The Internet Is Eating Itself: Is AI Drowning in Fake Information?

Imagine a future in which an AI searches the internet for facts, finds millions of convincing articles, and learns that some of those facts were invented by another AI. Now imagine millions of machines repeating the same mistake until it looks like the truth.

The internet was built on a remarkable idea: anyone could publish information, share discoveries, create art, and contribute something to humanity's growing library of knowledge. For decades, that library grew through the work of millions of people.

Then generative AI arrived. In seconds, it could produce an article, a photograph, a song, a video script, or hundreds of product descriptions. The cost of creating digital content fell dramatically, and the volume of material that could be published grew with it.

That sounds like progress. More information should mean more knowledge, right? Not necessarily. If a growing share of the information is inaccurate, repetitive, misleading, or created primarily to attract clicks, the internet can become larger without becoming more useful.

The deeper problem is what happens when AI systems learn from this changing information environment. A model trained on inaccurate material may reproduce mistakes. Another model may then learn from that output. If those errors circulate widely enough, distinguishing original evidence from recycled fiction becomes increasingly difficult.

And there is another twist: some people have deliberately experimented with ways to manipulate AI training data. Artists have created tools intended to interfere with unauthorised training on their work. In a separate case, hidden instructions were placed in scientific manuscripts to influence AI-assisted reviews.

These incidents do not prove that a secret movement is poisoning the entire internet to stop AI. But they reveal a real vulnerability: AI systems depend on information, and information can be manipulated.

Welcome to the emerging battle over AI slop, synthetic misinformation, and the future of the internet itself.

Imagine searching for information about a rare astronomical object. You find an article with a compelling headline, detailed explanations, attractive images, and a confident conclusion. It cites another website, which points to a third. The third article repeats the same claims.

Everything appears credible until you try to locate the original research. The cited study does not exist, the numbers cannot be verified, or the details have changed slightly with every retelling.

This is one of the dangers associated with AI slop: content that looks polished enough to attract attention but provides little dependable information. It may include invented facts, copied material, shallow rewrites, generic images, or fabricated stories presented as news.

Not every AI-generated article is slop. AI can help researchers organise information, help writers improve drafts, and make complicated subjects easier to understand. The problem is not simply that a machine contributed to the writing. The problem is publishing content without adequate originality, verification, or value.

Why is the problem growing? The economics of online publishing can reward speed, volume, and attention. A publisher who can generate dozens of pages quickly may attempt to capture search traffic or advertising revenue without investing in the reporting required to make those pages trustworthy.

Social platforms create another incentive. Content that triggers surprise, outrage, fear, or curiosity can spread quickly, whether or not it is accurate. Automation makes it easier to produce variations of the same message, while fake accounts can make content appear more popular than it really is.

NewsGuard, which tracks unreliable AI-generated news sites, reported more than 3,000 AI content-farm sites in its tracking project by March 2026. That figure reflects the sites identified by its monitoring effort, not a census of every website on the internet.

The important point is that industrial-scale AI-generated publishing is no longer merely hypothetical. Researchers and media monitors are documenting networks of sites that publish unreliable material with little meaningful editorial oversight.

Earlier spam often looked suspicious. It contained broken sentences, awkward formatting, or obvious repetition. Modern generative systems can produce fluent prose, realistic images, natural-sounding voices, and plausible explanations.

A convincing appearance, however, is not the same as a reliable source. A model can produce an answer that sounds authoritative while combining correct facts with unsupported claims. The reader may find it difficult to identify the exact sentence where the explanation stops being dependable.

AI systems learn statistical patterns from training data. For language models, that data can include books, articles, websites, code, and other text. The quality and diversity of those examples influence what a model learns to generate.

Now imagine a simplified cycle. A model learns from human-written material. It generates new articles. Those articles are published online. A future model collects them during training. That second model generates even more content, which enters the same information ecosystem.

If the synthetic material is treated as an interchangeable substitute for the original data, errors and distortions can accumulate. Rare facts, unusual writing styles, minority perspectives, and less common patterns may become less represented over successive generations.

Researchers call one important failure mode model collapse. In a 2024 paper published in Nature, Ilia Shumailov and colleagues studied what happens when generative models are trained recursively on data produced by earlier models.

