Reddit’s co‑founder Alexis Ohanian didn’t mince words: “so much of the internet is now just dead,” he told the TBPN podcast, describing a web increasingly populated by “botted” and “quasi‑AI” content, where genuine human signals are harder to find and harder to trust. That observation — echoed by OpenAI CEO Sam Altman’s candid note that he’s seeing “a lot of LLM‑run Twitter accounts now” — has slid what was once an internet fringe theory into mainstream concern. At the same time, independent audits, publisher complaints and platform telemetry paint a picture of an information ecosystem being reshaped by generative AI, scraping bots, and answer‑first search experiences that claim to save time while hollowing out the open web’s referral economy.
This feature examines whether the internet is “dead” in any meaningful sense, what evidence supports the claim, where the data are unclear or contested, and what practical options exist to preserve humans’ role in the information stack. The discussion crosses technical, economic and civic lines — from bot‑traffic metrics to AI training data scarcity, from publisher lawsuits over scraping to product designs that privilege AI summaries over links. For Windows users, site owners and IT professionals, the consequences are immediate: search and discovery patterns are changing, traffic models are under pressure, and new controls and protocols are being built precisely to push the web back toward human‑centered outcomes.
The “dead internet theory” started as an online conspiracy and thought experiment: that much of what appears online is generated by bots, not by authentic human authors. Once ridiculed, the idea gained traction as generative AI tools and bot networks became more powerful and easier to run at scale. Two separate but related technical trends have accelerated attention: the proliferation of automated scraping and retrieval bots that feed LLMs or generate content themselves, and the emergence of “answer‑first” or conversational search experiences (AI Overviews, chat assistants) that present synthesized answers instead of links. Together they change both supply (what’s produced) and distribution (how users encounter content).
What’s new is not just the existence of machine‑generated text — it’s the scale, and the effects. Bots now service everything from price scraping to automated posting and coordinated amplification. Meanwhile, search pages that once drove clicks to publishers increasingly present compact AI summaries that satisfy many users without sending them onto the source site. Publishers say this kills referral revenue; platforms say AI improves user experience. Independent studies and security vendors offer different measures of just how “botty” the web has become. The net result is that key pieces of the web’s attention economy are under stress.
Why measurements diverge: vendor definitions differ, vantage points differ, and “bot” is a catch‑all. Good bots (search engine crawlers, monitoring services) still matter; malicious bots and AI‑driven scrapers are the problem. For site owners, the practical outcome is the same: increased server load, the risk of scraped content being reused without attribution or payment, and rising costs to both detect and mitigate nonhuman traffic.
Why this matters beyond publishers: referral traffic has long been the discovery mechanism that funds niche creators, specialist blogs and investigative journalism. If answer‑first interfaces satisfy low‑ and medium‑complexity queries, the marginal clicks that once kept specialist publishers afloat vanish. The economic pressure can accelerate consolidation and reduce the diversity of voices on the open web.
What remains less certain is the pace and direction of systemic collapse. Will AI‑driven interfaces completely disintermediate publishers, or will market and regulatory responses rebalance incentives? Can decentralized or verifiably human systems scale without becoming exclusive? These are open questions that hinge on product design, business negotiations and public policy. Many claims about “running out of data” are conditional projections based on specific scaling assumptions; they are plausible scenarios, not certainties, and deserve cautious interpretation.
Another important nuance: not all automation is malicious. Good bots (indexers, accessibility tools, legitimate assistants) remain essential to the web’s function. The policy challenge is to manage intent and provenance — to let legitimate machine actors coexist while preventing stealthy scraping and deceptive amplification. Technical controls, business models and regulation will have to evolve in concert to steer outcomes.
What this moment demands is triage plus architecture. Triage: stop the worst abuses by improving bot management, clarifying legal rights around scraping, and supporting publishers that underpin public interest information. Architecture: build standards (like conversational web endpoints and provenance protocols), refine product UX to favor exploration and verification, and design means of demonstrating human presence without sacrificing privacy. From a Windows and publisher standpoint, the operational work starts now: instrument traffic, harden APIs, adopt standards that preserve provenance, and rethink discovery strategies in an era where answers often arrive without links.
The internet’s vitality has always depended on incentives — why someone writes, why someone hosts, why someone reads. AI has rearranged those incentives, not erased them. The next phase will be defined by how engineers, publishers, platform owners, regulators and users respond — whether they treat this as an opportunity to rebuild authenticity and sustainability, or as an inexorable shift toward curated, closed systems. The stakes include the future of independent reporting, the trustworthiness of public information, and even the kinds of software we install on our Windows desktops. The conversation that started as a podcast soundbite is now an urgent policy and product task: not to mourn a dead internet, but to prevent one from being engineered into obsolescence.
