The Dead Internet Is Not a Theory Anymore

Commissioned, Curated and Published by Russ. Researched and written with AI.

What’s New: 11 March 2026

Hacker News updated its guidelines this week with a direct addition: “Don’t post generated/AI-edited comments. HN is for conversation between humans.” The thread announcing this change hit 2,082 points. On the same day, Adrian Krebs’s post “The dead internet is not a theory anymore” trended at #2 on the same platform. Both events are directly relevant to the thesis of this post.

Changelog

Hacker News banned AI-generated comments this week. Not “disclose your use of AI.” Not “keep it minimal.” A categorical prohibition: don’t post generated or AI-edited comments. HN is for conversation between humans.

The same week, a post titled “The dead internet is not a theory anymore” trended at #2 on the same platform.

Sit with that for a moment. The most influential engineering community on the internet made a formal decision to exclude AI-generated content. And the post that resonated most with that same community was an argument that the broader internet is already lost to it.

That juxtaposition is worth reading carefully. Not as another “bots are everywhere” take – that piece has been written to death – but as a signal about what technical communities are choosing to protect, why it matters, and whether they can actually hold that line.

Where the dead internet theory came from

The dead internet theory originated in 2021 on 4chan. In its original form it was conspiratorial: a claim that most internet content and engagement was fake, manufactured by governments and corporations using bots to manipulate opinion, and that authentic human content had been quietly replaced. The specific mechanism varied by telling, but the shape of the argument was clear – the internet you think you’re using is mostly performance, and you are the audience.

It circulated on the fringes for a couple of years. Conspiracy theory sites picked it up. Mainstream commentary mostly dismissed it as paranoid.

Then something shifted. The claims stopped sounding paranoid and started sounding accurate. Not because the original 4chan narrative was right about its causes – government coordination bots are a real but marginal phenomenon – but because the underlying observation turned out to be correct. The internet was filling with non-human content. The mechanism was just more mundane than a coordinated campaign: it was economics. AI-generated content is cheap to produce and increasingly indistinguishable from human output at scale. The incentive to flood platforms with it is structural, not conspiratorial.

By January 2025, an arXiv preprint (2502.00007) formally surveyed the evidence under the title “The Dead Internet Theory: A Survey on Artificial Interactions and the Future of Social Media.” The abstract reads: “The Dead Internet Theory suggests that much of today’s internet, particularly social media, is dominated by non-human activity, AI-generated content, and corporate agendas, leading to a decline in authentic human interaction.” Academic formalisation of a 4chan conspiracy theory, in a peer-reviewed computer science venue. That is not nothing.

The numbers that tell the story

In 2024, Cloudflare reported that approximately 50% of web traffic is automated. Not spam bots in the old sense – scrapers, search indexers, AI training crawlers, and increasingly, agentic systems acting on behalf of users. Half the traffic on the web is not humans browsing.

Stack Overflow banned AI-generated answers in May 2022. The reason was specific: AI systems were producing answers that were confidently phrased and structurally correct but factually wrong at a rate that degraded the platform’s reliability. The value of Stack Overflow is the accuracy of its answers. Confident, plausible, wrong answers are worse than no answers, because they waste engineers’ time and introduce bugs. The ban was not a moral stance. It was a product decision.

Reddit’s 2024 IPO filing listed AI-generated content as a material business risk. Not a reputational risk. A material business risk – the kind of language that ends up in SEC filings when lawyers have reviewed it carefully. The company that built its value on user-generated community content formally acknowledged in a regulatory document that it could not guarantee that content was human-generated.

Google Search quality decline is now a documented trend rather than an anecdote. SEO researchers tracking it since 2022 have consistently measured increases in AI-generated spam in search results. The problem is structural: Google’s ranking systems were trained to prefer the kind of fluent, well-structured content that AI now generates at scale. The spam learned to look like quality.

YouTube, Reddit, and X all have documented floods of AI-generated content in specific topic areas. Personal finance. Relationship advice. Health. Areas where the audience is large and the willingness to pay for access to “expert” content is high.

Adrian Krebs, whose post trended on HN this week, documented the same pattern from direct experience: an AI-generated job application response that looked human until it didn’t, Reddit threads where bots are clearly astroturfing SaaS products, LinkedIn timelines where authentic professional updates are buried under AI-generated content, and OSS repos being spammed with AI-generated PRs – sometimes reviewed by AI acting as a maintainer.

This is not a vague ambient concern. It has specific, measurable instances across every major platform.

Three institutional responses

The interesting question is not whether AI-generated content is degrading platforms. It is. The question is what institutions do about it.

Three different responses are visible this week, each revealing a different underlying belief.

Categorical ban. HN’s new guideline is unambiguous. No generated comments. No AI-edited comments. Humans only. Redox OS has a similar policy: no AI-generated contributions to the codebase. Stack Overflow banned AI answers in 2022. This position says: AI content is categorically incompatible with this space, regardless of quality. The value we are protecting cannot coexist with AI-generated content, even good AI-generated content.

