The Dead Internet Feedback Loop and AI Model Collapse

By Julian Scaff

The Dead Internet Feedback Loop: a hall of mirrors where humans consume machine echoes, and authenticity dissolves into endless repetition. (Original image from Pexels, edited with Photoshop and Dall-E.)

The internet was once imagined as a digital commons: a boundless expanse where human creativity, knowledge, and connection could flourish. In its earliest forms, it resembled something close to this ideal. Amateur websites, forums, and blogs reflected the diversity and unpredictability of human culture. But over time, this open landscape was colonized by platforms whose business model hinged on attention capture, data extraction, and algorithmic manipulation. Today, we stand on the precipice of a new transformation, one in which the very presence of human beings on the internet is called into question.

This anxiety has been captured in the Dead Internet Theory, a conspiratorial yet oddly resonant idea that most of the content we encounter online is no longer generated by humans but by bots, corporations, and automated scripts. While overstated, the theory reflects a lived truth: the internet feels increasingly hollow, repetitive, and manipulated. Search results deliver SEO-optimized filler; social feeds recycle memes with uncanny sameness; and the signal of authentic human voices drowns in the noise of click farms, spambots, and AI-generated sludge.

Now, the rise of large language models (LLMs) threatens to deepen this trajectory. Already, generative AI systems are flooding the web with synthetic content that is nearly indistinguishable from human work. The risk is not only that we, as users, can no longer tell real from fake, but that future LLMs will be trained on their own output, a phenomenon sometimes called AI inbreeding. When a model ingests its own artifacts, the diversity and fidelity of the training data collapses. Errors are amplified, clichés reinforced, and subtle distortions snowball over time. Researchers call this process model collapse: a feedback loop in which the quality of generated text, images, and knowledge spirals downward, leaving us with degraded versions of degraded versions.

This is the Dead Internet Feedback Loop: synthetic systems generating synthetic content, which is then ingested by other synthetic systems, until the entire ecosystem is a hall of mirrors with no anchor to reality. In such a scenario, the internet ceases to be a repository of human experience and becomes a closed circuit of machine echoes. Humans remain present, of course, but as marginal actors, consumers of algorithmically generated spectacle rather than creators of meaning. The nightmare implicit in Dead Internet Theory is not that humans are gone, but that human expression is drowned out by automated artificial spectacle.

If this is the trajectory, the deeper problem lies not in the technology itself but in the political economy of the internet. Platforms and AI firms are locked in an arms race for scale and engagement, treating human attention as a resource to be mined and discarded. The internet, in this framing, is not a commons but a quarry. As long as this incentive structure dominates, the feedback loop will tighten, and our shared digital spaces will become ever more sterile.

Yet there are other paths. To prevent collapse and to reclaim the internet as a space for human thriving, we must ask not only how to manage AI but how to reimagine the digital world itself. A few speculative strategies emerge:

1. Data Provenance and Authenticity: Developing systems to track the origin of content, human-authored, AI-assisted, or machine-generated, could provide transparency and help preserve human contributions as distinct and valuable.
2. Digital Commons and Cooperative Platforms: Just as public libraries and universities preserve human knowledge against commodification, we need internet infrastructures that operate on cooperative or public-benefit models rather than extractive ones. These could prioritize community moderation, knowledge preservation, and genuine exchange.
3. Human-Centered Design of AI: Instead of maximizing clicks, LLMs and generative systems could be trained and tuned to enhance human creativity, dialogue, and collaboration, tools for augmentation rather than replacement.
4. Cultural Resistance and Renewal: Humans themselves must reclaim the act of authorship. Whether through creative subcultures, intentional communities, or educational initiatives, fostering spaces where authentic voices matter is essential.

The Dead Internet Feedback Loop is not inevitable, but it reflects a real danger: that in the pursuit of efficiency and scale, we will build an internet so hollow that it forgets the people it was meant to serve. To save it, we must insist on a digital order that values human connection over algorithmic engagement, creativity over repetition, and thriving over extraction.

The internet can still be reborn, not as a factory for attention, but as a living web of shared meaning. The challenge is not technical but moral: do we accept the collapse into automated simulacra, or do we choose instead to cultivate an internet where humanity, in all its messiness and brilliance, remains at the center?

Sources:

Hutchinson, Jamie. “The ‘Dead Internet Theory’ Makes Eerie Claims about an AI-Run Web. The Truth Is More Sinister.” The Conversation, May 20, 2024. https://theconversation.com/the-dead-internet-theory-makes-eerie-claims-about-an-ai-run-web-the-truth-is-more-sinister-229609.

Shumailov, Ilia, Yarin Gal, Nicholas Carlini, and Florian Tramèr. “The Curse of Recursion: Training on Generated Data Makes Models Forget.” arXiv, February 2025. https://arxiv.org/abs/2502.00007.

Zhao, Yuzhou, Stefano Ermon, and Percy Liang. “Model Collapse in Large Language Models.” Nature, 628 (2024): 65–72. https://www.nature.com/articles/s41586-024-07566-y.