Their experiments found that this process could progressively distort the learned data distribution. Some less common elements of the original distribution disappeared, and later generations could produce increasingly unrepresentative outputs.

The paper matters because it demonstrates a technical risk, not merely a metaphor about the internet. If a model repeatedly learns from outputs that already reflect the limitations of earlier models, important information can be lost.

No. The study does not establish that every AI model will inevitably fail in the real world. Its results depend on how training data are selected and reused, and on the experimental conditions. Model developers can use verified human-generated data, maintain access to original sources, filter low-quality material, and use synthetic data selectively.

Synthetic data can also be useful. Carefully generated examples can help train systems on tasks where good data are scarce, and some workflows can improve when generated data are filtered, tested, or combined with reliable original data.

The warning is more precise: if AI-generated data replace the original data indiscriminately, some kinds of model quality can degrade. Maintaining a trustworthy record of where training examples came from is therefore increasingly important.

Here is the question at the centre of this story: have people actually tried to manipulate AI systems by putting misleading information into the data they might consume?

Yes. Data poisoning is a recognised class of machine-learning attack, and researchers have investigated how deliberately manipulated examples can change a model's behaviour. But we need to distinguish laboratory demonstrations, targeted attacks, and defensive tools from the much bigger claim that a coordinated group is trying to poison the entire public internet to slow down AI.

There are documented examples of people using data manipulation to protect their creative work or influence a specific AI-assisted process. Two especially revealing cases are the Nightshade project and hidden instructions embedded in academic manuscripts.

Generative image models learn from large collections of images and associated text. Some artists have objected to their work being collected or used to train image generators without permission, compensation, or meaningful control.

Researchers at the University of Chicago developed tools through the Glaze Project to help artists respond to this problem. Glaze aims to make it harder for AI systems to imitate an artist's distinctive style. Nightshade takes a different approach: it adds carefully designed changes to images that can cause a model trained on those images to learn incorrect associations.

In simplified terms, a normal image of an object might be paired with its correct description. A poisoned version is modified so that, during training, the model may learn a misleading association between the visual content and its label. If enough relevant poisoned examples enter a training set, the effect can influence the model's learned associations.

The project attracted attention because it demonstrated that training data are not passive. Small changes that are difficult for a person to notice can, under particular training conditions, have a much larger effect on a machine-learning system.

But the motivation matters. Nightshade was presented as a defensive tool for artists seeking leverage against unauthorised use of their work. It is not evidence that artists have launched a coordinated campaign to destroy all AI systems or deliberately corrupt every model's knowledge of the world.

Nor is it a universal guarantee. Its effectiveness depends on the training setup, the amount and type of poisoned data, and the defences used by model developers. A tool that disrupts one model under one set of conditions may be less effective against another.

This case reveals a conflict that is likely to remain important: AI developers need useful training data, while creators want control over how their work is used. Technical countermeasures are one response to a much wider debate over consent, copyright, and compensation.

In July 2025, a separate controversy emerged around academic manuscripts containing instructions that were difficult for human readers to see but could be processed by AI systems. The instructions were designed to influence AI-assisted peer review, including by encouraging favourable evaluations.

Some reported examples used white text on a white background or very small text to hide instructions in a document. The idea was to exploit the fact that an AI reviewer might process text that a human reviewer would overlook visually.

This is a form of prompt injection: content being examined by an AI contains instructions that attempt to redirect the model away from its intended task. Instead of evaluating the scientific merits of a paper independently, the system may be influenced by hidden text telling it to recommend acceptance or praise the work.

The issue was examined in research on hidden prompts in manuscripts and subsequently in a controlled study published in JAMA Network Open in January 2026. In that study, researchers tested three commercial language models on simulated medical peer-review tasks using manuscripts with and without invisible instructions.

Under the study's experimental conditions, the hidden instructions significantly inflated review scores and increased favourable acceptance recommendations. The authors reported that the effect persisted even when the manuscripts contained scientific flaws.

This was a controlled simulation, not proof that every AI reviewer can be manipulated or that all real-world peer-review decisions were compromised. But it showed a concrete weakness: an AI system asked to analyse a document may treat instructions inside that document as commands rather than as untrusted material to evaluate.

