When Content Gets Cheap, What Gets Expensive? Five Media Explosions, and a Pattern the Data Will Not Support

Glasp Research, 2026-11-01. The four expectations and the conditions for abandoning them were committed to version control (in a private repository) before any outcome series was read. Ratios are computed at full precision and the printed multiples are rounded.

The question

Every few generations a technology makes content much cheaper to produce, and supply jumps. The printing press, the penny press, radio and the web each did it, and generative AI is doing it now. This post asks one question: each time supply exploded, what became scarce, and did value move toward it?

The answer most people expect comes from Herbert Simon in 1971: "a wealth of information creates a poverty of attention and a need to allocate that attention efficiently among the overabundance of information sources that might consume it" [1]. Michael Goldhaber made it an economy in 1997 [2]. If they are right, every explosion should move value from producing content to selecting it: editors, brands, distribution, recommendation. We wrote four expectations down in advance, with the conditions under which we would declare the pattern not to hold. One of those conditions has fired, which is the opposite of what we expected.

How this was measured

The historical numbers come from one public volume, the 1975 Census compendium Historical Statistics of the United States, Colonial Times to 1970 (HSUS) [4]. Its chapters are scanned images, so 174 values across 16 series were transcribed by eye with OCR as a cross-check, each value recorded with its series code, page, URL and access date, and no human has yet re-read them against the pages. A script builds every derived number from those values, so each figure below can be checked against the source using the series codes in the references. That buys reproducibility and nothing else: the comparison is descriptive, and nothing here separates "supply exploded, so value moved to selectors" from economies of scale, wire costs, spectrum licensing or regulation. Any cause named below is named as untested, the source's definition breaks are in Limitations, and of the five pre-registered historical cases these series cover three.

Mass newspapers: more titles, then fewer titles with more readers

From HSUS [4]. The source mixes "a small number of periodicals" into the 1850 to 1890 rows, so those levels are not pure dailies. The population denominator uses exact census values in census years and linear interpolation between them otherwise. Each line gives daily newspapers, circulation, copies per daily and copies per 1,000 people:

1850: 254 dailies, 758 thousand copies, 2,984 per daily, 33 per 1,000 people
1870: 574 dailies, 2,602 thousand copies, 4,533 per daily, 65 per 1,000 people
1880: 971 dailies, 3,566 thousand copies, 3,673 per daily, 71 per 1,000 people
1909: 2,600 dailies, 24,212 thousand copies, 9,312 per daily, 268 per 1,000 people
1929: 2,086 dailies, 42,015 thousand copies, 20,141 per daily, 347 per 1,000 people
1947: 1,854 dailies, 53,287 thousand copies, 28,742 per daily, 368 per 1,000 people
1963: 1,766 dailies, 63,831 thousand copies, 36,144 per daily, 342 per 1,000 people

From 1850 to 1909 dailies grew 10.2 times and circulation 31.9 times against population 3.97 times: ratios of 2.6 and 8.1. But 1909 is the peak year in titles, the endpoint that maximises that ratio; to 1929, the end of the window the pre-registration named, titles give 1.6 and circulation 10.5. Titles then fell 32 percent from the 1909 peak by 1963 while circulation rose 164 percent. Per head, readership did not keep climbing: copies per 1,000 people peaked at 368 in 1947 and were 342 by 1963, so titles fell by a third while readership roughly tracked population. The 19 percent dip in copies per daily from 1870 to 1880 sits inside the periodicals stretch and may be definitional.

Three dated findings describe what changed over the same decades. No source we found compares surviving papers with those that closed.

Trust became a product feature. "Independent" urban dailies, not tied to a party, rose from 11 percent in 1870 to 62 percent in 1920 [5].
Readers' attention was sold to advertisers. Advertising was 50 percent of newspaper income in 1880 and 64 percent in 1910, a figure Gentzkow, Glaeser and Goldin take from Emery and Emery (1988) [5, secondary for this number]. In an 1880 cross-section of 325 dailies, 10,000 more readers was associated with an advertising rate about 28 percent higher [5].
Brand signals multiplied. US trademark registrations rose from 121 in 1870 to 10,268 in 1920, though the series is volatile and tracks the trademark acts of 1881 and 1905 as much as branding [6].

