When half the population don't trust
Reasonable People #77: A Brexit-era survey of who voters trust and what it means for our information environment
I’m haunted by this opinion poll from nearly ten years ago.
You’ll remember 2016. The year establishment liberals in the US and UK woke up to the depth of resentment felt against them in the voting population1
The poll was from YouGov (a very respected UK polling outfit), and here it is as presented by then Head of Political at YouGov Joe Twyman on Twitter:
A representative sample of UK voters was asked: “Thinking about the EU referendum, how much do you trust what the following types of people say whether we should leave or remain in the European Union?”, and the results are split by those intending to vote Remain in the upcoming referendum, and those intending to vote Leave. The Referendum was held on 23rd of June 2016, so this is approx 10 days before voting.
“Net trust” or “Net approval” is a commonly reported summary from political surveys. Some individuals trust (or approve) and some distrust (or disapprove). The “net” statistic is made up of the percentage of the first subtracted from the other. This means that it moves between a theoretical minimum of -100 (everyone distrusts this thing) and +100 (everyone trusts this thing), but for all intermediate variables it disguises the underlying strength of feeling. So, for net trust of +0 it could be that everyone is on the fence, or it could be that exactly half of the people asked strongly trust and exactly half strongly distrust (what we could call polarisation).
The result is a presentation that takes individual level responses but constructs a population level description. If net score changes it means some subset of the population is flipping from one side to the other. I’ve seen these polls for years but never fully articulated this to myself, which is why I labour over this here.
You can see the full results from the YouGov pages, but, bearing in mind what the Net trust score hides as well as what it shows, here’s my simplified view in graph form:
Net trust goes from negative to positive on the horizontal x-axis, the groups asked about are stacked on the vertical y-axis, with scores for Remain voters shown in blue, Leave voters in red.
This is why I’m haunted by the poll results.
Remain voters trust some groups to tell them about the EU (Academics, Businesspeople, etc), and distrust others (Journalists, Actors). Fair enough. Leave voters, however, distrust everyone. There isn’t a single group they trust to talk to them about the EU. Academics are distrusted least, and foreign politicians most, which makes a kind of sense. Some figures they agree with the Leavers on (not trusting actors and sports people to tell them about the EU). The biggest discrepancies between Leave and Remain voters are International Organisations and Economists.
Still, the overall outcome is that no group that YouGov thought to ask about were trusted by Leave voters.
* *
Trust is the soil in which informed beliefs grow. You can’t verify information about any remote or complex topics entirely by yourself. At some point you have to make a choice about which information to believe, or to rely on someone else’s judgement on some aspect of the topic.
Trusting someone isn’t just a shortcut - it is a necessary delegation of cognitive labour. Just as trade is more efficient under conditions of trust, so is collective intelligence.
The only time there are no gains from trusting is when there is a real lack of trustworthiness - when there is nobody who it is justified in trusting. When these circumstance come about things have got pretty dire.
If Leave voters - as a group - tended not to trust any other group on the issue of the EU, it would explain why they were hard to reach (from the perspective of us establishment liberals), and why, in turn, the result came as a shock.
The Leave/Remain issue may be particular to British politics ten years ago, but that there is a segment of the population who can find nobody to trust remains as relevant as ever.
The difference in trust also speaks to how two halves of a population can, via their beliefs about who to trust, come to live in a distinct information environments.
* *
The same YouGov poll also asked about specific individuals, which, to my mind shows at least two things.
First, it shows that polling is more fun than academic research. With polling you can just throw in some specific names and you don’t need to provide a lengthy theoretical or methodological justification. Looking at the list of names I get the sense YouGov were just vibing, making a list of people without much of a formal plan (or a single woman). Here’s the list, and the results broken down as before:
These are the same respondents as for the first graph. Here the picture is more mixed - neither Remain nor Leave voters have much trust in the individuals asked about. Leave voters have net trust above zero for four people (Johnson, Lewis, Dyson and Farage); Remain votes for just two (Lewis and Corbyn, then leader of the Labour opposition).
The second thing the questions about individuals show is how reported attitudes to generalities may not map to specifics. Leave and Remain voters both placed “UK politicians” as among the lowest trusted group, but asked about specific individuals they both found a named UK politician as one of the individuals they most trusted, from those asked about, to tell them about the EU. Both groups agreed that they don’t trust newspaper journalists to tell them about the EU, but both identified a journalist (admittedly not a newspaper journalist) as their most trusted individual : Martin Lewis, MoneySavingExpert.com. You can see how he pitched his communication on the referendum in this 5 June 2016 blog post.
The result - where general sentiment isn’t mirrored in specifics - fits a pattern we observed in a recent study of online political adverts. If you ask people a general question about political advertising online, the typical response is skeptical - many people say political adverts are bad, often endorsing a view that they are unacceptable and should be banned. For our study, we showed participants - UK voters - the actual adverts used in the 2019 General Election, and - like this YouGov poll - the specifics tell a different story from the generalities. Faced with specific political adverts people overwhelmingly said the vast majority were acceptable. (thread about this study here).
The implication for the Trust poll is that we can’t conclude the Leave voters didn’t trust anyone, just that they don’t trust established groups and organisations. Michael Gove was mocked and misquoted at the time for saying that people have “had enough of experts”. The full quote is more reasonable, and - in the light of this poll - basically correct - people had “had enough of experts from organisations with acronyms”… “Organisations which are distant, unaccountable, elitists”
* *
A lack of trust creates an information void. If Leave voters didn’t trust organisations, most journalists or most politicians, where did they get their information?