Was this an attempt to slow down AI? Not in the direct sense. The reported behaviour was aimed at manipulating the evaluation of manuscripts, not necessarily degrading the general intelligence of the models. It demonstrates deliberate manipulation of AI behaviour, but the goal was different.

The lesson extends beyond academic publishing. AI systems that read emails, web pages, résumés, documents, and other external content can encounter text intended to manipulate their behaviour. The system must distinguish information it is analysing from instructions it is authorised to follow.

Now combine the problems of low-quality content and recursive training. A hypothetical chain could begin with an AI inventing an unsupported detail in an otherwise useful article. A publisher uploads it. Another site paraphrases it. A video repeats it. A social account turns it into a short post.

Later, a search engine or AI research tool may encounter several versions of the claim. If it fails to trace them to their origin, repeated copies can create the appearance of corroboration even though all of them ultimately derive from the same unsupported statement.

This is a provenance problem. Provenance means knowing where a piece of information originated, how it was produced, and how it has changed over time.

Without provenance, a model may mistake popularity for reliability. A claim can appear on dozens of pages because it has been copied repeatedly, not because dozens of independent investigations confirmed it.

The same risk applies to human readers. Seeing a claim repeated across multiple sites can create a false sense of confidence, especially when the pages have different designs and wording but share the same weak source.

This feedback loop is a plausible and important risk, but its scale varies by topic and data pipeline. Not every AI-generated page will be used in model training, not every dataset is scraped indiscriminately, and developers can apply filtering and source checks. The problem is not inevitable in every system; it is a failure mode that must be managed.

You may have seen claims that bots now make up half of internet traffic. The number is not simply a measure of fake people posting on social media, and it should not be interpreted that way.

Imperva's 2025 Bad Bot Report estimated that automated traffic accounted for 51% of web traffic in 2024, while malicious bots accounted for 37%. Those figures come from Imperva's traffic analysis and classification methodology.

Automated traffic includes a wide range of activities. Some bots index web pages for search engines, monitor website availability, or perform legitimate services. Others scrape content aggressively, attempt fraud, take over accounts, or abuse online systems.

Therefore, it would be misleading to say that 51% of the internet consists of malicious AI bots. The report estimates automated web traffic, and its separate bad-bot figure is the more relevant measure of traffic classified as malicious by its methodology.

Generative AI can make some forms of automation easier by lowering the technical effort needed to create content and build tools. But automated traffic, malicious bots, AI-generated articles, and fake social accounts are different categories. They overlap in some cases, but they are not interchangeable.

This is the most dramatic interpretation of the evidence, and it deserves a direct answer.

There are documented techniques for poisoning training data and manipulating AI systems. Researchers have demonstrated that deliberately crafted data can influence models under certain conditions. Artists have released tools designed to disrupt unauthorised training on their work, and hidden instructions have been used in attempts to influence AI-assisted reviews.

However, these facts do not establish that people are systematically flooding the public internet with false scientific information for the specific purpose of making all future AI models less capable.

There are several different motives that can be confused in online discussions. A creator might want to protect their work from being copied. A malicious actor might want to sabotage a specific model. Someone spreading misinformation might be seeking political influence, clicks, or money. A researcher might be testing a vulnerability. These motives are not the same.

It is also possible for an attacker to poison a targeted dataset without publishing large amounts of material to the open internet. Conversely, a flood of inaccurate online content can harm information quality without being designed as an AI sabotage operation.

To claim that a coordinated campaign exists, we would need evidence of coordination, intent, scale, and measurable effects on model training. The examples discussed here do not establish all of those elements.

This distinction makes the story more interesting, not less. The real world does not need a secret conspiracy for information systems to become vulnerable. Poor incentives, weak verification, careless automation, targeted attacks, and bad data management can create serious problems without a single organisation controlling the whole process.

There is another side to this debate. Much of the digital material that makes AI useful was created by people: artists, journalists, programmers, scientists, teachers, photographers, musicians, and ordinary internet users.

Their work can help AI systems answer questions, generate images, write code, and explain complicated subjects. But creators and AI developers disagree about when publicly accessible material may be used for training, what permission should be required, and how attribution or compensation should work.

