Key findings
Prevalence and concentration: In 2016, approximately 1 in 4 Americans visited a fake news website at least once, and the average American read only 5.5 articles on fake news sites. However, this consumption was highly concentrated: approximately 2.5% of web-page views related to hard news topics came from fake news sites. The distribution was skewed—a small number of highly politically engaged individuals consumed the vast majority of fake news content.
Audience characteristics: People who consumed fake news were overwhelmingly those who already consumed heavily skewed political news diets that aligned with their preferred candidate. When segmented by media diet polarization (using Facebook's public sharing data to estimate website slant), consumption of pro-Trump fake news surged dramatically in the rightmost decile (most conservative media diet)—over 60% of those individuals visited at least one pro-Trump fake news site. By contrast, pro-Clinton fake news remained a negligible portion of anyone's information diet.
Belief and credulity: In a related study, Nyhan exposed people to false claims from fake news sites (e.g., claims about the Trump administration's involvement in journalist Jamal Khashoggi's death, or about George Soros funding migrant caravans). People rated stories congenial to their partisan preferences as more accurate than non-congenial ones—a pattern that held not only for fake news but also for hyperpartisan mainstream sites and even legitimate news. This partisan bias in credibility judgment applied across all source types.
Limited evidence of electoral effects: Comparing candidate preferences from July 2016 (before fake news proliferated) to November 2016, people who consumed pro-Trump fake news did not shift differentially toward Trump relative to their pre-exposure preferences. Similarly, there was no measurable association between fake news exposure and turnout conditional on prior voting behavior (2012 turnout). A dose-response analysis splitting consumers into high and low fake news exposure showed no gradient of effect. An experimental study showing people inflammatory fake news headlines found effects on perceived accuracy of the false claims but no measurable impact on vote intention, polarization, or activism.
Role of social media distribution: Facebook was identified as the primary referral source for fake news websites in 2016. This was inferred by checking whether Facebook appeared in users' browsing history immediately before they visited a fake news site—a pattern that stood out compared to how Facebook referred users to hard news sites or other content sources. By 2018, this referral pattern had largely disappeared, suggesting either platform countermeasures or decreased user interest in sharing and seeking fake news.
Broader argument
Nyhan contextualizes these findings within a critique of moral panic about media and technology. He notes that similar catastrophic predictions surrounded radio and television, most of which proved unfounded in retrospect. The effects of new media on public opinion have historically been more nuanced than feared. He argues for moving beyond anecdotal scary stories (the "fake news is stealing the election" narrative) to careful measurement of: (1) how much exposure occurs, (2) who is actually exposed, (3) whether people believe false claims, and (4) whether exposure changes behavior. His data suggests a more complicated picture than the immediate post-2016 panic suggested—fake news is real and concerning but also concentrated, effects are harder to detect than headlines imply, and the problem may be more about intensifying existing partisan divisions among highly engaged voters than about swaying the undecided or changing election outcomes.
Methodological note
The research leverages survey respondents who consented to browser-tracking software, providing an objective measure of website visits rather than relying on self-report (which is error-prone). The limitation is that this data captures desktop/laptop visits but not mobile; similarly, it shows that people visited fake news sites but cannot reveal what they saw in Facebook newsfeeds (Facebook does not share such data externally). The sample is representative of the U.S. adult population on major demographics.
Notes
This talk stands out for its empirical rigor and methodological transparency. Rather than cite anecdotes or aggregate sharing statistics (which obscure who actually consumed content), Nyhan presents data on individual-level exposure via browser tracking. The finding that fake news consumption is concentrated among politically extreme users—not randomly distributed—constrains theories of fake news impact: if effect sizes are small and exposure is concentrated among the already-polarized, the mechanism must be amplification of existing preferences rather than persuasion of the undecided.
The null findings on electoral effects (candidate preference, turnout) are striking given the intensity of post-2016 media coverage suggesting fake news "stole" the election. Nyhan does not conclude that fake news had zero effects—the election margin was close, and elaborate targeted scenarios cannot be ruled out—but rather that measurable persuasive and mobilizing effects are not evident in population-level survey data. This invites a more nuanced conversation about what problem fake news actually poses: less "changing election outcomes" and more "degrading shared epistemic norms" and "intensifying polarization among the politically engaged."
The observation that Facebook's role as a referral source disappeared between 2016 and 2018 hints at platform countermeasures (e.g., deamplification, account removal) or user behavior changes, though the mechanism cannot be determined from the data alone.