Every figure in the essay comes from one of these sources. Each entry says what it shows and what it does not, and is sorted by the kind of claim it supports, so that an observed change is never dressed up as a measured effect.
[3] 1981Observed change
Museum of Broadcast Communications. Encyclopedia of Television.
The combined audience share of ABC, CBS and NBC was 90 for the 1980-81 programming year, and had fallen to 60 by 1993-94.
What it does not show. United States only. The entry does not say whether the figure is prime time only.
[4] 1983Observed change
Nielsen; Associated Press; Guinness World Records. Nielsen Insights, February 2010; Guinness World Records.
The M*A*S*H finale on 28 February 1983 averaged 105.97 million viewers with a 77 share of homes watching television, and stayed the most-watched US broadcast until February 2010.
What it does not show. One programme on one night; it shows the ceiling of shared viewing, not an ordinary evening.
https://www.nielsen.com/insights/2010/super-bowl-xliv-most-watched-super-bowl-of-all-time
[8] 2009Mechanism
Daniel Fleder; Kartik Hosanagar. Management Science 55(5), 697-712.
Common recommenders can raise each individual's diversity while lowering aggregate diversity: 'they often push users toward the same products.'
What it does not show. Analytical modelling and simulation, not observation of a live platform.
[6] 2015Measured outcome
Carlos A. Gomez-Uribe; Neil Hunt. ACM Transactions on Management Information Systems 6(4), article 13.
Netflix's recommender 'influences choice for about 80% of hours streamed at Netflix. The remaining 20% comes from search.'
What it does not show. The company's own unaudited figure, written by two of its executives.
[7] 2018Measured outcome
Joan E. Solsman, reporting Neal Mohan. CNET, 10 January 2018.
YouTube's chief product officer said at CES 2018 that recommendations drive more than 70 percent of the time people spend watching.
What it does not show. A reporter's paraphrase of a company statement; not independently measured.
https://www.cnet.com/tech/services-and-software/youtube-ces-2018-neal-mohan/
[9] 2014Measured outcome
Tien T. Nguyen; Pik-Mai Hui; F. Maxwell Harper; Loren Terveen; Joseph A. Konstan. Proceedings of WWW 2014, 677-686.
Recommenders expose users to a slightly narrowing set of items over time, and users who actually follow the recommendations narrow less.
What it does not show. One film-rating community; the narrowing it measures is small.
[10] 2020Measured outcome
Ashton Anderson; Lucas Maystre; Ian Anderson; Rishabh Mehrotra; Mounia Lalmas. Proceedings of The Web Conference 2020, 2155-2165.
Across more than 100 million users, algorithmically driven listening is associated with reduced diversity. Users who become more diverse do it by shifting toward organic listening. Generalists are up to 25 points less likely to leave and up to 35 points more likely to subscribe.
What it does not show. Association within one platform's logs, written by its own researchers.
[11] 2014Measured outcome
Kartik Hosanagar; Daniel Fleder; Dokyun Lee; Andreas Buja. Management Science 60(4), 805-823.
On real purchase data, personalisation did not fragment users: it widened their interests and created more commonality with others.
What it does not show. Music purchases with one recommender; a counterexample to the strong echo-chamber claim.
[12] 2018Measured outcome
Hannes Datta; George Knox; Bart J. Bronnenberg. Marketing Science 37(1), 5-21.
Adopting Spotify raised plays and the number of distinct artists heard: unique artists stayed about 32 percent higher after six months, and new-music discovery rose.
What it does not show. Counts how much variety a person samples, not whether anyone else heard the same thing.
[13] 2010Measured outcome
Victoria J. Rideout; Ulla G. Foehr; Donald F. Roberts. Henry J. Kaiser Family Foundation, January 2010.
71 percent of American 8- to 18-year-olds had a TV in their bedroom, and they spent 7 hours 38 minutes a day with entertainment media.
What it does not show. A 2008-2009 survey, before smartphones were universal; the private screen has since moved into the pocket.
[14] 1979, 2010Measured outcome
A. C. Nielsen Company; The Nielsen Company. Nielsen, 1979; Nielsen Insights, April 2010.
In 1978 an estimated 48 percent of US television homes had two or more sets. By 2010 the average home had 2.93 sets, and only 17 percent had just one.
What it does not show. Counts sets, not who watched them together.
https://www.nielsen.com/insights/2010/u-s-homes-add-even-more-tv-sets-in-2010/
[15] 2019Measured outcome
GfK MRI, reported by Wayne Friedman. MediaPost, 15 January 2019.
48 percent of US television viewing time was co-viewing in 2019.
What it does not show. A trade survey of the television set only; phones and laptops are outside it.
[18] 2012Measured outcome
James G. Webster; Thomas B. Ksiazek. Journal of Communication 62(1), 39-56.
Nielsen data across 236 TV and internet outlets shows 'extremely high levels of audience duplication', overlapping attention rather than isolated groups of loyalists.
What it does not show. Measures whether audiences overlap across outlets, not whether they watched the same thing at the same time.
[19] 2025Measured outcome
Nielsen. Nielsen News Center, 17 June 2025.
In May 2025 streaming was 44.8 percent of US television viewing, more than broadcast (20.1) and cable (24.1) combined, for the first time.
What it does not show. United States, one month; streaming includes live sport on streaming services.
https://www.nielsen.com/news-center/2025/streaming-reaches-historic-tv-milestone
[20] 2022Measured outcome
Ofcom. Ofcom, 17 August 2022.
In 2021, UK 16- to 24-year-olds spent 53 minutes a day with broadcast television; people over 65 spent 5 hours 50 minutes, about seven times more.
What it does not show. United Kingdom, 2021 data; broadcast TV only, not all shared viewing.
[5] 1971 to 1999Law
Canadian Radio-television and Telecommunications Commission. CRTC public notices.
Canadian content rules for AM radio took effect in January 1971, using the MAPL test (music, artist, performance, lyrics). The requirement for commercial stations rose from 30 to 35 percent of popular music from 3 January 1999.
What it does not show. A quota on what is played, not a measure of what listeners discovered.
[16] 2022Law
European Parliament and Council of the European Union. Official Journal of the European Union, L 277.
Article 27 requires platforms to explain the main parameters of their recommender systems. Article 38 requires very large platforms to provide 'at least one option for each of their recommender systems which is not based on profiling'.
What it does not show. A right to switch off profiling, not a right to shape the ranking.
[17] 2023Law
Parliament of Canada. Statutes of Canada 2023, chapter 8, assented to 27 April 2023.
Brings online streaming under the Broadcasting Act, and lets the CRTC set conditions on 'the presentation of programs ... for selection by the public, including the showcasing and the discoverability of Canadian programs'.
What it does not show. Discoverability for Canadian programming specifically, not unfamiliar culture in general.
https://laws-lois.justice.gc.ca/eng/annualstatutes/2023_8/FullText.html
[1] 2001 to 2017Argument
Cass R. Sunstein. Princeton University Press; New York Review of Books, 25 April 2002.
Argues that a system of free expression needs 'unanticipated, unchosen encounters and a range of shared experiences', and calls for an 'architecture of serendipity'.
What it does not show. A normative argument about what a democracy needs, not a measurement of exposure.
[2] 2001Argument
Michael Suk-Young Chwe. Princeton University Press, 2001; reissued 2013.
Coordination needs common knowledge: not only that people saw something, but that each knows the others saw it. Super Bowl advertising is valuable because viewers know many others are seeing precisely what they see.
What it does not show. A theory of coordination, illustrated with cases. It does not measure how much common knowledge a society has.
https://press.princeton.edu/books/paperback/9780691158280/rational-ritual