Advertising stayed a roughly constant share of GNP

Advertising as a percentage of GNP, from HSUS [4]: 1.79 in 1880, 2.90 in 1900, 3.42 in 1909, 3.32 in 1929, 2.09 in 1940, 2.00 in 1950, 2.37 in 1960 and 2.01 in 1970.

Across 15 benchmark years advertising stayed between 1.36 percent (1945, wartime) and 3.42 percent (1909) of GNP, while daily circulation grew more than tenfold and radio and television arrived. Douglas Galbi, on his own site and using the Coen advertising estimates with GDP, finds 1.8 to 2.3 percent outside the 1920s and the war through 2007 and concludes that new media "merely redistributed" advertising rather than increasing it [8, self-published, built on estimates]. That is consistent with attention being the binding constraint, and equally consistent with advertising being a fixed share of commercial activity for unrelated reasons.

Radio and television: the selector's share did not rise

By the end of 1938, 341 of 660 commercial stations were network affiliates, and NBC and CBS controlled "more than 85% of the total night-time wattage" [9]. That is a level in one year, and the series holds no earlier value to compare it with. What it does show, from 1935 on, is the network share falling. In percent, at five-year steps:

Network share of radio advertising, 1935 to 1970: 55.8, 52.3, 46.7, 32.4, 15.4, 6.2, 6.5, 4.5
Network share of TV advertising, 1950 to 1970: 49.7, 52.7, 49.2, 49.2, 46.7
TV share of all advertising, 1950 to 1970: 3.0, 11.1, 13.3, 16.5, 18.7

It is tempting to credit television. The data does not allow it: of the 51.3 point fall from 1935 to 1970, 23.4 points, 46 percent of it, had happened by 1950, when television was 3.0 percent of all advertising and zero in every earlier year. The public record offers an alternative this run cannot test, the FCC chain broadcasting rules of 1941 and RCA's proposed sale of the Blue Network [9]. The TV network share, meanwhile, held between 46.7 and 52.7 percent for twenty years; it did not rise.

The web, where the concentration is documented but not as a rise

Google held more than 90 percent of a £7.3 billion UK search advertising market in 2019 [11]; open-web display was cheap from the start, "on the order of $2 CPM" for large publishers in early 2007 [10]; and Alphabet, Amazon and Meta took 56.1 percent of global ad spend outside China in 2025, later forecast at 59.7 percent for 2026 [12; 31, secondary]. These are levels from different countries, vendors and years, not a measured rise. For print the compiled dataset has nothing at all.

What the pre-registered tests say

The pre-registration states the abandonment condition verbatim: "if in at least two of the five historical cases E3 fails (the intermediary share does not rise within a generation), or E2 fails (unit price of attention rises in real terms as supply explodes), the 'scarcity moves to curation' pattern is not a general law and the article says so." The two-case threshold attaches to E3 only.

E1, supply outgrows attention. Data: against population, titles 2.6, or 1.6 to 1929; circulation 8.1, or 10.5; radio stations 50.1 from 1922 or 2.6 from 1925; TV stations 4.9; US book titles 4.3 but not comparable across the definition breaks, and 0.9 in the definition-consistent 1920 to 1950 window. Against set-owning households, stations per household fell, to 0.25 and 0.43 of the start level. Verdict: proxy-sensitive and not directly tested; it generally holds against population, except for books in the definition-consistent 1920 to 1950 window, and reverses for radio and TV against set-owning households.
E2, the real price of a unit of attention falls. Not tested; see below. Verdict: untested.
E3, the intermediary share rises. Data: computable in two of five cases. Radio 55.8 percent (1935) to 32.4 (1950) to 6.2 (1960); TV 49.7 (1950) to 46.7 (1970). Verdict: fails in both. Stop condition met.
E4, a trust or brand premium appears. Data: independent dailies 11 to 62 percent; trademarks volatile; an 1880 circulation-to-rate association. Verdict: untested. Trust and brand signals became more prevalent, but no premium attributable to them was measured.

