The Dead Internet, Measured

Is the internet dying, taken over by machines that write for other machines? What can be measured: who reads the web, who writes it, who answers instead of linking, and where the people went.

Essay · the web · October 2026

On 5 January 2021 a user called IlluminatiPirate posted a long thread on Agora Road, a small forum, under the title “Dead Internet Theory: Most of the Internet is Fake”. “The Internet feels empty and devoid of people,” it began, and it went on to claim that “the U.S. government is engaging in an artificial intelligence powered gaslighting of the entire world population.”[1] The evidence was a collection of impressions: forums that felt repetitive, posts that seemed to come from “someone or something”, the same jokes everywhere.

The thread would have stayed obscure if The Atlantic had not written about it that August, under a title that summed up the mood: “Maybe You Missed It, but the Internet ‘Died’ Five Years Ago”.[2] Since then the phrase has travelled far beyond the forum. Retellings usually date the death to around 2016, although the original post never says the internet died, only that something changed on some forums in those years.[2]

As a conspiracy theory, it does not hold. There is no evidence of a government plan to replace people with programs, and those who wrote about it at the time called that part a paranoid fantasy, while granting that the feeling behind it was real.[2] But five years later, several of the things the thread could only feel can be counted. This essay sets the theory aside and asks the measurable questions instead: how much of the web is read by machines, how much is written by them, how often a machine now answers in place of a page, what happens when machines learn from machines, and where the people are.

Machines reading

Most visits are not people

Imperva, a security company owned by Thales that watches traffic on many large sites, reported that in 2024 automated traffic overtook human traffic for the first time in a decade, at 51 per cent, and that 37 per cent of all traffic came from malicious bots.[3] Its report for 2025 put bots at more than 53 per cent and people at 47.[4] Cloudflare, which carries a large share of the web’s traffic, measures it differently, by requests for web pages, and found something similar at the end of 2025: on 2 December, 47 per cent of requests for web pages came from people and 44 per cent from ordinary bots, while bots run by AI companies averaged about 4 per cent over the year.[5] Neither is a census of the whole web; both are large samples, and they agree that people are now roughly half.

Small sites see more. This blog’s own counter, studied in Mostly Machines, recorded 20,900 views in a month in which readers other than the author made perhaps 160 to 320: about one view in a hundred.

The new arrival is the AI crawler, which fetches pages to train language models or to answer users’ questions.[a] What makes it different from a search engine is what it gives back. A search engine indexes a page and sends readers to it; a crawler for a model takes the text and, in most cases, sends nobody. Cloudflare compares the pages each company crawls with the visitors it refers. In 2025 Google crawled between 3 and 30 pages for every visitor it sent; OpenAI as many as 3,700 in March; Anthropic, the maker of Claude, between about 25,000 and 100,000 in the second half of the year, after peaks of 500,000 earlier.[5] Training, not search, accounted for 80 per cent of AI crawling over the year to July 2025.[6]

Figure 1. Pages crawled for every visitor sent back to the site, 2025, on a logarithmic scale. Grey: search engines; red: AI companies. Bars show the lower value and the shaded part the range the source gives; OpenAI is its March peak, Anthropic its range for the second half of the year. Source: Cloudflare Radar 2025 Year in Review.[5]

Sites have started to close the door. By 2024 the owners of about 28 per cent of the most actively maintained sources in C4, a widely used training collection of web text, had restricted AI crawlers in their robots.txt files, the notices that tell crawlers what they may fetch.[7] In July 2025 Cloudflare began blocking AI crawlers by default on new domains, and offered sites a way to charge them per page, which it called “pay per crawl”.[8] Its chief executive put the change in terms of what the old bargain no longer pays: “With OpenAI, it’s 750 times more difficult to get traffic than it was with the Google of old.”[8]

Machines writing

About half of new articles, and a growing share of everything else

The most careful estimate of machine-written text on the open web comes from Graphite, a marketing company, which sampled 55,400 English articles from Common Crawl, the public archive of the web, and ran each through three AI detectors. Articles written mainly by AI were about 36 per cent of new ones a year after ChatGPT launched, 48 per cent a year later, and have stayed near 50 per cent since: “relatively stable, near 50%, over the last five quarters.”[9] The detectors wrongly flagged fewer than 2 per cent of articles written before ChatGPT existed, but the study did not test human articles drafted with AI, so the boundary between the two is blurred.[9] A different measure, from the SEO company Ahrefs, found some AI content in 74 per cent of 900,000 new web pages in April 2025, but only 2.5 per cent written purely by AI; most were mixed.[10]