Source: Windows Central Is the internet already dead? Reddit co-founder say bots and AI are taking over
This feature examines whether the internet is “dead” in any meaningful sense, what evidence supports the claim, where the data are unclear or contested, and what practical options exist to preserve humans’ role in the information stack. The discussion crosses technical, economic and civic lines — from bot‑traffic metrics to AI training data scarcity, from publisher lawsuits over scraping to product designs that privilege AI summaries over links. For Windows users, site owners and IT professionals, the consequences are immediate: search and discovery patterns are changing, traffic models are under pressure, and new controls and protocols are being built precisely to push the web back toward human‑centered outcomes.
Background / Overview
The “dead internet theory” started as an online conspiracy and thought experiment: that much of what appears online is generated by bots, not by authentic human authors. Once ridiculed, the idea gained traction as generative AI tools and bot networks became more powerful and easier to run at scale. Two separate but related technical trends have accelerated attention: the proliferation of automated scraping and retrieval bots that feed LLMs or generate content themselves, and the emergence of “answer‑first” or conversational search experiences (AI Overviews, chat assistants) that present synthesized answers instead of links. Together they change both supply (what’s produced) and distribution (how users encounter content). What’s new is not just the existence of machine‑generated text — it’s the scale, and the effects. Bots now service everything from price scraping to automated posting and coordinated amplification. Meanwhile, search pages that once drove clicks to publishers increasingly present compact AI summaries that satisfy many users without sending them onto the source site. Publishers say this kills referral revenue; platforms say AI improves user experience. Independent studies and security vendors offer different measures of just how “botty” the web has become. The net result is that key pieces of the web’s attention economy are under stress.
The Evidence: What’s real, and what’s contested
Bot traffic: numbers that don’t quite agree
Security firms disagree on exact numbers, but the direction is unanimous: automated traffic has grown significantly in recent years. Imperva’s 2025 Bad Bot Report (published by Thales) found that automated bots accounted for roughly 51% of web traffic in 2024 — the first time bots reportedly outpaced humans — with a large fraction classified as “bad bots” used for scraping, credential stuffing and fraud. By contrast, Cloudflare’s application‑security reports have consistently reported that about 30–32% of application traffic is bot traffic, while noting that a significant proportion of bot activity is malicious or questionable. Those differences matter: they reflect the diversity of measurement methods, network vantage points and the distinction between HTTP application traffic and the full scope of global internet flows. The headline takeaway is not a precise percent, it’s the trend: automated activity has grown to a scale where it materially alters how sites are visited and measured.Why measurements diverge: vendor definitions differ, vantage points differ, and “bot” is a catch‑all. Good bots (search engine crawlers, monitoring services) still matter; malicious bots and AI‑driven scrapers are the problem. For site owners, the practical outcome is the same: increased server load, the risk of scraped content being reused without attribution or payment, and rising costs to both detect and mitigate nonhuman traffic.
AI Overviews and the collapse of clickthroughs
Perhaps the most quantifiable change in user behavior has been documented around “AI Overviews” — synthesized summaries delivered on top of search results. A Pew Research Center analysis found that when an AI Overview appears, users click through to external links far less frequently: click rates fell from roughly 15% without an AI summary to about 8% with one, and only about 1% of AI Overview visits included a click on a cited source. The same report found more sessions ended after seeing an AI Overview. Publisher analytics firms have echoed those findings, reporting traffic drops of 50% or more for queries where an AI summary replaces or crowds out traditional results. The business implication is stark: fewer clicks means less ad revenue and fewer subscription signups driven by search referrals.Why this matters beyond publishers: referral traffic has long been the discovery mechanism that funds niche creators, specialist blogs and investigative journalism. If answer‑first interfaces satisfy low‑ and medium‑complexity queries, the marginal clicks that once kept specialist publishers afloat vanish. The economic pressure can accelerate consolidation and reduce the diversity of voices on the open web.