E2 is untested rather than refuted. The only per-unit series this run could build, newspaper advertising dollars per daily copy, cannot carry the test, for four reasons. Its numerator covers all newspapers including weeklies and Sunday editions while its denominator counts daily copies only. Annual revenue per copy is not a price per unit of attention unless advertising volume per copy held constant, which nothing here measures. The window, 1935 to 1967, begins 26 years after the title peak. And the sign depends on the deflator: in 1967 dollars it rises 1.64 times, from $45.4 (1935) through a dip to $41.4 (1947) to $74.3 (1967), while as a share of nominal GNP per capita it falls from 3.29 to 1.82 percent.

E3 is the result. M4 as pre-registered, the share of a medium's revenue or audience held by the selecting layer, is computable from these series in exactly two of the five cases: radio and television. Newspaper consolidation is producer concentration, not an intermediary share, and no wire service or chain revenue series was available; print books and the web have none. In both computable cases the share does not rise within a generation of the explosion, which is the pre-registered threshold. The radio series begins after the networks were built, so it cannot show whether their share rose earlier.

So, as pre-registered: the pattern that scarcity moves to curation is not a general law. The published claims that it does are levels and cross-sections from single years and countries, not measured rises. What these series support is narrower: US advertising expenditure stayed within 1.36 to 3.42 percent of GNP at the 15 available benchmarks rather than growing in proportion to measured media supply, and the selecting layer's share can be large without rising.

The AI explosion

Every published share of AI content comes from a commercial detector run by the company reporting it: Surfer for Graphite's 2025 study, Pangram, GPTZero and Copyleaks for its revision, Ahrefs' own bot_or_not, Deezer's own classifier. Graphite's 2025 detector was validated only against GPT-4o output, at a 4.2 percent false positive rate, the same order as the margin by which AI was said to have overtaken human writing.

Web articles. Graphite's October 2025 study classified 43,000 new English articles from Common Crawl with one detector and reported AI passing human writing in November 2024 [14]. Graphite's own page now records that study as superseded; its three-detector revision, reported in May 2026, puts the primarily-AI share at 49.9 percent in Q1 2026, level with human writing rather than ahead [15, secondary; Axios returned 403 in this run].
Web pages. Ahrefs, running its own detector on 900,000 new pages sampled one per domain in April 2025, found AI content in 74.2 percent of them; only 2.5 percent were purely AI and 71.7 percent mixed [17].
Music. Deezer reports fully AI-generated tracks passing 50 percent of daily uploads in June 2026 at about 90,000 a day, a figure the same release elsewhere gives as a monthly average, against about 10,000 a day early in 2025 [18, the 2025 figure secondary].
A projection, not a measurement. The claim that 90 percent of online content would be synthetic by 2026 came from a 2022 Europol report quoting unnamed experts; Europol's January 2024 replacement no longer contains the sentence [19].

The pre-registered 10 percent supply threshold is exceeded in each measured category, though these samples do not estimate AI's share of all new content. Consumption is another matter. Deezer reports AI tracks at over half of daily uploads but 1 to 3 percent of streams [18]. Those shares have different denominators, a daily flow against plays across a catalogue that is overwhelmingly pre-AI, so the gap is indicative only. For text there is no consumption measure at all; the closest is Graphite's finding that 86 percent of articles ranking in Google Search are human-written, a ranking share, from the single-detector method the 2026 revision corrected [16, secondary].

