Propaganda is dangerous, but not because it is persuasive

Reasonable People #52: I pick at the claim that propaganda "doesn't work".

Hugo Mercier has claimed Propaganda (Almost) Never Works. Let’s look at what he means and see if we can be convinced.

This is a short version of his claim:

“when it comes to mass persuasion attempts, failure is the rule rather than the exception….advertising, authoritarian speeches, and political campaigns all fail to convince the overwhelming majority of their audience”

(quoted from the short persuasion.community piece linked above)

Mercier has also written a scholarly article How gullible are we? A review of the evidence from psychology and social science (2017) and a book Not Born Yesterday: The Science of Who We Trust and What We Believe. These give additional context, and broaden the scope his argument, but the sections on propaganda rely on the same evidence.

The general account is that people are often portrayed as victims of mass persuasion attempts, but wherever we look for evidence, mass persuasion looks like it is usually ineffective.

Evidence cited includes:

• This study, which suggests that most TV advertising campaigns have an effect which is indistinguishable from zero.

• This study, a meta-analysis of 49 field experiments in political campaigning, which suggests an typical effect of zero.

Perhaps most compelling is the account of Nazi propaganda. The Nazi are often portrayed as the inventors of modern propaganda, weaving a dark spell over the ordinary people which allowed them to take power, keep power and pursue their nightmare vision.

Not so, says Mercier, and he has Ian Kershaw, pre-eminent historian of Nazi Germany, on his side. Kershaw notes that historians have assumed Nazi propaganda had a “monopoly hold over a largely defenceless population”, but using records from the Nazi era, he shows that the Nazi’s themselves were frustrated by the *lack* of success their propaganda efforts had, often pivoting away from unsuccessful messages rather than being able persuade people. Policies such as euthanasia for the handicapped were abandoned after popular resistance, resistance undampened by extensive propaganda. Nazi propaganda failed to make people antisemitic who weren’t already, and failed to make people like the Nazi party, claims Kershaw.

Other studies concur. This study of influences on antisemitism among Germans who were young during the Nazi era concludes that schooling, not media propaganda, was the driver. Another striking example recruited by Mercier is this study - Radio and the Rise of The Nazis in Prewar Germany - which used the differences in radio signal strength in different German cities as a way of quasi-experimentally testing the effect of Nazi propaganda. They show that availability of Nazi radio propaganda didn’t drive antisemitic incidents 1929–1934, rather it was the availability of radio propaganda in cities which already had a larger proportion of Nazi voters where antisemitic incidences occured. In other words, the propaganda mobilised people, but it didn’t persuade people who weren’t already antisemitic.¹

How convincing should we find this?

First, let’s be careful about the strength of the claim. It isn’t that propaganda has no persuasive effect. The claim is merely to rule out a large or common persuasive effects. The evidence cited acknowledges this: both the advertising and political campaigning reviews discuss exceptions to the typical “zero effect” pattern: some brands do run successful ad campaigns, and some political campaigns are persuasive. The TV ad review says that 80% of brands get zero effect from advertising, so presumably the evidence is compatible with a non-zero effect for 20%². The political campaigning review says that there is evidence of campaign persuasion when campaigns focus on persuadable voters AND the opposition has adopted an unpopular position (e.g. the opponent has come out as pro-life and the campaign focusses on finding pro-choice voters and these voters against that candidate).

Small effects can be very consequential - if you’re a firm with small profit margins or a politician in an election which is decided by a tiny percentage of voters. Even marginal persuasive effects could be very consequential. It depends a lot on what else is operating.

Second, let’s be careful about the scope of the claim.

Mercier allows that propaganda has other functions than persuasion, and in the book goes into more detail of what those might be. Propaganda may not persuade, or at least persuade with some magically strong effect, but dispelling that illusion is only part of recognising what it does do. Borrowing from Mercier and others, my ad hoc list of functions of propaganda beyond persuasion includes:

• mobilisation and coordination (encourage and recruit people to your cause who already agree with you, and/or let’s them find each other)

• signalling permission (shows the government won’t punish certain actions which might normally be illegal)

• crowding out dissent (degrades the public conversation, changes the topic, or absorbs everyone’s energy in discussions which are ultimately distracting).

• sowing distrust, confusion, animosity (hinders coordinated action against the propagandist).

• helps make dissent less visible (e.g. by signalling what will be punished)

• loyalty test or status display (some propaganda is so patently untrue that repeating it is a strong sign that you are willing to humiliate yourself in service to those in power).

All this is more than enough that propaganda should worry us, even if we’re convinced that we don’t need to worry about it persuading people particularly.

This study - Propaganda and conflict: Evidence from the Rwandan genocide - came out the year before study of radio propaganda in Nazi Germany, in the same journal, and uses a similar method. The author uses a detailed map of radio signal strength in Rwanda to test the degree to which radio propaganda was associated with ethnic violence during the 1994 genocide. The results are pretty stark:

The results show that the broadcasts had a significant effect on participation in killings by both militia groups and ordinary civilians. An estimated 51,000 perpetrators, or approximately 10% of the overall violence, can be attributed to the station.