Other kinds of content show the same curve, steeper. Deezer, the music streaming service, has tagged fully AI-generated tracks since January 2025, when they were about 10,000 a day. By April 2026 they were almost 75,000 a day, 44 per cent of new uploads, and in June 2026 more than half.[11] Almost nobody listens to them: they make 1 to 3 per cent of streams on Deezer, and up to 85 per cent of those streams were fraudulent, plays by bots to collect royalties.[11] Spotify said in September 2025 that it had removed more than 75 million “spammy” tracks in a year.[12] Amazon limited self-published authors to three new books a day in 2023, citing abuse.[13]

Figure 2. Share of new items that are machine-made. Articles: English articles in Common Crawl written mainly by AI, by quarter, set at zero when ChatGPT launched in November 2022 (Graphite). Music: fully AI-generated tracks among daily uploads to Deezer, as reported by the company; the April 2025 to January 2026 values are from press coverage of its releases, the last point is “more than 50 per cent”. Sources: Graphite; Deezer.[9][11]

Some of it is written to deceive. NewsGuard, which rates news sites, counted 3,749 “AI content farms” in 16 languages by June 2026: news sites that publish mostly machine-written articles without saying so.[14] The United States banned fake reviews, including those “AI-generated”, in a rule that took effect in October 2024.[15] How these are used to simulate public opinion is the subject of The Manufactured Public.

One number that is not part of the evidence: the claim that “90 per cent of online content” would be synthetic by 2026, which is widely attributed to Europol. Europol did print it in a 2022 report, as an estimate by “experts”, but its source was a 2020 book on deepfakes, which quoted the chief executive of an AI video company forecasting that synthetic video might make up 90 per cent of all video within three to five years.[16] A sales forecast about video became an expert estimate about everything. Europol’s revised report of January 2024 no longer contains the sentence.[16]

Machines answering

The answer instead of the link

For twenty years the bargain of the open web was simple: sites let search engines read them, and search engines sent readers back. Even before AI, that was weakening. In 2024, by one panel’s count, almost 60 per cent of Google searches in the United States and Europe ended without a click on anything.[17] AI summaries at the top of the results push further. The Pew Research Center tracked the browsing of 900 American adults in March 2025: on searches that showed an AI summary they clicked a result on 8 per cent of visits, against 15 per cent when there was none, and clicked a link inside the summary only 1 per cent of the time.[18]

Figure 3. What 900 American adults did on Google results pages with and without an AI summary, March 2025 (68,879 searches). Source: Pew Research Center.[18]

The places that produced the answers feel it. In October 2025 the Wikimedia Foundation, after improving its bot detection and finding that much of an apparent surge of readers earlier that year was bots “built to evade detection”, reported that human pageviews were about 8 per cent lower than a year before, and said it believed this reflected “the impact of generative AI and social media on how people seek information”.[19] The Foundation’s own public data show the decline continuing: views by people of the English Wikipedia were 7 per cent lower from March to September 2026 than in the same months of 2025, and 14 per cent lower than in 2016, though changes in bot detection make long comparisons uncertain.[20]

Figure 4. Monthly pageviews of the English Wikipedia not classified as bots, January 2016 to September 2026. The dotted line marks the improvement of bot detection in 2025; earlier months may still include bots that evaded it. Source: Wikimedia pageviews API, agent type “user”, retrieved 28 October 2026.[20]

Stack Overflow, the question-and-answer site where programmers solved each other’s problems for fifteen years, shows the most extreme case. New questions had been falling slowly since 2014, from about 207,000 a month to 106,000 just before ChatGPT. Then they collapsed: 52,000 a month a year later, 9,600 in May 2025, 963 in September 2026, one for every 215 asked in March 2014.[21] The answers that trained the models are no longer being written, because people now ask the models.

