Dead Internet Theory: Bots, AI Content and the Trust Problem

Watch the first Theories Elucidated documentary

The Truth About The Dead Internet: It’s Worse Than You Think is the first public video on @theorieselucidated. This article follows its central questions while adding qualifications to statistics and analogies that can otherwise sound more certain than the evidence supports.

The illusion: bot traffic measures how many people are real

Dead Internet Theory claims that automated activity and synthetic content have displaced much of genuine online interaction. Documented bot activity is real; the sweeping claim that authentic human participation has largely vanished is not established by a traffic percentage.

Imperva’s 2016 report reported a 51.8% bot share of measured online traffic. Traffic requests are not people, social accounts, comments or packets across every global fibre cable. Search crawlers and other useful automated services also contribute requests. That number cannot establish that more than half of internet users were fake.

Trap one: cheap production can reward volume over verification

The documentary’s economic argument is that inexpensive content and engagement can be produced at scale. When publishers or advertisers reward raw volume, unreliable material can compete for attention. That is a plausible incentive mechanism, not proof that every advertisement is served to a bot or that every automated publisher earns a profit.

A polished paragraph and a busy comments section do not establish independent expertise. Conversely, awkward prose does not prove a writer is human. Style alone is an unreliable detector. Follow specific claims back to original reporting, data or documents.

Trap two: repeating generated data can narrow a model’s view

Shumailov and colleagues’ 2024 Nature paper demonstrates model collapse under recursive training conditions: learned distributions can lose information, including rare outcomes. The study supports concern about indiscriminate reuse of generated material. It does not establish a universal rule that 10% synthetic data inevitably destroys every AI model.

Gerstgrasser and colleagues’ 2024 study found that retaining original real data while accumulating synthetic data could avoid collapse in their experiments. Training method, data quality, filtering and evaluation matter. The documentary’s photocopier analogy is useful for understanding accumulated distortion, but it is not a law covering every training setup.

What the headline numbers cannot establish

A forecast that synthetic content might reach a certain share is different from a measured count of all content now online. We do not treat the documentary’s 90% claim as an established global statistic. Likewise, text published before late 2022 is not guaranteed to be free of automation, fabricated material or machine-generated content.

An old archive can preserve valuable evidence, but its date does not make every statement true. Original sources still need context, methodology and checks against other evidence.

The actionable rule: follow one source chain today

Choose one claim in your feed. Identify who first made it, find the underlying document and check its date, sample and units. Ask whether later articles added evidence or simply repeated each other. A number without a denominator or method is a reason to investigate further.

For creators, show the documents, calculations and uncertainties behind your work. For readers, keep a small set of sources with a record of useful corrections. Human judgment remains valuable when it is made visible through evidence. Watch the documentary and follow Theories Elucidated for the full discussion.

Educational content, not financial advice.

Keep exploring

AI Voice Cloning Scams: How to Verify an Emergency Call · Agentic AI in Finance: What Changes for Entry-Level Jobs? · Scarcity Marketing: Why Waitlists Make Products Feel Valuable

Sources

• Theories Elucidated — first documentary, original video
• Imperva — Bot Traffic Report 2016
• Shumailov et al. — AI models collapse when trained on recursively generated data, Nature, 2024
• Gerstgrasser et al. — Is Model Collapse Inevitable? 2024