There’s no control for the extent of anti-Tutsi sentiment beforehand, so the study simply doesn’t speak to the extent to which those engaged in slaughter were persuaded by radio propaganda, or merely mobilised by it. For the victims, the distinction hardly matters.

References

Mercier, H. (2017). How gullible are we? A review of the evidence from psychology and social science. Review of General Psychology, 21(2), 103-122.

Kershaw, I. (1983a). How effective was Nazi propaganda. In D. Welch (Ed.), Nazi propaganda: The power and the limitations (pp. 180 –205). London, UK: Croom Helm.

Adena, M., Enikolopov, R., Petrova, M., Santarosa, V., & Zhuravskaya, E. (2015). Radio and the Rise of the Nazis in Prewar Germany. The Quarterly Journal of Economics, 130(4), 1885-1939. https://doi.org/10.1093/qje/qjv030

Yanagizawa-Drott, D. (2014). Propaganda and conflict: Evidence from the Rwandan genocide. The Quarterly Journal of Economics, 129(4), 1947-1994. https://doi.org/10.1093/qje/qju020

Catch-up service

RP#51: How persuasive is AI-generated propaganda? Bullet-time and bullet-point review of a new research report

RP#49: Complex problems and the rush to judgement Smart answers, vs smart processes for finding answers

RP#48: How do you fact-check a deep story? Disinformation campaigns as cooperative work across different groups of actors

PAPER: AI model GPT-3 (dis)informs us better than humans

I saw this paper - AI model GPT-3 (dis) informs us better than humans - mentioned in relation to the last edition of RP (“How persuasive is AI-generated propaganda?”).

From the abstract:

We evaluate whether recruited individuals can distinguish disinformation from accurate information, structured in the form of tweets, and determine whether a tweet is organic or synthetic, i.e., whether it has been written by a Twitter user or by the AI model GPT-3. The results of our preregistered study, including 697 participants, show that GPT-3 is a double-edge sword: In comparison with humans, it can produce accurate information that is easier to understand, but it can also produce more compelling disinformation. We also show that humans cannot distinguish between tweets generated by GPT-3 and written by real Twitter users.

That AI can pass as human here doesn’t worry me too much. On twitter lots of humans behave caricatures of themselves. Decades ago, the sexbot Jenny18 passed the Turing Test in the wild. Creating AI tweets which are the same quality as human tweets may be harder, but I’d bet it isn’t that harder.

The actual methods are quite hard to follow, but from what I can tell (“We measured how accurately participants recognized whether a tweet was containing disinformation or accurate information (disinformation recognition score, range 0 to 1”) participants found it easier to spot of a fake tweet contained disinformation, or if it was accurate, compared to human generated tweets.

You could read this as showing that human tweets are harder to categorise than AI generated tweets - which is the opposite of worrying, since it suggests that the best disinformation will still be written by humans (currently).

As far as I could tell the study did not measure how humans felt about the (dis)information, nor any measure of persuasion or behaviour change, so the abstract line about AI producing “compelling disinformation” is simply overclaiming.

There are a number of other things about the study data presentation which confuse me (Figure 3??), so I’m motivated not to read too much into their results.. They share their data and analysis code, so if I really wanted to know what I thought I’d reanalyse their data and produce discrimination metrics showing how participants balance spotting disinformation vs accurate tweets (like this).

Overall conclusion: no evidence AI is persuasive, no additional reason to worry more about AI generated content than you probably had before.

Spitale, G., Biller-Andorno, N., & Germani, F. (2023). AI model GPT-3 (dis) informs us better than humans. Science Advances, 9(26), eadh1850. DOI: 10.1126/sciadv.adh1850

Open Data and Code: https://osf.io/9ntgf/

PLATFORM: BBC Verify

New technology to show why images and video are genuine launches on BBC News

'Content credentials' will show how BBC journalists have verified content's authenticity.

This is the work done day in day out by BBC News teams, but this is the first time it’s been explained to audiences in this way.

This approach helps us counter disinformation, AI-generated deepfakes and other forms of fake or manipulated content. In the future it will also help audiences around the world differentiate between real and fake BBC content when they see what might appear to be BBC stories on external sites.

This is what an example looks like, a video from the story : Haiti violence: Haiti gangs demand PM resign after mass jailbreak (from 5th March 2024).

If you click on the blue “How we verified this” button, it expands to show:

Thanks to Laurence Dierickx for the example

Story link: https://www.bbc.co.uk/mediacentre/2024/content-credentials-bbc-verify

And finally…

Credit: Dave Coverly / speedbump.com

Comments? Feedback? Persuasive falsities? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

END

1. The uncomfortable corollary of this is recognise that the horrors of the holocaust were not just some dark spell cast by the Nazis, and to portray them as such is an exculpatory narrative. This EconTalk episode with Haviv Rettig Gur discusses the long European history of antisemitism, and notes that the success of the Nazi’s antisemitic policies crucially depended on support from local populations.

2. And gives us an evidence-informed update to the old saw of “I know 50% of my advertising budget is wasted, I just don’t know which 50%”…So now it should be “I know 80% my advertising budget is wasted etc”