Where value is moving

Selection and distribution, already concentrated, and now shifting inside search. Pew's March 2025 panel of 900 US adults found users clicked a regular result in 8 percent of visits where an AI summary appeared against 15 percent without one, a comparison of visit types rather than before and after [20]. Google organic referrals to publishers fell 33 percent globally and 38 percent in the US from November 2024 to November 2025, on Chartbeat data across more than 2,500 sites [21].
Human communities as licensed supply. Reddit's revenue went from $804.0M (2023) to $2,202.5M (2025), and its 10-K attributes the growth in other revenue to content licensing agreements executed in 2024 and 2025 [22]. Its January 2024 contracts were reported at $203.0M over two to three years [23, secondary]; other deals are reported but unconfirmed, News Corp and OpenAI at possibly over $250M over five years [24, secondary] and Dotdash Meredith at at least $16M a year [25, secondary].
Preference without a measured price premium. In a YouGov survey across Argentina, Denmark, France, Japan, the UK and the US, 62 percent are comfortable with news made entirely by humans and 12 percent with news made entirely by AI [27]. Trust in news overall is 37 percent, the lowest since the measure began in 2015 [28]. We found no public data showing people pay more for human-made content, and provenance marking is spreading without one: Google's Pixel 10 attaches C2PA Content Credentials to every photo from its camera app [26]. Open-web prices did fall in 2025, on two national vendor barometers not comparable to each other: French display CPM down 8 percent, UK eCPM down 10.54 percent [13, secondary].
Not tested. The pre-registration named the opposite possibility, that value moves to the generators while platform share stays flat. No model-maker revenue figure was compiled, so that is neither met nor ruled out.

Where the analogy breaks

In earlier explosions the party that ranked content was usually not the party that made it: stations were not the networks, websites were not Google. Now the companies that rank content also generate it, and an AI answer sits between the reader and the source. Producer-owned selection is not itself new, as the AP was owned by the newspapers it supplied [7, secondary] and the first English review periodical was founded by a bookseller [32, secondary]; what is new is that the ranking party generates a substitute for the thing it ranks. The Pew and Chartbeat figures are consistent with pressure on the pipe that historically paid producers, and with selectors keeping bargaining power while using producer content as source material and as training or retrieval data, but neither isolates AI as the cause and this run does not measure where value accrued.

One smaller break is speed. Taking 1450 as the conventional starting date for European printing, the first scholarly review journal followed in 1665 [3, secondary], about 215 years later, and radio network affiliation reached the levels recorded for 1938 by the FCC within the early network-radio period [9], while AI went from about 10,000 fully AI tracks a day in early 2025 to about 90,000 in June 2026 on one platform [18, the 2025 figure secondary]. Some filtering work now falls on platforms and volunteer communities: a preprint flagged over 5 percent of new English Wikipedia articles in August 2024 using a detector set at a 1 percent false positive rate [29], and English Wikipedia adopted a speedy-deletion criterion for unreviewed LLM articles in August 2025 [30, secondary]. Whether that burden favours the largest platforms is a hypothesis; no filtering-cost series was found for any era.

What this suggests

A selecting layer can hold half a medium's revenue for decades without its share rising, and can lose almost all of it for reasons a descriptive comparison cannot name. So "value moves to curation" is better treated as a hypothesis with two failed tests behind it than as a law. Three numbers are worth watching: the AI share of new supply against the AI share of consumption, remembering the mismatched denominators; the share of ad spend held by the top platforms; and publisher referral traffic from search.

Limitations

The historical series are American, from one volume, transcribed by eye from scanned pages and not yet re-read by a human, so transcription error is possible; the series codes in reference 4 let any value be checked against the source.
Four definition breaks affect the headline numbers: periodicals in the 1850 to 1890 newspaper counts, pamphlets in the 1880 to 1919 book titles with further breaks at 1959 and 1967, authorized rather than on-air radio stations for 1933 to 1948, and a decade-average GNP at 1880. Whether 1890 GNP is also an average is unverified.
Population and set-owning households are weak stands-in for attention, and no time-use series was compiled, so the pre-registered attention denominator is unmeasured. Only one per-unit price series could be built, and it cannot test E2. M4 was computable in two of five cases.
Cambridge Core, the AEA PDF server, Taylor and Francis, Axios and Getty's newsroom blocked automated access and were skipped rather than worked around. Axios being blocked is why the 49.9 and 86 percent figures are secondary here; Cambridge Core being blocked is why the per-century book totals are unused.
Several AI-era deal values are press reports and are marked secondary, and every AI supply share comes from a commercial detector run by the party reporting it.
This is a descriptive comparison. It does not show that supply growth caused value to move, in either direction.