Figure 5. New questions on Stack Overflow per month, selected months, on a logarithmic scale. Counts are questions still on the site; deleted ones are not included, which understates older months more. Source: Stack Exchange API, retrieved 28 October 2026.[21]

Part of the web is being paid for instead. Reddit licensed its content to Google in February 2024, in a deal reported at about 60 million dollars a year, and to OpenAI three months later; publishers including the Associated Press, the Financial Times and News Corp have signed similar agreements.[22] Others went to court: The New York Times sued OpenAI and Microsoft in December 2023 for using its articles without permission.[23]

Machines learning

When models learn from models

If half of new text is written by machines, the next machines will learn partly from it. A paper in Nature in July 2024 showed what can happen when they learn only from it: trained generation after generation on their predecessors’ output, models lose the rare cases first, as “tails of the original content distribution disappear”, and end up producing a narrow, degraded version of what they started with.[24] The authors called it model collapse.

Another group answered that the experiment assumed the worst case, in which each generation’s output replaces the human data. If the synthetic text is added to the original rather than replacing it, collapse is avoided, in theory and in their experiments.[25] Since nobody deletes the old web, the second scenario is closer to reality. The open question is a practical one: how much fresh human writing remains to be added, if fewer people write in public and more of what they do write is closed to crawlers.

The people

Where the people are

On the other side of the ledger, there have never been more people online. The International Telecommunication Union estimates that about 6 billion people used the internet in 2025, three-quarters of the world.[26] WhatsApp alone has more than 3 billion monthly users.[27] The most visited sites in September 2026 were Google, YouTube, Facebook, Instagram and, fifth, ChatGPT; Wikipedia was eleventh.[28]

What has changed is where people spend that time and what they read. Across the 48 countries in the Reuters Institute’s survey, the share of people who watch news on social video platforms each week rose from 52 per cent in 2020 to 65 per cent in 2025, and in the United States social and video networks overtook both television and news websites as a source of news.[29] Yet most people still prefer reading the news to watching it, 55 per cent against 31.[29] Long writing found new channels: Substack, which delivers it by email, passed five million paid subscriptions in March 2025.[30]

The movement is away from the open, public web of forums, blogs and comment sections, towards private groups, feeds and inboxes. On a small scale, this blog shows the same pattern: other people left 160 comments on its posts between 2011 and 2016 and 17 between 2020 and 2026, while its links travelled through Facebook, where whatever conversation followed them stayed. The people did not disappear. They moved to places that crawlers cannot read and counters cannot see.

Indoors

The web moves indoors

The move has had names for more than a decade. In 2012 Alexis Madrigal, then at The Atlantic, noticed that most of the visitors arriving at his magazine’s site from other people’s recommendations came with no trace of where they had come from: links pasted into emails, chats and text messages. He called it “dark social”, and estimated that it brought 56.5 per cent of The Atlantic’s social traffic, more than Facebook and Twitter together.[31] The figure overstated it, as he later found that some of the untraceable visits came from Facebook’s own mobile apps, but the idea stuck.[31]

In 2019 two writers described the same thing as a change of climate. Yancey Strickler, a co-founder of Kickstarter, borrowed an image from Liu Cixin’s science fiction: the open web had become a “dark forest”, where speaking in public draws predators, trolls and marketers, so people retreat to “newsletters and podcasts”, “Slack channels, private Instagrams, invite-only message boards”.[32] Venkatesh Rao called the retreat the “cozyweb”: “slacks, messaging apps, private groups”, a space whose content, he warned, is “poorly addressable, poorly searchable, and very vulnerable to bitrot.”[33]

The largest social network said the same thing that year, as a strategy. In March 2019 Mark Zuckerberg wrote that Facebook and Instagram had been “the digital equivalent of a town square”, but that people “increasingly also want to connect privately in the digital equivalent of the living room”, and that “the future of communication will increasingly shift to private, encrypted services.”[34] Six years later Meta’s own data, cited by a federal judge in the antitrust case brought against it, showed how far the shift had gone. In January 2025 Americans spent 17 per cent of their time on Facebook looking at content from friends, and 7 per cent on Instagram; on Facebook that share had fallen by almost a quarter in two years, on Instagram by more than a third. “Users have become far less likely to post publicly,” the judge wrote, “and instead primarily share content using private messages, either in the app or over text.”[35] The head of Instagram, Adam Mosseri, put it plainly in 2025: “there are way more photos and videos shared into DMs than into stories.”[36]