The answer is no doubt complex, but my hint would be to follow what we know about trust. We trust those who we believe have our values at heart. Another study I was involved in asked people about who they trusted to tell them about pollution risks in their home. The results showed that many people trusted experts, like independent scientists, but everyone trusted their friends and family. Digging into these attitudes a bit, we saw that people’s perception of expertise wasn’t distorted. In general, they believed the scientists were experts (even if they didn’t trust them), and knew their family and friends weren’t experts (even though they did trust them). The predictor of trust was the belief that an adviser had their interests at heart. Those that didn’t trust the scientists thought the scientists would be too cautious in estimating pollution risk, or wouldn’t be transparent about what they knew. Everyone trusted friends and family because everyone knew that their friends and family where on their side and would be concerned about potential risks to them from pollution (study link, my summary here).
Now, even more than in 2016, we get views from friends and family via social media. The apps and algorithms which populate our feed mix updates from offline friends with those friends who we only know online, and updates from those we feel an ‘asymmetrical intimacy’ with, celebrities and Influencers who are part of our daily lives but have know knowledge of us personally in return. Social media launders information - good and bad - through the social graph. It supercharges the dynamic of an urban myth, where a story is always told by someone you know, and presented as traceable back to an immediate contact (“This thing happened to my friend’s cousin…”). Information arrives with a personal connection, via our social network, or via the proxy social network we’ve co-curated with the algorithm.
All of this is to say that the new dynamics of our information environments are confusing, and intimately tied up with - and reflected by - who we elect to trust. The votes of 2016 augured in new political eras in the US and UK. Polls like this one could have told us how things had shifted even before the votes occurred.
I’m writing more for Reasonable People while on my career break. Upgrading to a paid subscription will encourage this
Keep reading for more on what I’ve been reading on deliberation, AI and fact-checking.
PAPER: Democratic forecast: Small groups predict the future better than individuals and crowds.
Dezecache et al (2022) report a replication of the result I wrote about previously in : How to outsmart a crowd of 5000 people in 4 minutes.
Abstract
Predictions pose unique problems. Experts regularly get them wrong, and collective solutions (such as prediction markets and super-forecaster schemes) do better but remain selective and costly. Contrary to the idea that face-to-face discussion hinders collective intelligence, social deliberation improves the resolution of general knowledge problems, with four consensually agreed answers outperforming the aggregate knowledge of 5,000 nondeliberating individuals. Could discussion help predict the future in an efficient, cheap, and inclusive way? We show that smaller groups of lay individuals, when organized, come up with better predictions than those they provide alone. Deliberation and consensus made individual predictions significantly more accurate. Aggregating as few as two consensual predictions did better than classical “wisdom of crowds” aggregation of 100 individual ones. Against the view that discussion can impair decision-making, our results demonstrate that collective intelligence of small groups and consensus-seeking improves accuracy about yet unknown facts, opening the avenue for efficient, inclusive, and inexpensive group forecasting solutions.
Dezecache, G., Dockendorff, M., Ferreiro, D. N., Deroy, O., & Bahrami, B. (2022). Democratic forecast: Small groups predict the future better than individuals and crowds.Journal of Experimental Psychology: Applied, 28(3), 525–537. https://doi.org/10.1037/xap0000424
related
Pescetelli, N., Rutherford, A., & Rahwan, I. (2021). Modularity and composite diversity affect the collective gathering of information online. Nature Communications, 12(1), 3195. https://doi.org/10.1038/s41467-021-23424-1
AI Update: Why language models hallucinate
A new report from OpenAI was reported as admitting that hallucinations are inevitable (which those in the know knew already), or that the cure for hallucinations was to train models to say “I don’t know” (it might reduce the number of hallucinations, but it wouldn’t remove them completely).
There’s a good explainer from Wei Xing in the Conversation: Why OpenAI’s solution to AI hallucinations would kill ChatGPT tomorrow. Xing makes the convincing point that we consumers have got used to language models having the answer to everything. We simply don’t want them to say “I don’t know”, and the perceived commercial benefits would evaporate if they did this too much.
“In short, the OpenAI paper inadvertently highlights an uncomfortable truth: the business incentives driving consumer AI development remain fundamentally misaligned with reducing hallucinations. Until these incentives change, hallucinations will persist.”
OpenAI: Why language models hallucinate (5 September 2025)
Computerworld.com: OpenAI admits AI hallucinations are mathematically inevitable, not just engineering flaws (18 September 2025)
Indicator: Briefing: Finally, some data about Meta’s Community Notes program
A little slow, a little late. And a lackadaisical engagement with the transparency that Twitter/X have made such a feature of their deployment of the system
Watch this space for some future work on Community Notes that I’m really excited about.
Link: Finally, some data about Meta’s Community Notes program
Catch-up
The costs of being open-minded
The intelligence is in the user
END
Comments? Feedback? Ingredients for Martin Lewis’s secret sauce for bipartisan trust? I am
tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online
Don’t think for a moment that I don’t recognise myself as an establishment liberal. I’ll be first against the wall regardless of who leads The Revolution.