Training data scarcity and model quality risks
Several industry observers and data scientists now warn that the era of cheap, high‑quality, human‑generated public training data is constrained. Reports and analyses from research groups like Epoch AI and comments from enterprise data executives indicate that relying on indiscriminate web scraping will eventually yield diminishing returns: the pool of high‑quality, non‑synthetic, legally reusable human text is finite under prevailing training practices. That has two consequences: AI labs may increasingly rely on synthetic data — which risks “model collapse” if models train on recycled machine‑generated material — or shift to proprietary datasets that advantage incumbent firms with large data stores. Both paths change incentives and the shape of future models. These projections are contingent and technical — they are plausible but far from inevitable.The human side: young people and the shift to AI
Academic and survey research shows that younger users are rapidly incorporating AI tools into research, homework and everyday life. Multiple studies from 2024–2025 report that university students and adolescents increasingly use ChatGPT, Snapchat’s My AI, and other assistants as an initial information source or a social companion. That doesn’t mean kids trust AI more; often it means they use it for convenience, drafting, ideation or quick answers while still appreciating its limits. Importantly, short interventions in AI literacy do not reliably prevent over‑reliance — a worrying signal for educators and IT decision‑makers. For public information ecosystems, the upshot is cultural: more people start their searches and their social conversations with a machine rather than with a human‑authored page.Voices from the center: Ohanian, Altman and publishers
- Alexis Ohanian argued that platforms filled with low‑quality, automated content reduce the value of attention and that the next generation of social media will need to be verifiably human — a phrase that captures both the desire for authenticity and the challenge of achieving it without stifling participation. He pointed to the migration of meaningful conversations into private group chats as evidence of both the problem and a nascent solution.
- Sam Altman, whose company built one of the most influential large language models, admitted he’s now seeing networks that feel “fake” and suggested the creeping reality of the dead internet theory — an unusual public pivot for an industry leader. His comment elevated a once‑marginal idea into mainstream debate and fueled urgent conversations inside AI labs and platforms about attribution, provenance and platform health.
- Publishers and journalists have been blunt: AI Overviews and scraping bots have already reduced referral traffic, they say, and some legacy outlets have pursued litigation or refused access. The New York Times, Forbes and others have publicly pushed back, either legally or through technical defenses, while trade associations and regulators in the EU and UK contemplate systemic remedies. Those clashes are reshaping the commercial terms of web publishing and are likely to inform policy decisions in the near term.
Risks and consequences — short term and structural
- Journalism and niche publishing: Reduced referral economics can hollow out specialist reporting, local journalism and independent blogs. Fewer eyeballs translate into less revenue and thinner coverage for issues that large platforms don’t prioritize.
- Information quality and hallucination risk: Generative AI systems sometimes “hallucinate” — inventing plausible but false details. When these systems are both a major source of answers and a vehicle for amplifying their own outputs (synthetic data feeding new models), the result can be degraded knowledge layers that are harder to correct.
- Economic concentration: If future training data increasingly depends on proprietary enterprise stores or licensed publisher collections, the advantage accrues to firms that can pay for or control those datasets — reinforcing winner‑take‑most dynamics already visible in cloud and compute markets.
- Platform manipulation and trust erosion: Automated amplification and bot-generated “engagement” can distort signals used by ranking and recommendation algorithms, making it harder for humans to find trustworthy sources and easier for bad actors to game visibility.
- Civil and regulatory fallout: The combination of legal disputes over copyright, technical efforts to block scrapers, and formal complaints about AI Overviews has pushed the problem from the lab into courts and regulatory bodies — a messy transition that could produce either helpful rules or overbroad restrictions.
Responses, fixes and deterrents under development
The internet is not inevitably doomed. Engineers, publishers and platform operators are building countermeasures — some incremental, some structural — that aim to preserve human value and restore healthy incentives.Technical protocols and publisher tooling
- NLWeb and AutoRAG: Microsoft’s NLWeb and Cloudflare’s AutoRAG are practical efforts to let websites expose authoritative, structured natural‑language endpoints that answer conversational queries with provenance. In short: rather than letting agents scrape and summarize HTML unpredictably, sites can provide machine‑readable answers and context designed for agents and assistants. That’s a steps‑toward restoring attribution and control for publishers. Adoption matters: these tools need to be widely deployed to be effective.
- Bot management and “AIndependence” controls: CDN and security providers are offering tools to block or label AI crawlers and to monetize or rate‑limit bot traffic. These countermeasures can reduce unwanted scraping but also raise questions about which bots should be allowed and who decides.
Business and licensing strategies
- Publisher partnerships and licensing: Some publishers are negotiating licenses with AI vendors that provide compensation for training or retrieval use. These commercial arrangements can underwrite the journalism ecosystem if they’re widespread and fairly priced. The alternative — unilateral scraping by model vendors — has already provoked lawsuits and contract disputes.
