Hotz argues we've crossed an irreversible threshold analogous to Eternal September 1993, but worse because there's no pre-slop shore to retreat to. He catalogs hallucinated READMEs, fake Stack Overflow answers, and AI-authored Amazon books as evidence the contamination is already ambient and total.
Hotz's central argument is that slop poisons the write side, not just the read side: with Common Crawl already 40-60% machine-generated, models released in 2027 will be trained partly on the output of 2024's models. This creates a recursive distributional drift that compounds with each generation rather than self-correcting.
The editorial extends Hotz's framing by arguing this is structurally different from prior eternal-September moments. When producing plausible text costs nothing, equilibrium volume goes to infinity, and the symptom is visible even in code review — PRs that compile, read cleanly, and follow conventions while being substantively hollow.
George Hotz published a blog post titled The Eternal Sloptember on May 24, 2026, riffing on the old Usenet term 'Eternal September' — the moment in 1993 when AOL opened its gates and the internet's signal quality never recovered. Hotz's thesis: we are now in the AI-content equivalent of that moment, and unlike September 1993, there is no shore to swim back to.
The post (sitting at 251 points on Hacker News at time of writing) is short, blunt, and characteristically uninterested in hedging. Hotz catalogs the visible symptoms: GitHub repositories with READMEs that confidently describe features that don't exist, Stack Overflow answers that hallucinate function signatures, Amazon listings for books whose authors are themselves generated, and a Google search experience where the top three results are increasingly indistinguishable from each other because they are, in fact, the same LLM output paraphrased through three different content farms.
What makes the piece land — and what separates it from the usual 'AI bad' discourse — is Hotz's argument about feedback loops. The slop isn't just polluting the read side of the internet; it's poisoning the write side, because the next generation of models trains on a web that is already 40-60% machine-generated by some estimates. Common Crawl snapshots from 2025 already show measurable distributional drift versus 2022. The models that get released in 2027 will have learned, in part, from the outputs of the models released in 2024.
The original Eternal September was a cultural event. Eternal Sloptember is an information-theoretic one. When the marginal cost of producing plausible-looking text drops to roughly zero, the equilibrium volume of plausible-looking text approaches infinity. Anyone who has tried to evaluate a junior engineer's PR in 2026 knows this viscerally: the code compiles, the comments are grammatically immaculate, the commit message hits all the conventional-commits beats, and absolutely none of it indicates whether the author understood the problem.
Hotz isn't the first to make this observation — Maggie Appleton's 'expanding dark forest' essay from 2023 covered similar ground, and Ted Chiang's 'ChatGPT is a blurry JPEG of the web' has been making rounds for three years. What's new in Hotz's framing is the resignation: he's not predicting a tipping point, he's announcing that the tipping point already happened and we're now negotiating with the steady state.
The community response on HN splits roughly into three camps. The first agrees with Hotz and is busy building personal trust networks — curated RSS, friends-only Discord servers, paid newsletters from named humans. The second argues this is just nostalgia: the web was always 90% garbage, and the SEO content farms of 2015 were arguably worse than the LLM farms of 2026 because at least the latter occasionally surface something useful. The third — smaller but vocal — is technical: they point to cryptographic provenance (C2PA, signed commits, attestation-based publishing) as a structural fix that doesn't depend on humans winning an arms race against generators.
The technical camp has the strongest case but the worst distribution problem. C2PA adoption outside of major camera vendors and a handful of news organizations remains essentially zero. GitHub's sigstore integration is brilliant infrastructure that almost no consumer-facing tool surfaces. The provenance layer exists; the UX layer to make humans care about it doesn't.
There's also a measurement problem hiding in plain sight. We don't actually have good public benchmarks for 'how much of the indexed web is LLM output.' The numbers floating around — 40%, 57%, the AWS researchers' 57.1% figure from machine-translation analysis — are estimates with wide error bars. Anyone making strong claims about the trajectory should be honest that the underlying data is fuzzy. Hotz mostly is.
If you ship software for a living, three things follow from taking Hotz seriously.
First, your documentation pipeline is now a security boundary. If your team copy-pastes from ChatGPT into your internal wiki, you are introducing hallucinated API signatures, made-up config flags, and confidently-wrong architectural diagrams into the artifact that new hires will trust most. The fix isn't 'ban LLMs from docs' — that ship sailed — it's instituting a verification step where every LLM-authored claim about your own system is cross-referenced against the source code or a runtime probe before it lands.
Second, dependency selection needs a freshness-and-humans signal, not just a stars-and-downloads signal. A repo with 8,000 stars, a polished README, and zero human-written issues from the last six months is now a plausible slop artifact, not a plausible production dependency. The heuristic that worked in 2020 — 'this is popular, it must be real' — is broken. Look for human review traces: substantive PR discussions, maintainer responses with personality, conference talks, blog posts that cite specific commits.
Third, your hiring loop probably already has this problem. Take-home exercises are now solved instantly. Resumes are written, reviewed, and submitted by agents. The signal you used to extract from 'did this candidate write a clean README' is gone. Live problem-solving, paired debugging, and asking candidates to defend specific decisions in code they claim to have written are no longer optional — they're the only reliable channel left.
The optimistic read on Sloptember is that markets eventually price in scarcity, and the scarcest thing on the 2026 internet is verifiable human attention applied to a real problem. Substack, paid Discord communities, and curated link-blogs are early indicators of where the value is migrating. The pessimistic read — and the one Hotz seems to hold — is that the open web, as a thing you could trust to surface a credible answer to a technical question, is now a historical artifact, and the replacements will be smaller, more gated, and less open than what they replaced. Either way, the developers who win the next five years will be the ones who invested in their own trust networks before the slop made the public ones unusable.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.