

AI detection tools had a fundamental credibility problem by 2025. Every study that evaluated them found high rates of both false positives (classifying human-written text as AI-generated) and false negatives (classifying AI text as human). The false positive problem was particularly damaging because students and writers were being accused of using AI tools for work they had written themselves.

The deeper issue is that detection is an asymmetric problem. An AI system trained to avoid detection can generate text that evades classifiers. A detection system trained on that text improves and catches the next generation of evasion. The cycle never resolves in favour of detection because the generator has an inherent advantage.

Turnitin, one of the largest academic plagiarism detection services, reported in 2025 that its AI detection feature had been used to flag over 22 million papers. It also acknowledged that students with demonstrated false positives had legitimate grounds to appeal, which its administrators had no reliable way to evaluate.

Watermarking: the technical alternative

The technically more promising approach is watermarking, where a statistical signal is embedded in AI-generated text at generation time. Google's SynthID technology, initially developed for AI-generated images, was being adapted for text. Several AI labs were working on similar approaches. The idea is that rather than trying to detect AI content after the fact, you embed a traceable signal during generation.

ItemValue
AI detectors
How: Classify text using trained classifiers
Limitation: High error rates, gameable
Statistical watermarking
How: Embed token bias patterns at generation
Limitation: Editing removes the signal
Cryptographic watermarks
How: Embed hash in generation metadata
Limitation: Requires API, not for open models
Provenance metadata
How: C2PA content credentials on files
Limitation: Easy to strip, voluntary adoption
Human attestation
How: Signed verification of human authorship
Limitation: Not scalable, trust problem
Fig. 2. Approaches to AI content provenance — how each works and its primary limitation.

The SEO and creator economy impact

Google's search quality team was clearly aware of the problem and was updating its ranking algorithms to deprioritise thin content that matched patterns associated with AI generation. This created a peculiar incentive structure: sites with AI content had a declining return on that content over time as Google devalued it, which pushed them to produce more content to compensate, which degraded quality further.

For human content creators, the situation was demoralising in a specific way. Writing a detailed, expert article that took a week to research and compose now competed in search rankings against a thousand AI-generated articles on the same topic. The market for undifferentiated informational content had effectively collapsed. The value of genuinely original reporting, primary research, and expert analysis remained, but it required producers to find direct audiences rather than relying on search discovery.

Model collapse: the long tail problem

Researchers had been warning since 2023 about a phenomenon called model collapse: if AI models are trained on data that increasingly consists of AI-generated content, performance degrades over generations because the training data loses the diversity and specificity of human-produced information. By 2025 this was no longer a theoretical concern. Several research groups had demonstrated degradation empirically on models trained with progressively higher proportions of synthetic data.

The uncomfortable conclusion is that the viability of the next generation of AI models depends partly on the continued existence of a web with substantial human-authored content. The AI industry was, in a sense, consuming the resource it depended on. The labs were aware of this and were exploring agreements with news organisations and publishers for access to verified human-authored text, though the economic terms of these agreements remained contentious.