Four kinds of place have taken the conversation in. The first is private messaging. WhatsApp has more than 3 billion monthly users,[27] and Telegram says it has “significantly over” a billion;[37] between them they carry the family, club and work groups that once lived in forums and comment threads. Messages on WhatsApp have been encrypted end to end by default since 2016, group chats included,[38] so what is said there is not indexed and cannot be read by anyone outside. News travels there too: across the 48 countries of the Reuters Institute’s survey, about a fifth of people use WhatsApp for news each week, 21 per cent in Portugal, 36 per cent in Brazil and 46 per cent in India.[39]

Figure 6. Share of people who used WhatsApp for news in the previous week, early 2025, and, where the report gives it, for any purpose. The Indian sample is mainly English-speaking online news users, not the whole population. Source: Reuters Institute Digital News Report 2025.[39]

The second is the conversation with a chatbot. ChatGPT is now the fifth most visited site in the world,[28] and the questions that used to be asked on Stack Overflow, where anyone could read the answer, are now asked in private and answered for one person. The third is the inbox: Substack alone counts more than 35 million subscriptions, free and paid,[30] and much of what they carry never appears on an open page. The fourth is the logged-in platform: closed groups, and community servers such as Discord’s, reported to have more than 200 million monthly users at the end of 2025.[40] Discord’s instructions to search engines exclude its channels from crawling;[41] only this week did the company announce that, from mid-December 2026, some messages from public channels of its larger open servers, those with at least a thousand members, may appear in web search, while “private servers, DMs, and GDMs are always private.”[41]

There is a counter-current, and it shows how much the open remainder is now worth. For years people have added the word “reddit” to their Google searches to find answers written by people; Google’s own head of search acknowledged in 2023 that users were “not quite happy” when Reddit went dark during a protest.[42] From late 2023 Google began showing Reddit far more often, and by mid-2024 Reddit’s visitors who were not logged in, most of them arriving from Google, had grown by 74 per cent in a year.[42] Reddit then licensed the same conversations to Google and OpenAI.[22] The public places that remain have become both more visible and more valuable, and they are being sold.

Lost and gained

What is lost indoors, and what is gained

The first thing lost is memory. The open web was never a good archive: the Pew Research Center found in 2023 that 38 per cent of the pages that existed in 2013 could no longer be reached, a quarter of all pages from the decade had gone, and 11 per cent of the references in English Wikipedia articles led nowhere.[43] Even tweets vanished: 18 per cent of those Pew collected in the spring of 2023 were no longer visible a few months later.[43] But a dead link at least shows that something was there, and much of it survives in archives that crawl the public web. A conversation in a closed group was never public, never crawled and never archived; when the group is deleted or the phone is lost, it leaves no trace at all.

Figure 7. Share of web pages from each year that could no longer be reached in October 2023. The decline is not smooth: pages from 2018 and 2019 fared worse than those from 2017. Sample of about a million pages from Common Crawl. Source: Pew Research Center.[43]

The second is shared knowledge. When a programmer solves a problem in a project’s Discord server, the answer helps the people in that server that day; when the same answer was posted on a public forum, it helped everyone who searched for the same problem for years. A study of programmers’ chat noted that “coding examples, idiomatic patterns, and common mistakes are discussed daily” on Gitter, Slack and Discord, and that “there is a lack of techniques and tools to reliably mine, index, and retrieve coherent content from such platforms.”[44] Tools exist whose only purpose is to copy Discord help threads onto public pages so that search engines can find them. The collapse of new questions on Stack Overflow, shown above, is the same loss seen from the other side.

The third is the ability to see what is spreading. In public, a false claim can be found, counted and answered. In closed groups it can only be studied by researchers who join them. During Brazil’s 2018 election, a team that monitored 364 public political WhatsApp groups found that 85 images later rated false by fact-checkers, fewer than 1 per cent of all the images shared, appeared in 44 per cent of the groups; of the 47 most shared images that made a factual claim, only 10.6 per cent were true.[45] In India, a study of more than 5,000 political groups in 2018 and 2019 classified 13 per cent of the images shared as misinformation.[46] WhatsApp’s response was to make forwarding slower: a limit of five chats at a time, first in India in 2018 and worldwide from January 2019, and, from April 2020, one chat at a time for messages already forwarded many times, which the company said cut such messages by 70 per cent.[47] It cannot read the messages, by design, so it can only slow them down.

The fourth is the supply of public human writing, which the models of the previous sections need. Epoch AI, a research group, estimated in 2024 that the stock of public human text of usable quality is about 300 trillion tokens, roughly words, and that at the rate models are growing it would be fully used between 2026 and 2032.[48] Every conversation that moves from a forum to a group chat is one fewer that can be added to that stock, which brings the scenario of models learning mostly from models a little closer. And the last is measurement itself: a site’s statistics cannot see sharing that happens in private, which is why, on a small blog, most visits arrive with no known origin.

