-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA256 About 2 years ago in Korea, ‘AI side jobs’ were a thing. One would create multiple Internet blogs (tistory.com was notorious for this) and then generate a bunch of articles with ChatGPT, monetizing any traffic it attracted. It was so popular that there were even paid video tutorials on ‘automating blogs to get rich quick’. I was quite amazed at the time how much hype it was getting. And then there’s this rule of economics that any discrepancy between the nominal and real value of an asset leads to people seeking to profit from it. The gap here was caused by ChatGPT dramatically lowering the investment costs of writing a web article (time, effort). With hundreds of machine-written articles and automated tagging systems, these websites tended to do fairly well on SEO, leading to some generating hoards of income using such methods. But Economics 101 steps in again; the same speculative forces also tend to close the gap between the values. It was not long before those blog ‘content farms’ stopped yielding noticeable profit. Sites that at first were rockin’ with a few banner ads are now bloated with loads of popups and anchor ads. The more these nuisances, the less readers tend to visit. And they’re not without tools! Who would want to visit and read those ad-stuffed internet blogs when they can just send an agent to do it instead? Leading to less traffic, more bots, less revenue. The 2024 ‘blog automation’ craze ended like that. It wouldn’t have been much problematic if this story was just some lesson of *there is no such thing as a free lunch*, but it’s far worse than that. To me it seems that these very actors were the ones that changed the future of the web. They’re the ones who killed the Internet. ## Lemons take over Do you remember the 2010s? There used to be a time when people would read up articles found by a Google search and then put together information inside their heads. It sounds really painstaking written like that, but it wasn’t quite. One of the biggest reason is that the *quality* of information at hand was noticeable at a glance: “Does the site look trustworthy? pass. The author’s profile? seems ok. Are the images scraped from somewhere else?” and such. But today those same heuristics just aren’t relevant anymore. Now that AIs can build a nice-looking website in ~10 minutes, where on earth would be a crappy website? Image generating models can create author profiles, cover images in a blink of an eye. There are even loads of [fake news sites](https://cybelangel.com/blog/ai-generated-fake-news-sites/) found across the network. Moreover, it tends to take a lot of effort to ‘detect’ AI-written text. We all have the experience of reading an internet blog when in the middle of reading something starts to feel *off*: that feeling of awkward phrasing (even when grammar is errorless!) and the feel of AI prose. At that point you realize you’ve wasted ~10 mins of time reading that slop article. When readers have no idea if the content I’m reading is written by a thoughtful human or generated by a single-sentence prompt, they simply decide not to make the bet. A classic example of a [Market for Lemons](https://en.wikipedia.org/wiki/The_Market_for_Lemons). The writer of course knows how much effort he put into that article, but the reader has no idea–until he reads it. If in the past the various signaling channels for conveying content quality were operational, LLMs have effectively neutralized much of them. And in a situation of severe information asymmetry, Peaches cannot hold a hand to Lemons. The effort of those who write good content gets undervalued, and they now have no incentive to write online. ## A positive feedback loop > garbage in, garbage out. This phenomenon affects AI power users as well – negatively. Haven’t you had the experience, where you click on that ‘reference’ your LLM provided just to see another AI-written article on the internet? To mitigate problems with being trained with outdated data, LLMs use a technique called [RAG](https://en.wikipedia.org/wiki/Retrieval-augmented_generation) to fetch the recent information. It effectively works like giving them Google search bars; referencing internet stuff to give answers with the newest data. The main idea about RAGs was to let LLMs use the Internet (or whatever source) as a good *source of truth* while writing. But *who* writes those text these days? The concept of feeding GenAI models with its own outputs as training data is [known to be a bad idea](https://arxiv.org/abs/2307.01850). While RAGs are not training mechanisms, a similar problem may potentially occur. Especially hallucination! While hallucinations of LLMs itself has decreased substantially over the years, a circulation environment we just built poses a graver problem. AIs, when fed inaccurate information generated by other models, primarily with methods like RAG, would facilitate an environment where false information is created and spread exponentially. Kurzgesagt had a recent [video](https://www.youtube.com/watch?v=_zfN9wnPvU0) about misinformation spreading through AI slop content being the biggest threat to human knowledge. ## So what? As usual, any novel idea I think of isn’t really novel: a bunch of researchers around the world have already done quantitative works on the topic and written papers about it. And sure they talked about solutions! Here’s an excerpt from the ACM2026 paper [*Retrieval Collapses When AI Pollutes the Web*](https://arxiv.org/abs/2602.16136), written by three researchers from Naver Labs. > (…) as Agentic AI begins to autonomously publish content, defense mechanisms must evolve from static text analysis to behavioral fingerprinting, identifying and isolating agents that systematically produce high-entropy, low-factuality streams. So we need to change our system, of course. One big issue leading to the AI slop problems of today is that search engines fail to filter them out - a clear limitation of traditional ranking systems. So by reinforcing our search systems, we could be able to put the task of determining whether a text is AI slop or not upon the search engine, not the user. Anthropic recently announced that [they’ll be putting watermarks to Claude output](https://www.anthropic.com/news/claude-text-watermark) starting August 2026, and it’s not irrelevant to this. Instead of relying on brittle methods like those AI detection tools, we could intervene from the generation process to make AI-written text distinguishable. While this clearly isn’t a perfect solution on its own, it reflects a growing concern that AI writing should be made identifiable. - --- After the blind optimism with the advent of LLMs went away, we’re now faced with the threats AI systems pose to the human knowledge base. But that wouldn’t mean removing AI entirely – it bears potential and benefits simply too large to ignore. Still, the task of tackling the problems of AI slop and restoring the Internet into a ‘treasury of human knowledge’ would be left for humans to complete. -----BEGIN PGP SIGNATURE----- iKYEARYIAE4WIQQmP48mzPgsad1+4+LxH/R5ARn9kAUCao/3/BsUgAAAAAAEAA5t YW51MiwyLjUrMS4xMiwwLDMUHGhlbGxvQGJhbWJhbmdhbi5uZXQACgkQ8R/0eQEZ /ZBHvAEAxF3icWq/qQdG1fp7tdNRg60bREBpqT4lC9TezT7JQBwA/R0XyFcZkbEg oT6xJ1kbhVmdgM26pamcxH4w3ZlsmAkG =7uqY -----END PGP SIGNATURE-----