Verification and “proof of life”
- Human verification mechanisms: Ohanian’s “verifiably human” vision can take many forms: cryptographic attestations, rate‑limited public posting channels, or reputation systems that reward proven human participation. Any such system must balance privacy, inclusivity and friction; heavy‑handed identity requirements risk excluding marginalized voices.
Product design tradeoffs
- Search and assistant UX: Designers can adjust how AI summaries are presented — including clearer provenance, obvious links to sources, friction for high‑stakes queries, and explicit “read more” nudges. These UX choices influence whether an answer becomes a black box or a gateway to further exploration. The Pew study suggests design choices dramatically alter user click behavior; product teams can use that leeway to favor discovery over closure.
What Windows users, site owners and IT teams should do now
- For site owners and publishers
- Implement bot management and rate‑limiting to reduce unwanted scraping costs.
- Explore NLWeb/AutoRAG or equivalent to expose structured answers and control provenance.
- Monitor search referral trends closely; diversify traffic sources beyond search.
- For enterprise IT and Windows admins
- Audit third‑party crawlers and APIs hitting corporate sites and apply adaptive WAF rules.
- Treat AI agents as a new class of client: build rate limits, API keys and observability for agent traffic.
- Consider internal governance for enterprise data if you plan to offer it to third‑party models.
- For everyday users
- Keep a skeptical stance toward single‑answer outputs; use AI for drafting and ideation, not final verification.
- Favor sources with clear provenance on high‑stakes topics (health, legal, finance).
- Protect group chats and private spaces with strong platform hygiene — end‑to‑end encryption, admin controls, and clear norms around AI‑generated messages.
- For developers and product teams
- Design assistant UX that surfaces citations and encourages exploration (not just closure).
- Build prompt‑level provenance and logging for outputs used in public‑facing features.
- Invest in detection and labeling techniques that help users tell human‑authored content from machine outputs.
A critical appraisal: strengths, limits and unknowns
The arguments that “the internet is dead” are rhetorically powerful but deserve nuance. There is strong evidence that automated traffic has increased and that AI summaries alter user behavior; those are defensible, measurable claims supported by independent reports. There is also a credible risk that training on synthetic or recycled AI outputs could degrade model quality over time — a legitimate technical worry raised by data‑scarcity analyses.What remains less certain is the pace and direction of systemic collapse. Will AI‑driven interfaces completely disintermediate publishers, or will market and regulatory responses rebalance incentives? Can decentralized or verifiably human systems scale without becoming exclusive? These are open questions that hinge on product design, business negotiations and public policy. Many claims about “running out of data” are conditional projections based on specific scaling assumptions; they are plausible scenarios, not certainties, and deserve cautious interpretation.
Another important nuance: not all automation is malicious. Good bots (indexers, accessibility tools, legitimate assistants) remain essential to the web’s function. The policy challenge is to manage intent and provenance — to let legitimate machine actors coexist while preventing stealthy scraping and deceptive amplification. Technical controls, business models and regulation will have to evolve in concert to steer outcomes.
Conclusion — not a eulogy, but an emergency lamp
The phrase “the internet is dead” captures a real cultural and technical anxiety: much of the web now includes substantial machine‑generated content, and the ways we find answers are changing faster than rules for attribution, compensation and provenance. But the web is not a single organism that can die; it’s a layered ecosystem of protocols, markets and social practices that can be repaired, legislated and redesigned.What this moment demands is triage plus architecture. Triage: stop the worst abuses by improving bot management, clarifying legal rights around scraping, and supporting publishers that underpin public interest information. Architecture: build standards (like conversational web endpoints and provenance protocols), refine product UX to favor exploration and verification, and design means of demonstrating human presence without sacrificing privacy. From a Windows and publisher standpoint, the operational work starts now: instrument traffic, harden APIs, adopt standards that preserve provenance, and rethink discovery strategies in an era where answers often arrive without links.
The internet’s vitality has always depended on incentives — why someone writes, why someone hosts, why someone reads. AI has rearranged those incentives, not erased them. The next phase will be defined by how engineers, publishers, platform owners, regulators and users respond — whether they treat this as an opportunity to rebuild authenticity and sustainability, or as an inexorable shift toward curated, closed systems. The stakes include the future of independent reporting, the trustworthiness of public information, and even the kinds of software we install on our Windows desktops. The conversation that started as a podcast soundbite is now an urgent policy and product task: not to mourn a dead internet, but to prevent one from being engineered into obsolescence.
Source: Windows Central Is the internet already dead? Reddit co-founder say bots and AI are taking over