What is gained is real too, and it is the reason people moved. Privacy is the first: messages that cannot be read by a platform cannot be sold, leaked by it or demanded from it. Strickler’s explanation was psychological: “The dark forests grow because they provide psychological and reputational cover”, a place to speak without performing for strangers or being attacked by them.[32] The court in Meta’s case recorded the same preference: “People have increasingly become less interested in blasting out public posts that hundreds of others can see.”[35] There is some evidence that it is better for people as well. A study of 1,910 Facebook users, linked to their actual activity, found that receiving “targeted, composed communication from strong ties” was associated with improved well-being, while scrolling through friends’ broadcasts and collecting likes was not; one of its authors worked at Facebook, and the study was not an experiment.[49]

The trade, then, is between the people inside and everyone else. Those in the group gain trust, privacy and a smaller, kinder audience. Those outside, the stranger with the same question next year, the researcher, the fact-checker, the historian, the search engine and the next generation of models, lose access to what was said. The open web worked because a great many people wrote in public for no one in particular. A web written for friends is warmer, and much harder to find.

What is being tried

Labels, signatures and prices

Three kinds of response are under way. The first is labelling. The Coalition for Content Provenance and Authenticity, whose members include Adobe, Google, Microsoft, OpenAI and Sony, has a standard for “Content Credentials”, a signed record of where an image or file came from and how it was edited.[50] Leica built it into a camera in 2023, and Google into every photo taken by its Pixel 10 phones in 2025.[50] It is not foolproof: Nikon suspended the feature on one camera in 2025 after a researcher found a way to make it sign an image it should not have.[50] How such signatures work, and how they fail, is explained in The Forged Signature.

The second is law. Article 50 of the European Union’s AI Act requires providers of systems that generate text, images, audio or video to mark their output “in a machine-readable format and detectable as artificially generated”, and requires those who publish deepfakes to say so. Most of the article applies from 2 August 2026; under a provisional agreement of May 2026, the marking duty was postponed to 2 December 2026.[51]

The third is price: licensing deals, Cloudflare’s blocking by default and its pay-per-crawl experiment, all attempts to restore the old exchange, in which those who take from a site give something back.[8][22]

Coda

Not dead, but changing hands

Measured, the dead internet theory is wrong where it is most dramatic and close to right where it is most mundane. There is no evidence of a plan, and there are more people online than ever. But about half of all web traffic is automated, about half of new articles are written mainly by machines, more than half of new music uploads on one large service are generated, and the machines that read the most send back the fewest readers. The forums and question sites where people answered each other in public are emptying, partly because the answers now come from models trained on them.

Whether this is a death or a migration depends on what one meant by the internet. If it was the network, it has never been busier. If it was the open, public web, written by people for other people and found through links, a growing part of it is now written by machines, read by machines, and summarised by machines for people who no longer click through. The people are still there, but increasingly in places the rest of the web cannot see.

On method and tools

This piece was written collaboratively with Claude Opus 5.5 (Anthropic): human specification and editorial direction; machine research, analysis and drafting. Five research agents gathered and checked the sources on 28 October 2026; quotations were read in the original except where the list of sources says otherwise. The Wikipedia and Stack Overflow series were pulled from the public Wikimedia and Stack Exchange APIs on the same day, and the figures come from a script, scripts/dead_internet.py. Anthropic, which makes Claude, appears in Figure 1 with the highest crawl-to-referral ratio; the figure reports Cloudflare’s numbers as published. Several sources are companies that sell detection or security products; their figures are samples of their own customers’ traffic or their own tools’ judgements, and are presented as such.

Cover: a poem written and signed by Henri Maillardet’s automaton, about 1800, reproduction by the Franklin Institute, via Wikimedia Commons, public domain, cropped.

Authored by: Luis Matos Ferreira — Physicist, Developer, Writer

Go deeper

Ideas this essay uses without stopping to explain them, and where this blog does. The letters in the text point here; the links go straight to the section.

[a] How a model learns from text

What training on web text does, step by step, and why so much text is needed. How a Language Model Learns, “Walking downhill in fog”.

