The internet is mutating, again. What was once a system of human-readable pages, hyperlinks, and visible sources has begun to vanish beneath the smooth interface of a chatbot. In its place is emerging something less tangible but more radical: a web where machines no longer index knowledge for users, but reinterpret it, synthesize it, and present it without citation or context. This is not Web 3.0, the long-promised semantic web of intelligent agents and linked ontologies. Nor is it Web3, the blockchain-fueled vision of digital decentralization. This is something more subtle, more powerful, and far more monopolistic. Call it Web 3.5: the AI-mediated web where content is generated, recomposed, and delivered not for readers, but for prompters and optimized not for truth, but for plausibility.

Web 3.5 is not defined by a new protocol or consensus layer. It is defined by a shift in epistemic architecture. Search engines no longer return links; they return answers. Personal blogs are skimmed and paraphrased before the user ever sees them. The academic paper, the community post, the article, the tutorial—all become ingredients in a vast statistical stew, processed and served by large language models (LLMs). These models—ChatGPT, Gemini, Claude, and others—are trained on the residue of the old web, and now function as its interpreter. But in doing so, they change the nature of information itself: flattening nuance, erasing origin, and replacing citation with summary. The interface has changed, but so has the authority.

The underlying business model is not new. As with Web 2.0, the user remains the product. However, in Web 3.5, the real product is the archive of collective human knowledge: scraped, filtered, and internalized by models whose training data are opaque and whose output is neither transparent nor accountable. Publishers, educators, and creators are discovering that their content is being absorbed wholesale into proprietary models. The economic asymmetry is stark. Labour is collective, but capital is centralized. Google’s search generative experience can summarize your website without ever sending traffic to it. OpenAI can ingest millions of documents and output paraphrases that bypass attribution altogether. Lawsuits are beginning to mount, but they only scratch the surface of a deeper epistemic enclosure.

This enclosure is not merely commercial; it is also geopolitical. Most foundational models are trained in English, using Western-sourced data, and governed under American legal norms. Voices from the Global South, indigenous knowledge systems, and non-dominant languages are almost entirely excluded—either because the content isn’t available in machine-readable formats, or because it doesn’t exist in the public digital sphere to begin with. Web 3.5 thus risks becoming a new colonial archive, one where the knowledge of the many is mined, modelled, and monetized by the few. The result is not a democratization of information, but a consolidation: a narrowing of what counts as knowable, computable, and sayable.

Supporters of generative AI argue that this shift is inevitable—and even beneficial. Machines can now summarize complex topics, answer user queries instantly, and personalize learning at scale. The productivity gains are real. But the trade-offs are profound. When machines speak to machines, humans no longer interact with the web as a rich, diverse, messy ecosystem of voices. Instead, they receive a simulated response—confident, coherent, but often hallucinated, unverified, and contextless. The web becomes a ghost of itself: responsive, but disembodied; efficient, but untraceable. What’s lost is not just citation, but deliberation. Discovery becomes convenience. Inquiry becomes autocomplete.

It is tempting to see this as a purely technological transition. But Web 3.5 is, above all, a governance failure. Public institutions—the libraries, universities, research councils, and archives that once served as epistemic anchors—have been largely absent from the construction of the AI web. Instead, the infrastructure of modern knowledge is being built by private firms operating under incentives of scale, speed, and shareholder return. Without intervention, Web 3.5 will lock in a future where epistemology is outsourced, and history is rewritten in latent space. To prevent this, governments and civil society must reclaim the digital commons—not only by regulating AI, but by investing in open, transparent, and plural knowledge infrastructures.

That means building public LLMs trained on licensed, diverse, and inclusive datasets. It means supporting minority languages, local archives, and alternative ontologies. It means enforcing data provenance and attribution standards. Above all, it means treating knowledge not as a commodity but as a public good that requires stewardship, not just computation. A parallel can be drawn with the rise of public broadcasting in the 20th century: when private media monopolies threatened to define reality, democracies responded by building institutions of trust. The same must now happen in the realm of machine learning.

The web, for all its faults, once offered a degree of epistemic pluralism. It was a place where dissenting voices could coexist, hyperlinks fostered transparency, and readers could trace claims to sources. Web 3.5 offers something smoother but far more brittle: a world where answers are easy but opaque and where the machinery of knowledge is hidden behind proprietary walls. This is not progress. It is simulacra certainty—and it demands resistance.

Whether Web 3.5 becomes a tool of liberation or control depends not on the brilliance of engineers but on the courage of institutions. It is not too late to act, but time is short—and the machines are already talking.