Every time I take this framework into a room, someone offers media literacy. It sounds reasonable and costs the person proposing it nothing, which is part of why it arrives so early. I have spent enough time being annoyed by the reply to owe it the strongest version of itself, so here is the strongest version, followed by what I think is wrong with it.
The strongest version is a 2020 paper in PNAS by Andrew Guess and six colleagues. They took the intervention behind the largest media literacy campaign in the world, a set of tips on spotting false news that had been pushed to users in fourteen countries, and tested it properly: preregistered survey experiments run around recent elections in the United States and India. Discernment between mainstream and false headlines improved by 26.5% in a nationally representative American sample and by 17.5% in a highly educated Indian online sample. In the United States the gain was still measurable several weeks later. That is a real result produced by serious people, and I am not going to wave it away.
The mechanism is where it gets uncomfortable, and the authors state it plainly. The intervention reduced the perceived accuracy of mainstream headlines as well as false ones. It simply reduced the false ones more. Discernment improved as arithmetic. The tips lowered trust across the board, and the gap widened in the preferred direction because the two effects were unequal. Nobody came away with a way to recognize a true story.
The programs studied here work by making people believe less, and then hope the arithmetic favors the truth.
Once you are looking for that shape, it turns up elsewhere in the literature. In 2023 Ariana Modirrousta-Galian and Philip Higham reanalyzed five studies of two inoculation games, Bad News and Go Viral!, in the Journal of Experimental Psychology: General. Earlier work had compared how participants rated headlines before and after playing, and reported gains. What that work had not done was separate discrimination, the ability to tell true from fake, from response bias, the overall tendency to answer “fake” to anything. Using ROC curves from signal detection theory, the reanalysis found that where comparable true and fake items were used, the games did not improve discrimination. They produced more “false” responses to everything. The authors’ own word for the possible net effect is counterproductive.
The obvious objection is that this is a Western finding about Western games, and it has been tested. A 2023 study had reported that Bad News improved discrimination in an Indian sample. In 2025 Tina Seabrooke, with the same two authors, ran a preregistered replication with Indian participants and counterbalanced which headlines appeared before and after the game. With the counterbalancing in place, neither discrimination nor bias moved significantly. Run the way the original had been run, discrimination still did not shift, and the conservative bias did.
Now set that beside week fourteen. Chesney and Citron’s sharper claim was that the liar’s dividend grows as the public becomes better educated about fakes, because a better-informed audience makes each subsequent denial more plausible. If the working ingredient in these programs is generalized doubt, those programs are manufacturing the raw material the dividend runs on. A population trained to distrust everything can be told that anything is fabricated.
There is a distributional problem underneath as well. Guess’s team found no effect at all in a representative sample from a largely rural part of northern India where social media use is far lower, and the Indian gains did not persist the way the American ones did. Whatever gains there were went to people already online.
The rubric reading is straightforward. Media literacy is denominator work aimed at R. Last week’s post argued that R is institutional and slow, assembled out of teacher quality, press freedom, and public trust over decades. Media literacy as it is usually proposed is neither institutional nor slow. It is a leaflet, a game, a classroom module and the modest budget line that goes with it. It is popular because it is cheap, requires no regulation, blames no platform, and offends nobody in the room where the decision gets made. Those properties get a proposal approved. None of them makes it work.
None of this argues for doing nothing. The version worth paying for gives doubt somewhere to go: provenance, a checkable chain from a claim back to its origin, verification a reader can actually perform. The three questions Kari Kivinen describes in the interview cited last week, from the FaktaBaari school materials he worked on, ask who is behind a claim, what evidence supports it, and what other sources say. That is literacy terminating in an action. Tips that terminate in suspicion have produced, in the strongest study this note has walked through, a public that believes less of everything.
Next week: Herbert Simon in 1971, and the half of his sentence that usually gets cut.
— J.W.B.
Note 16 of 24