# JUNK SITE: The Digital Echo Chamber of AI-Generated Web Spaces Canonical URL: https://jeanyoon.ch/oi/texts/junk-site-the-digital-echo-chamber-of-ai-generated-web-spaces Type: essay / research note Author: Jeanyoon Choi (최정윤) Date: 2025-02-06 ## Summary Junk Site is Jeanyoon Choi's 2025 proposal for a multi-device environment that continuously generates working websites, one per output channel, driven by a mobile voice interface and blink input. Its subject is a possible closed loop in which language models generate websites that return as training data, increasing quantity without novelty. The title adapts Rem Koolhaas's Junkspace to the web. The proposal pairs a framework for generating site structure, content, design and search metadata with interaction research, and argues for keeping the circuit open to human invention. ## 요약 (Korean) 최정윤의 『Junk Site』 제안서로, 음성과 눈 깜빡임 입력에 따라 웹사이트를 끝없이 생성하는 멀티 디바이스 작업을 통해 LLM과 생성 콘텐츠 사이의 폐쇄 루프를 렘 콜하스의 '정크스페이스'에 빗대어 다룬다. ## Concepts defined - **Junk Site**: A proliferation of chaotic, self-replicating web spaces mass-produced by language models, homogenised by the circulation of data between generated content and model training; a web counterpart to Koolhaas's Junkspace. ## Questions this answers ### What is Jeanyoon Choi's Junk Site? Junk Site is a proposed multi-device artwork that continuously generates functioning websites across several screens, controlled by voice context and eye blinks on a phone. It addresses how AI-generated web content may feed back into language-model training. ### How does Junk Site relate to Rem Koolhaas's Junkspace? Jeanyoon Choi translates Junkspace to the web: as air conditioning and escalators enabled homogeneous commercial spaces to spread, generative technologies enable apparently different but structurally homogeneous websites to proliferate. ### What is the risk of a closed loop between language models and AI-generated content? The proposal argues that if generated outputs become the main source for further generation, data grows in quantity but not novelty, and new discourse and design languages may struggle to enter. It calls for an open loop sustained by human invention. Topics: ai-generated content, ai training data, website generation, multi-device web artwork, voice user interface, blink interaction, search engine optimisation References: [Rem Koolhaas](https://en.wikipedia.org/wiki/Rem_Koolhaas), Junkspace, [Nassim Nicholas Taleb](https://en.wikipedia.org/wiki/Nassim_Nicholas_Taleb), [OpenAI](https://en.wikipedia.org/wiki/OpenAI) ## Full text Junk Site is a proposal for a multi-device environment that continuously generates functioning websites. Each output channel contains a different site. A mobile voice interface supplies changing contexts; a blink triggers a design update; an automated process navigates through the sites, perhaps once a second. The subject is not merely automated UI design. Language models and coding assistants can produce landing pages, navigation, nested pages, text and search metadata: the components of an entire web service. Similar processes already suggest the mass generation of posts, images and videos. What happens when this material returns as training data? I am interested in a possible closed circuit: language model → generated website → language model. The quantity of material increases while its novelty may not. Trash in, trash out, repeated without an encounter with something outside the loop. The title refers to Rem Koolhaas’s Junkspace. In my reading, technologies such as air conditioning and escalators enabled commercial spaces to expand and repeat across the world. Junk Site translates that concern into the web: technologies of generation enable a proliferation of apparently different but structurally homogeneous digital environments. Where does human creativity sit within this circulation? The question extends beyond websites to generated text, images and video, but websites offer a particularly tangible multi-device form in which to encounter it. As a research project, this has two connected strands. One is a framework for generating website structure, context, content and visual design. The other is interaction: extracting a context from a voice exchange and using a blink as an input that changes the generated environment. The critical title need not become the label of every technical contribution. The longer-term challenge concerns both training material and human creation. If generated outputs become the dominant source for further generation, new discourse and new design languages may struggle to enter the circuit. I want to argue for an open relation in which human invention continues to introduce material that cannot be reduced to recycling the existing dataset. The proposed framework would create actual programmed websites with usable URLs, rather than design-tool mockups. Given a context, keyword or site category, it would generate page content, navigation and relationships between pages; produce multiple visual treatments; and construct paths and metadata for search discovery. Its practical usefulness is secondary to the artistic question. The installation would let participants generate sites in real time and encounter the scale of that proliferation. The original research notes also proposed flooding the web with search-optimised sites as a speculative intervention, beginning with mutually referencing pages about multi-device web artwork. That was a proposed strategy within the artwork, not an account of an action already carried out. The goal is to make a possible future of self-recycling content available to experience, and to ask what would be required to keep the circuit open. References in the original notes include Koolhaas’s Junkspace and Taleb’s discussion of data retrieval and generation. --- Jeanyoon Choi (최정윤) is a Korean computational artist and researcher (b. 1999), PhD candidate at KAIST XD Lab, making multi-device web artworks. Profile: https://jeanyoon.ch/oi/en. Index: https://jeanyoon.ch/llms.txt.