Related essays on this blog

1. Mostly Machines — this blog’s own visitors, counted three ways: mostly machines.
2. The Accelerant — social media as an accelerant, and generative AI as the second one.
3. The Manufactured Public — fake accounts, bots and influence operations.
4. How a Language Model Learns — what the text taken from the web is used for.
5. The Forged Signature — digital signatures, and what happens when they can be forged.
6. Bargain, Then Rip-Off — why platforms turn on their users.

Sources

1. IlluminatiPirate, “Dead Internet Theory: Most of the Internet is Fake”, Agora Road forum, 5 January 2021.
2. K. Tiffany, “Maybe You Missed It, but the Internet ‘Died’ Five Years Ago”, The Atlantic, 31 August 2021 (not seen in the original; described via the English Wikipedia article “Dead Internet theory”, which also summarises its reception).
3. Thales, “2025 Imperva Bad Bot Report”, press release, 15 April 2025.
4. Thales, “AI-driven bot attacks surged, according to Bad Bot Report”, press release, 29 April 2026.
5. Cloudflare, “Cloudflare Radar 2025 Year in Review”, 15 December 2025 (data from 1 January to 2 December 2025).
6. J. Tomé, “The crawl-to-click gap”, Cloudflare blog, 29 August 2025.
7. S. Longpre et al., “Consent in Crisis: The Rapid Decline of the AI Data Commons”, Data Provenance Initiative, arXiv:2407.14933, July 2024.
8. M. Prince, “Content Independence Day: no AI crawl without compensation!”, Cloudflare blog, 1 July 2025; W. Allen, “Introducing pay per crawl”, 1 July 2025.
9. Graphite (J. Paredes, D. Druck, M. Benson, E. Smith), “AI Now Writes as Many Online Articles as Humans”, 15 May 2026 (replacing its study of 14 October 2025).
10. R. Law, “74% of New Webpages Include AI Content (Study of 900k Pages)”, Ahrefs, 19 May 2025.
11. Deezer, “AI-generated tracks now represent 44% of all new uploaded music”, 20 April 2026, and “AI music exceeds 50 percent of daily uploads”, July 2026; earlier shares as reported by Digital Music News, Music Business Worldwide and MusicTech.
12. Spotify, “Spotify strengthens AI protections”, 25 September 2025.
13. Amazon KDP announcement of 18 September 2023, as reported by Ebook Friendly, “Amazon KDP: publishers can no longer publish more than 3 books a day”, 22 September 2023.
14. NewsGuard, “AI Tracking Center”, updated 23 June 2026.
15. US Federal Trade Commission, “Federal Trade Commission Announces Final Rule Banning Fake Reviews and Testimonials”, 14 August 2024; 16 CFR Part 465, effective 21 October 2024.
16. Europol Innovation Lab, Facing reality? Law enforcement and the challenge of deepfakes, 2022, p. 5 (archived copy), citing N. Schick, Deepfakes: The Coming Infocalypse, 2020; revised version of January 2024; the origin of the figure as traced by J. Walker, Reason, 27 January 2023 (the book not seen).
17. R. Fishkin, “2024 Zero-Click Search Study”, SparkToro, 2 July 2024 (Datos clickstream panel).
18. Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results”, 22 July 2025.
19. M. Miller, “New user trends on Wikipedia”, Wikimedia Foundation, Diff, 17 October 2025.
20. Wikimedia Foundation, pageviews API, monthly aggregate for en.wikipedia.org, agent type “user”, retrieved 28 October 2026; own calculations.
21. Stack Exchange API, /2.3/questions on stackoverflow, total per month, retrieved 28 October 2026; own query.
22. Google, “An expanded partnership with Reddit”, 22 February 2024; Reuters, 22 February 2024, as excerpted by Nieman Lab (the 60 million figure, from one anonymous source); Reddit, “Reddit and OpenAI build partnership”, 16 May 2024; S. Guaglione, “A timeline of the major deals between publishers and AI tech companies in 2025”, Digiday, 1 January 2026.
23. M. Bergen, “New York Times sues OpenAI, Microsoft over use of its stories”, Bloomberg, 27 December 2023 (case 1:23-cv-11195, S.D.N.Y.).
24. I. Shumailov, Z. Shumaylov, Y. Zhao, N. Papernot, R. Anderson and Y. Gal, “AI models collapse when trained on recursively generated data”, Nature 631, 2024, pp. 755–759.