AI-use disclosure

This post was researched and written by an AI system (Claude, run by Glasp as an automated research task), which searched public sources, transcribed public tables, wrote the script and drafted the text. No Glasp-internal data were used, and no personal data appears in it. Independent AI reviews checked the methods and the sources. A human review of the numbers against the cited pages is pending.

References

Simon, H. A. (1971). Designing Organizations for an Information-Rich World. In M. Greenberger (ed.), Computers, Communications, and the Public Interest, Johns Hopkins Press, pp. 40-41. https://gwern.net/doc/design/1971-simon.pdf
Goldhaber, M. H. (1997). The attention economy and the Net. First Monday 2(4). https://doi.org/10.5210/fm.v2i4.519
History of Information, Journal des sçavans (secondary). https://www.historyofinformation.com/detail.php?id=2281
U.S. Bureau of the Census (1975). Historical Statistics of the United States, Colonial Times to 1970. Series R 93-105, R 192-217, R 244-257; T 444-471; F 1-5; E 135; A 2. All accessed 2026-10-30. https://www2.census.gov/library/publications/1975/compendia/hist_stats_colonial-1970/hist_stats_colonial-1970p2-chR.pdf https://www2.census.gov/library/publications/1975/compendia/hist_stats_colonial-1970/hist_stats_colonial-1970p2-chT.pdf https://www2.census.gov/library/publications/1975/compendia/hist_stats_colonial-1970/hist_stats_colonial-1970p1-chF.pdf https://www2.census.gov/library/publications/1975/compendia/hist_stats_colonial-1970/hist_stats_colonial-1970p1-chE.pdf https://www2.census.gov/library/publications/1975/compendia/hist_stats_colonial-1970/hist_stats_colonial-1970p1-chA.pdf
Gentzkow, M., Glaeser, E., and Goldin, C. (2004). The Rise of the Fourth Estate. NBER WP 10791. https://www.nber.org/papers/w10791
Historical Statistics of the United States, Millennial Edition, Table Cg108-109 (trademarks registered and renewed, 1870-2002). https://hsus.cambridge.org/HSUSWeb/toc/treeTablePathIdCg108-109.html
Britannica, Associated Press (secondary). https://www.britannica.com/topic/Associated-Press
Galbi, D. (2008). U.S. Advertising Spending as % of GDP (self-published; built on the Coen Structured Advertising Expenditure Dataset, itself a set of estimates). https://galbithink.org/ad-spending.htm Galbi, D. (2009). The effects of new advertising media. https://www.purplemotes.net/2009/06/07/the-effects-of-new-advertising-media/
NBC v. United States, 319 U.S. 190 (1943). https://www.law.cornell.edu/supremecourt/text/319/190
Evans, D. S. (2008). The Economics of the Online Advertising Industry. Review of Network Economics 7(3). https://www.cs.princeton.edu/courses/archive/spr11/cos448/web/docs/week5_optional1.pdf
UK Competition and Markets Authority (2020-07-01). Online platforms and digital advertising, final report. https://assets.publishing.service.gov.uk/media/5fa557668fa8f5788db46efc/Final_report_Digital_ALT_TEXT.pdf
WARC (2025-12-11). Global ad market prospects upgraded. https://www.warc.com/en/press/press-releases/25-12-11_global-ad-market-prospects-upgraded-to-8-but-growth-concentrated-in-Big-Tech-platforms
Journal du Net (2026-02-12), Alliance Digitale barometer (secondary). https://www.journaldunet.com/adtech/1547943-programmatique-la-degringolade-du-display-et-de-la-video-sur-l-open-web/ ExchangeWire (2026-02-11), Pubstack State of Programmatic UK 2025 (secondary). https://www.exchangewire.com/blog/2026/02/11/state-of-programmatic-uk-2025-volume-up-value-down-for-uk-publishers/
