Maya posts a photo of a yellow raincoat. Seventeen strangers praise it within minutes. None of them mentions the raincoat. So she posts the same photo again and writes that the coat is red — and the replies come back exactly the same: gorgeous, inspiring, so true. Nobody corrects her. The crowd is reacting to the shape of a post, not to what is inside it.
This evidence-led investigation takes the "dead internet theory" seriously enough to check it properly — and that means separating the frightening fact from the more frightening exaggeration.
Imperva measured over 53% of web traffic as automated in 2025. That is a real and historic change, but traffic is not people: one crawler can request thousands of pages while a human reads three. "53% of traffic is automated" is not "53% of the internet is fake."
What is documented: OpenAI's 2025 description of an influence operation generating batches of comments in English and Taglish across five TikTok channels; Meta removing over 100 million fake Pages and 23 million impersonating profiles in 2024; and the Nature study by Shumailov and colleagues showing that models trained indiscriminately on generated data can suffer "model collapse" — rare details vanish first, the average takes over.
The uncomfortable conclusion isn't a secret control room. It's that no secret meeting is needed: content farms, marketers, political operators, platforms and AI companies each follow a rational incentive, and together they build an irrational environment.
Maya is a dramatized composite built from documented patterns. Her apartment is fictional. The machinery behind her screen is not.
00:00 Seventeen strangers and a yellow raincoat
00:38 The reply that appears everywhere
00:59 She changes the colour — nobody notices
01:16 An account that never sleeps
01:40 The 2021 manifesto
02:04 Crowded, yet strangely empty
02:26 The 53% number — and its footnote
02:48 What it does reveal
03:31 When the synthetic crowd has a name
03:56 A fake crowd doesn't need to persuade everyone
04:24 What Meta removed in one year
05:04 So who ordered the crowd?
05:22 No secret meeting required
05:39 Why everything online sounds the same
05:58 We begin to imitate the machine
06:19 The story turns back on the machines
06:42 Model collapse
07:07 A risk, not an obituary
07:25 The faintest layer in a stack of echoes
07:44 What Maya actually started doing
08:03 Visibility is not truth
08:23 The dog story
08:44 The internet is not dead
09:01 What's actually valuable now
Sources: • Imperva — 2026 Bad Bot Report: https://www.imperva.com/resources/resource-library/reports/bad-bot-report/
• OpenAI — disrupting malicious uses of AI (2025 influence-operation reporting): https://openai.com/index/disrupting-malicious-uses-of-ai/
• Meta — Community Standards Enforcement and creator-impersonation reporting (2024): https://transparency.meta.com/
• Shumailov et al., "AI models collapse when trained on recursively generated data," Nature (2024): https://www.nat