25. M. Gerstgrasser, R. Schaeffer et al., “Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data”, arXiv:2404.01413, 2024.
26. International Telecommunication Union, “Facts and Figures 2025”, press release, 17 November 2025.
27. I. Mehta, TechCrunch, 1 May 2025, reporting Meta’s earnings call of 30 April 2025.
28. Similarweb, “Top Websites Ranking”, September 2026, retrieved 28 October 2026.
29. N. Newman et al., Reuters Institute Digital News Report 2025, “Executive summary and key findings”, 17 June 2025.
30. Tubefilter, “Substack passes 5 million paid subscribers”, 12 March 2025.
31. A. C. Madrigal, “Dark Social: We Have the Whole History of the Web Wrong”, The Atlantic, 12 October 2012, and his 2014 follow-up with Chartbeat (both described from reposts and reports; the originals not opened).
32. Y. Strickler, “The Dark Forest Theory of the Internet”, May 2019.
33. V. Rao, “The Extended Internet Universe”, 24 May 2019.
34. Meta, “A Privacy-Focused Vision for Social Networking”, 6 March 2019; the full note quoted via J. Constine, TechCrunch, 6 March 2019.
35. FTC v. Meta Platforms, Civil Action No. 20-3590 (JEB), memorandum opinion, US District Court for the District of Columbia, 18 November 2025, citing Meta’s internal data for January 2025.
36. A. Mosseri, interview with P. Kafka, Business Insider, 21 May 2025 (republished on Yahoo).
37. TASS, “Telegram’s active users above one billion monthly — Durov”, 19 March 2025 (the company’s own figure).
38. WhatsApp, “End-to-end encryption”, 5 April 2016.
39. N. Newman et al., Reuters Institute Digital News Report 2025, full report, question Q12b and country pages for Portugal, Brazil and India.
40. Storyboard18, “Discord files confidentially for IPO: reports”, 7 January 2026, after Bloomberg (the user figure unattributed in the report).
41. Discord, robots.txt, retrieved 28 October 2026; K. Duncombe-Smith, “Making Content from Discoverable Servers Easier to Discover”, Discord blog, 27 October 2026.
42. K. Hurler, “The Reddit blackout made Google worse”, Gizmodo, 27 June 2023, after CNBC’s report of leaked audio; J. Vanian, CNBC, 14 August 2024, republished by NBC New York.
43. A. Chapekis, S. Bestvater, E. Remy and G. Rivero, “When Online Content Disappears”, Pew Research Center, 17 May 2024.
44. M. Raglianti, “Evolution of Crowd-Sourced Documentation in Developers Discord Conversations”, BENEVOL 2022.
45. G. Resende, P. Melo, H. Sousa, J. Messias, M. Vasconcelos, J. Almeida and F. Benevenuto, “(Mis)Information Dissemination in WhatsApp: Gathering, Analyzing and Countermeasures”, Proceedings of The Web Conference 2019, pp. 818–828.
46. K. Garimella and D. Eckles, “Images and misinformation in political groups: evidence from WhatsApp in India”, HKS Misinformation Review, 7 July 2020.
47. WhatsApp, “More changes to forwarding”, 19 July 2018, updated 21 January 2019, and “Keeping WhatsApp Personal and Private”, 7 April 2020; the 70 per cent figure from a WhatsApp spokesperson, TechCrunch, 27 April 2020.
48. P. Villalobos, A. Ho, J. Sevilla, T. Besiroglu, L. Heim and M. Hobbhahn, “Will we run out of data? Limits of LLM scaling based on human-generated data”, Epoch AI, June 2024.
49. M. Burke and R. E. Kraut, “The Relationship Between Facebook Use and Well-Being Depends on Communication Type and Tie Strength”, Journal of Computer-Mediated Communication 21(4), 2016, pp. 265–281 (abstract only seen).
50. C2PA (members of the steering committee); Adobe, “Leica launches world’s first camera with Content Credentials built-in”, 26 October 2023; Google, “Pixel 10 and Content Credentials”, 10 September 2025; PetaPixel, “Nikon can’t fully solve the Z6 III’s C2PA problems alone”, 22 September 2025.
51. Regulation (EU) 2024/1689 (AI Act), Article 50 and Article 113; William Fry, “EU AI Act omnibus deal reached”, 12 May 2026, on the provisional agreement of 7 May 2026.