Graphite (2025-10-14). More articles are now created by AI than humans. Superseded, per a notice on the page itself. The page URL was not re-fetched in this round, so that it still resolves to the cited study is unverified. https://graphite.io/five-percent/more-articles-are-now-created-by-ai-than-humans
Axios (2026-05-15), reporting Graphite's three-detector revision (secondary; returned 403 in this run). https://www.axios.com/2026/05/15/human-vs-ai-written-articles
Axios (2025-10-14), reporting Graphite's search-ranking figure (secondary; returned 403 in this run). https://www.axios.com/2025/10/14/ai-generated-writing-humans
Ahrefs (2025-05-19). What percentage of new content is AI-generated? https://ahrefs.com/blog/what-percentage-of-new-content-is-ai-generated/
Deezer (2026-07-21). AI music exceeds 50 percent of daily uploads. https://newsroom-deezer.com/2026/07/ai-music-exceeds-50-percent-daily-uploads-deezer/ Deezer (2026-04-20). AI-generated tracks represent 44% of new uploaded music. https://newsroom-deezer.com/2026/04/ai-generated-tracks-represent-44-of-new-uploaded-music/
Europol Innovation Lab, Facing reality? Law enforcement and the challenge of deepfakes (2022; January 2024 revision, which no longer contains the sentence). https://www.europol.europa.eu/cms/sites/default/files/documents/Europol_Innovation_Lab_Facing_Reality_Law_Enforcement_And_The_Challenge_Of_Deepfakes.pdf Original sentence via Futurism (2022, secondary). https://futurism.com/the-byte/experts-90-online-content-ai-generated
Pew Research Center (2025-07-22). Google users are less likely to click on links when an AI summary appears. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
Reuters Institute (2026-01-12). Journalism, Media and Technology Trends and Predictions 2026. https://reutersinstitute.politics.ox.ac.uk/journalism-media-and-technology-trends-and-predictions-2026
Reddit, Inc. Form 10-K for FY2025. https://www.sec.gov/Archives/edgar/data/1713445/000171344526000022/rddt-20251231.htm
TechCrunch (2024-02-22), Reddit licensing (secondary). https://techcrunch.com/2024/02/22/reddit-says-its-made-203m-so-far-licensing-its-data
Engadget (2024-05-22), News Corp and OpenAI (secondary; WSJ wording is "could be", never confirmed by the parties). https://www.engadget.com/openai-will-reportedly-pay-250-million-to-put-news-corps-journalism-in-chatgpt-214615249.html
Engadget (2024-11), Dotdash Meredith and OpenAI (secondary). https://www.engadget.com/ai/openai-will-pay-dotdash-meredith-at-least-16-million-per-year-to-license-its-content-191756662.html
Google (2025-09-10). Trusted images with C2PA Content Credentials. https://blog.google/security/pixel-android-trusted-images-c2pa-content-credentials/
Reuters Institute (2025-10-07). Generative AI and News Report 2025 (YouGov, June to July 2025; Argentina, Denmark, France, Japan, UK, US). https://reutersinstitute.politics.ox.ac.uk/generative-ai-and-news-report-2025-how-people-think-about-ais-role-journalism-and-society
Reuters Institute (2026-06). Digital News Report 2026. https://reutersinstitute.politics.ox.ac.uk/sites/default/files/2026-06/DNR 2026 FINAL_2.pdf
Brooks, C., Eggert, S., and Peskoff, D. (2024). The Rise of AI-Generated Content in Wikipedia. arXiv preprint, not peer reviewed. https://arxiv.org/abs/2410.08044
English Wikipedia, Artificial intelligence in Wikimedia projects, speedy-deletion criterion G15 (secondary). https://en.wikipedia.org/wiki/Artificial_intelligence_in_Wikimedia_projects
PPC Land, reporting WARC's Q3 2026 forecast (secondary). https://ppc.land/warc-forecasts-social-media-ad-spend-gains-21-3-to-394-6bn-in-2026/
Wikipedia, The Monthly Review (London), founded 1749 by the bookseller Ralph Griffiths (secondary). https://en.wikipedia.org/wiki/The_Monthly_Review_(London)