Engagement amplification

What it looks like

This is the tactic on the list that nobody performs. There is no author. A ranking system predicts what will get a response and orders your feed by that prediction, and because hostility gets a response, the most divisive available version of a story is systematically promoted.

The consequence for a reader is not that the feed lies. It is that the feed is a sample, and the sample is biased in a direction you can name.

Why it works

The cleanest audit is a preregistered study in which 806 people had their engagement-ranked timeline captured alongside their own reverse-chronological one, and rated the posts from each. Against the chronological baseline, the ranked feed amplified out-group animosity, partisanship, and anger both expressed by authors and felt by readers.1 The finding that matters most: participants did not prefer the political posts the algorithm selected for them. Engagement and satisfaction came apart.

What the research does not support is the stronger popular claim. In 2023 a set of large studies run with Meta’s cooperation switched consenting users to a chronological feed for three months during the 2020 US election. Doing so substantially changed what people saw and how long they spent — and did not significantly move measured polarization, political knowledge or key attitudes over that window.2 A companion paper studying 208 million users found that ideological segregation rises at every step from potential exposure to actual exposure to engagement.3 Curation concentrates what you see; three months of un-curating it did not move survey-measured attitudes.

Those results carry two documented caveats, and a site about verification has to state them. The project’s own independent rapporteur judged the work rigorous but argued it should not be the model for future research, because it was conducted under "independence by permission" from the platform.4 And a 2024 letter led by researchers at Indiana University noted that Meta deployed temporary emergency ranking changes in November 2020 — inside the study window — which were rolled back the following March.5 The headline null result may describe an algorithm that no longer existed three months later.

Documented cases

A ranking change with a stated pro-social purpose

In 2018 Facebook reweighted its feed toward what it called meaningful social interactions, intended to favour exchanges between friends and family. Internal documents later reported by journalists recorded staff concluding the change had "unhealthy side effects on important slices of public content, such as politics and news," and recorded publishers telling the company that the ranking was not rewarding what it claimed to reward.6

This is reporting on leaked internal documents rather than peer-reviewed research, the documents are partial, and the company disputed parts of the characterisation. It is included because the mechanism it describes is the one this page is about: nobody decided to promote divisiveness. The weights did.

A platform measuring, and reducing, its own recommender

In January 2019 YouTube announced it would reduce recommendations of "borderline content and content that could misinform users in harmful ways," giving as its own examples videos promoting a phony miracle cure and content claiming the earth is flat.7 In December 2019 it reported a 70 percent average drop in watch time of that content arriving from recommendations to non-subscribers in the US.8

That is a company stating in its own words that its recommender had been driving substantial watch time to health misinformation, and that it could turn that down when it chose to. The 70 percent figure is the company’s own, unaudited, with no published baseline volume, and no external party has replicated it.

What to do about it

• Assume you are seeing the tail, not the distribution. A post reached you because it was predicted to get a reaction. It is not a sample of what most people think or what most coverage says.
• Go to the publication’s own front page. Reading a story in its outlet’s ordering rather than the feed’s removes the selection filter at no cost.
• Switch to a chronological or following-only feed where one exists. It reliably changes what you see and how long you spend. On the evidence, treat it as an exposure control rather than a cure.
• When you see the angriest possible framing of a story, go find it reported plainly and see what survives the translation.
• Do not read engagement counts as consensus. In the one study that asked, users did not even prefer the content the ranking selected for them.

Sources

Every figure on this page comes from one of these. Each entry names the publisher, the date and the sample size, so you can check it rather than take our word for it.

1. Milli, S., Carroll, M., Wang, Y., Pandey, S., Zhao, S. & Dragan, A. — "Engagement, user satisfaction, and the amplification of divisive content on social media"PNAS Nexus 4(3):pgaf062 · 5 March 2025 · preregistered, n=806https://academic.oup.com/pnasnexus/article/4/3/pgaf062/8052060
2. Guess, A. M. et al. — "How do social media feed algorithms affect attitudes and behavior in an election campaign?"Science 381(6656):398–404 · 27 July 2023 · 20,000+ consenting users, 3-month windowhttps://doi.org/10.1126/science.abp9364
3. González-Bailón, S. et al. — "Asymmetric ideological segregation in exposure to political news on Facebook"Science 381(6656):392–398 · 27 July 2023 · 208 million US usershttps://doi.org/10.1126/science.ade7138
4. Wagner, M. W. — "Independence by permission"Science 381(6656):388–391 · 27 July 2023 · the project’s independent rapporteurhttps://doi.org/10.1126/science.adi2430
5. Grabowicz, P., Menczer, F. et al. — letter on the temporary ranking changes inside the study windowScience eLetter · 27 September 2024 · summarised by the Observatory on Social Media, Indiana Universityhttps://osome.iu.edu/research/blog/science-eletter-social-media-algorithms-can-curb-misinformation-but-do-they
6. Hagey, K. & Horwitz, J. — "Facebook Tried to Make Its Platform a Healthier Place. It Got Angrier Instead."The Wall Street Journal, 15 September 2021, as submitted to the US House Committee on Energy and Commerce · reporting on leaked internal documents, not peer-reviewedhttps://docs.house.gov/meetings/IF/IF16/20211201/114268/HHRG-117-IF16-20211201-SD012.pdf
7. YouTube — "Continuing our work to improve recommendations on YouTube"YouTube Official Blog · 25 January 2019https://blog.youtube/news-and-events/continuing-our-work-to-improve/
8. YouTube — "The Four Rs of Responsibility, Part 2: Raising authoritative content and reducing borderline content"YouTube Official Blog · 3 December 2019 · company-reported figure, unauditedhttps://blog.youtube/inside-youtube/the-four-rs-of-responsibility-raise-and-reduce/
9. Merrill, J. B. & Oremus, W. — "Five points for anger, one for a ‘like’"The Washington Post · 26 October 2021 · reporting on leaked internal documents, not peer-reviewedhttps://www.washingtonpost.com/technology/2021/10/26/facebook-angry-emoji-algorithm/