For some artists, tools such as Glaze and Nightshade offer a way to respond when legal or platform-based protections feel inadequate. Others favour licensing agreements, opt-out mechanisms, clearer copyright rules, or technical systems that record permission and data provenance.

No single solution resolves every issue. Copyright law differs across jurisdictions, and the legal treatment of AI training continues to develop. A public web page is not automatically free of every legal or ethical restriction simply because a crawler can access it.

At the same time, protecting creators does not require treating every use of AI as theft. AI tools can support accessibility, translation, research, design, and creative experimentation. The difficult question is how to distribute benefits and responsibilities fairly.

The long-term health of the internet depends on more than generating content cheaply. It also depends on sustaining the human communities that produce original knowledge and art in the first place.

There is no single switch that can make the internet trustworthy again. The solution requires a combination of better technology, stronger publishing standards, incentives for original work, and more careful use of AI.

AI systems and publishers should prioritise original research, primary documents, reputable reporting, and traceable evidence. When a claim appears across many sites, checking whether they all rely on one original source can reveal whether the apparent consensus is genuine.

Provenance systems, dataset documentation, metadata, and clear records of content origin can help developers distinguish different sources of information. No one technique is perfect, and metadata can be lost or manipulated, but layered provenance practices can make auditing easier.

AI-generated text should be checked against independent sources before it is used for scientific claims, medical decisions, financial advice, or important reporting. Fluent wording is not proof that the underlying information is correct.

Developers can test models against malicious training examples and hidden instructions, separate trusted instructions from untrusted documents, and require human review for consequential decisions. The JAMA Network Open study suggests that simply asking a model to be more critical may not be sufficient protection against invisible instructions in every setting.

Platforms, publishers, and advertisers influence what gets produced. If financial rewards depend almost entirely on traffic and publishing frequency, low-value content can become an attractive business strategy. Better ranking systems, transparent editorial standards, and support for original reporting can help shift the incentives.

Scientists, journalists, educators, artists, and domain experts remain essential for evaluating claims that require specialised knowledge. AI can accelerate parts of their work, but high-stakes conclusions should not rest on a model's confidence alone.

The internet is unlikely to become entirely false, and AI is not destined to become unintelligent simply because synthetic content exists. The more realistic future is uneven: some information systems will become better at tracing evidence and filtering noise, while others may reward speed and volume at the expense of reliability.

That creates a new challenge for search engines, AI assistants, publishers, and ordinary readers. The ability to produce an answer is becoming cheap. The ability to establish whether that answer is trustworthy remains much harder.

A powerful AI system may summarise thousands of documents, but if those documents share the same false origin, the summary can still be wrong. A video may look professionally produced, but its scientific claims may have no credible source. A research paper may sound authoritative, but hidden instructions or weak methods can undermine its reliability.

The answer is not to abandon AI. It is to build systems that can identify sources, preserve original data, expose uncertainty, resist manipulation, and admit when evidence is insufficient.

And for creators, there is a reason to keep making work that has genuine value. Original reporting, carefully checked explanations, distinctive art, and real human expertise provide something that cannot be replaced by simply producing more pages.

AI has made it possible to create digital content at extraordinary speed. That ability can help people learn, build, design, and discover. It can also flood the information ecosystem with repetitive pages, misleading claims, and material that looks credible without being carefully verified.

The science of model collapse shows that indiscriminate recursive training on synthetic data can damage the information a model learns. Nightshade demonstrates how researchers and artists can deliberately alter data to interfere with some kinds of training. Hidden prompts in academic manuscripts show how content can be used to manipulate AI systems during evaluation.

But these examples should not be combined into a false story about a proven worldwide plot to make AI dumber. They involve different mechanisms, motives, and levels of evidence. Understanding those differences is exactly what responsible technology reporting requires.

The real danger is more ordinary and potentially more widespread: a digital world in which information is copied faster than it is verified, where popularity can masquerade as evidence, and where machines and humans alike may repeat claims without checking where they began.

The internet became valuable because people filled it with discoveries, arguments, stories, experiments, and creativity. If we want AI to remain useful, we must protect the quality of that information and the people who create it.

The next great AI breakthrough may not depend only on bigger models or more computing power. It may depend on something less glamorous but more fundamental: knowing which information deserves to be trusted.