*Published July 26, 2026* --- *What a 90%-automated knowledge system taught me about trusting anything online* --- Yesterday my knowledge system committed 192 changes. I made 7 of them. The other 185 came from a fleet of agents that lint my notes, import my highlights, cross-link my topics, and file reports while I sleep. I built this deliberately, over a year, and it works. It is also, at personal scale, the thing everyone is worried the internet has become. --- ## The Dead Internet, Briefly The Dead Internet Theory started as a conspiracy: most of what you see online is no longer human. The full version doesn't hold up, but a leaner one does: a growing, hard-to-detect share of it is now machine-made, and that alone is enough to break how credibility used to work. Venkatesh Rao pushed it further in "Dead Forest Theory": the living human internet collapsed into private interiors (Discords, group chats) whose inner life can't be reconstructed from outside. What remains public is a "perpetual visibility engine," recycling content and narratives, creating no common knowledge among strangers. LLMs, in his framing, are radiation from the corpus before the collapse: fossil-record knowledge, no access to the interior. When a hard-to-detect share of what you read is machine-generated, "does this look credible?" stops working. What works is "who is answerable, and how do I check?" Trust re-attaches to origin, not to truth. A verified source can still be wrong, or lying. --- ## The Microcosm My personal knowledge base is a dead internet with one human in it. That is an unfair vantage point. About 9,800 topic notes, most drafted by agents from captured articles and threads, maintained by loops that commit around the clock. An 804-commit week, more than 90% automation. Read that the way people read the internet, and I'd conclude my main interest in life is note formatting. For a while, my instruments did exactly that. --- ## Three Ways I Fooled Myself This month I ran an adversarial review of my vault's analysis tools: three reviewers, told to attack. What they found is a catalog of how machine-generated volume corrupts self-knowledge. **My "revealed priorities" were a cron schedule.** One tool infers what I care about from recent commits. Verified live: 22 of the last 30 were bots, a portrait of my automation presented as my character. **My "recent journals" were one folder.** A sampling command sorted file paths instead of dates, so "the last 14 journal entries" returned 14 files from the alphabetically-last folder. Every constraint map built on it rested on a single life domain. Wrong data, clean exit code. **AI opinion laundered itself into fact.** An AI-written portrait praised a trait; a later analysis cited it as evidence; a third document cited both as "the behavioral record." A leverage score one tool invented got quoted as if measured. Each hop upgraded speculation toward fact. Nobody lied; the provenance evaporated, one reference at a time. If this happens inside one vault run by someone watching for it, I no longer need convincing about what happens at internet scale, where nobody is checking. --- ## The Discipline The fixes were provenance rules, not smarter models. Three of them, replicable. **Label origin at write time.** Every AI-generated analysis now opens with a header: hypothesis, not record. Unlabeled output gets cited as fact by the next reader. **Classify the author before reading the signal.** Commits and activity mean nothing until split into "the human did this" and "the fleet did this." My tools filter automation before inferring anything about me. The internet-scale version is the one no reader can currently run. **Verify adversarially before trusting.** Three agents told to attack caught all this, checked against the live system before I acted. Charlie Marsh said the same about code: the most effective guardrail against slop is still talented humans doing thoughtful review. Human attention is scarce. Spend it at the choke points. You don't need a fleet to start. Label your chatbot's output as machine output, and trace one citation to a named origin before repeating a claim. --- ## The Hard Question Here is the part I have not resolved, and the reason this essay exists. Most notes in my vault were drafted by agents. I chose the sources, set the standards, and my lint gates every note on citation. If provenance means "a human typed this," my vault fails. If it means "a named person is answerable, and the chain of custody is inspectable," it passes better than most things I read online. I think the second definition is the real one. A signature on a photo tells you which camera fired, never whether the photographer was honest. Provenance has always attached to the accountable curator, not the hand on the keyboard. Curation without instruments is no longer enough, because the volume being curated is machine-made and self-plausible. My claim to authorship isn't that I wrote every note. It's that I run the audits that catch it lying to me, and publish the human/bot split. One embarrassment keeps me honest: my site stays open to crawlers, feeding the fossil public I'm describing. --- ## Reseeding Rao borrows Gall's Law: a complex system that works evolves from a simple system that worked. Real repair is underway too: provenance signing now ships in actual hardware, and the EU AI Act's synthetic-content labeling rules go binding in August 2026. But adoption is still poor, and no single technique is sufficient alone. Reseeding doesn't replace that repair. It's the part you can run today, at a scale of one, while the larger system catches up. I used to think my vault's provenance discipline was housekeeping. I now think it's that smaller working system. The internet I grew up with asked, "is this interesting?" The one we have now rewards a different question: who says so, and what would catch them being wrong? --- ## Sources - Dead Internet Theory: [Wikipedia](https://en.wikipedia.org/wiki/Dead_Internet_theory); measurement discussion in [arXiv:2502.00007](https://arxiv.org/abs/2502.00007) — only the leaner "growing, hard-to-detect share" version holds up under scrutiny - Venkatesh Rao, ["Dead Forest Theory"](https://contraptions.venkateshrao.com/p/dead-forest-theory), Contraptions, 2026 - Center for Democracy & Technology, ["The Promise and Risk of Digital Content Provenance"](https://cdt.org/insights/the-promise-and-risk-of-digital-content-provenance/) — provenance authenticates origin, not truth - EU AI Act synthetic-content labeling requirements (binding August 2026): [Resemble AI guide](https://www.resemble.ai/resources/complete-guide-to-eu-ai-act-watermarking-requirements-for-generative-ai) - Content Authenticity Initiative, ["The State of Content Authenticity in 2026"](https://contentauthenticity.org/blog/the-state-of-content-authenticity-in-2026) — provenance signing in hardware (Pixel 10, Sony) - Charlie Marsh on human review as the guardrail against AI slop: [x.com/charliermarsh/status/2042053725183762939](https://x.com/charliermarsh/status/2042053725183762939) - The vault incidents are from my own audit records, 2026-07-06: three-reviewer adversarial review of six analysis tools, fixes committed same day. Private git history, not a public link — the kind of chain-of-custody gap this essay's own closing question points back at.