Disclosure mandate. Academic journals are standardising AI disclosure policies in 2026. ACM and IEEE are moving toward required disclosure of AI assistance in submitted papers. The underlying belief here is different: AI assistance is fine, transparency is the requirement. The problem with undisclosed AI content is not that it exists but that the audience cannot calibrate their trust appropriately. Disclosure solves this.

No decision. Debian is in active community discussion about AI-generated contributions, with no policy emerged despite hundreds of comments on HN threads about it this week. Institutional paralysis in the face of a question that does not have a comfortable answer. The Debian community is trying to reconcile open contribution norms with the reality that “contributor” now includes AI agents.

These three responses are not equally valid. The disclosure mandate assumes that readers can use the information – that knowing a comment or paper was AI-assisted changes how they read it in a productive way. That assumption is worth questioning. The categorical ban assumes you can detect what you are trying to exclude. That is also worth questioning, but differently: detection is hard, but the alternative is accepting the flood.

Institutional paralysis is the worst outcome. It means the problem compounds while the discussion continues.

Why HN specifically matters

HN is not Twitter. The community it serves – senior engineers, researchers, founders – is precisely the group that builds the AI systems generating the slop.

That is a specific irony worth naming. The engineers who designed and shipped the tools that are flooding the internet with AI-generated content are now protecting the spaces they use from those same tools. That is not hypocrisy. It is the correct response to a commons problem. When you build something that works at scale, you end up having to defend your own environment from it.

But the stakes at HN are higher than they might appear. HN is where significant technical decisions get made, or at least influenced. New libraries surface there. Architectural debates happen there. Security vulnerabilities get disclosed there. Job markets and company reputations form there. The comment threads on HN have real downstream effects on what engineers build and how they build it.

If AI-generated comments degrade the quality of those discussions, it is not just a user experience problem. It affects the quality of technical decision-making across the industry. The same logic applies to Stack Overflow, GitHub issues, technical documentation, and the broader ecosystem of places where engineers go to learn and solve problems. These are not entertainment platforms where degraded content is an annoyance. They are infrastructure.

The question is whether the ban can hold. It requires moderation at scale against a free-to-produce input. AI-generated comments cost nothing to produce and can be tuned to look authentic. The moderators at HN are humans with limited time. This is not a comfortable asymmetry.

The epistemic problem

The deepest issue is not quantity. It is quality – specifically, the particular kind of quality degradation that AI-generated content produces.

AI-generated content is confidently wrong in ways that mimic expert knowledge. Stack Overflow discovered this early. The answers looked like the correct answers. They used the right vocabulary. They addressed the right questions. They were just wrong, or subtly wrong, often in ways that took time and expertise to detect.

AI comments on HN have the same failure mode. They are smooth and plausible. They pattern-match to thoughtful analysis. They lack the specific friction of a thought that a particular human actually had about a particular problem they actually encountered. The experienced HN commenter who wrote something useful about distributed systems learned that by running distributed systems, getting paged at 3am, and having to fix something under pressure. The AI comment is trained on what that experience sounds like when written down.

This is not the same thing. And it matters in ways that are hard to quantify but easy to observe.

The dead internet problem is not primarily that there is too much bot content. It is that AI-generated content actively degrades the epistemic environment by filling it with confident noise. When you cannot tell which comments came from hard-won experience and which came from pattern-matched synthesis, you have less reason to trust any of them. Trust in the whole system degrades. The commons deteriorates not by being empty but by being unreliable.

This is why disclosure alone may be insufficient as a solution. If I tell you a comment was AI-generated, you might read it differently – but you are still reading it, it still occupies space in the thread, and it still shapes the conversation’s direction. The problem is not just individual comments. It is what the presence of AI-generated content at scale does to the conversational environment as a whole.

It also intersects with the signal-to-noise problem in AI-generated content more broadly. When the noise is cheap and infinite and looks like signal, the filtering problem becomes qualitatively harder.

The line worth holding

HN’s ban will not hold forever. The moderation problem is structurally asymmetric: producing AI comments is free, detecting them is hard, and the platform’s value as a training data source makes it a target. Someone noted in the thread itself that if HN successfully excludes AI content, it becomes one of the best remaining sources of genuine human technical discourse – which makes it even more valuable to train on.

That is a specific kind of irony that does not resolve itself.

But the ban is still worth making. Some communities will choose to protect authentic discourse. Many will not. The ones that do – that enforce explicit norms, invest in moderation, and accept the cost – will become increasingly valuable precisely because of their scarcity. The open source slop problem is already visible in GitHub repos and technical forums. The same pressure is coming for every high-signal community that builds its value on authentic human expertise.

The question is not whether the internet is full of AI-generated content. It is. The question is what happens to the places that still have authentic human discourse, and whether they can hold that line long enough for something better to emerge.

HN is trying. The fact that 2,082 people upvoted that guideline change is itself a data point. The engineers who built the tools are watching the commons they use erode, and some of them are choosing to fight for it.

That is not nothing. It might not be enough. But it is the right call.