Cross-country survey experiment · MultiPoD
What gets lost when politics speaks a second language?
For MultiPoD, a Horizon Europe project, we tested in seven countries whether reading politics in a second language costs citizens understanding, and whether offering a choice of language helps.
MultiPoD deliverable D1.1, revised 2 October 2026 · Ethos Lab (MultiPoD Work Package 1)
Chapter 1
A question for a multilingual Europe
MultiPoD, a Horizon Europe project, is building an online space where citizens from across Europe can deliberate on policy across language borders. Before designing how the platform handles language, the consortium needed evidence, and Ethos Lab led the study.
Whether citizens take in politics as well in a second language as in their own was an open question.
A debate you cannot follow is not a debate you can join.
The easy default is to publish in English and assume people cope. We wanted to measure what that default costs, so we built an experiment into a survey.
Chapter 2
A fair test in seven countries
We surveyed 8,195 adults through online panels in Austria, Belgium, France, Greece, Portugal, Spain and the United Kingdom. People who speak at least one language beyond their first read the same short, factual article on the EU AI Act, in one of three ways: in their first language, in their second language with no choice, or in a language of their choosing. The experiment uses the 4,630 of them who could be given any of the three versions.
A random draw decided which version each person read: not the respondent, and not us. That makes the three groups alike, on average, in every way except how they met the article. So if one group understands less than chance alone would explain, the way they read is the reason.
Six factual questions followed straight away. We also asked how difficult the text felt and recorded how long each person spent on the page: the first shows what people understood, the other two what it cost them to get there.
Comparison group
First language
The article in their first language
Group 2
Forced second language
The same article in their second language, with no choice
Group 3
Choice of language
The same article, in a language they chose
What we measured, one for each question
- Correct answers to six factual questions (main outcome)
- How difficult the text felt (1 to 5) and time spent reading
- The same measures, for the group offered a choice of language
Every specification on the page
This study was not pre-registered, and we say so. The hypotheses came from earlier research, and we report every specification we estimated: unadjusted, with demographics and country fixed effects, and with full controls. All three appear in chapter 3.
Then we estimated the effects three ways, and kept every one.
Chapter 3
Three questions, three answers
Each question had its own measure. Here is what each one showed.
The question
Do people understand less when they read politics in a second language?
The answer
Yes. About half a correct answer less.
Reading in a second language, with no choice, cost 0.51 correct answers out of six: about 14% of the first-language group's average of 3.7. Accounting for differences between respondents barely changes this.
−0.51
correct answers out of 6, forced second language
The question
Does it cost them more effort?
The answer
Yes. More effort, less understanding.
The same group found the text harder, by 0.25 points on a 1 to 5 scale, and spent about 13% longer reading it.
+0.25
difficulty, points on a 1 to 5 scale
The question
Does offering a choice of language help?
The answer
Yes. Choice avoided the penalty.
Outside the UK, more than nine in ten people offered a choice picked their national language. They scored slightly higher than the first-language group and found the text no harder, though reading took them a little longer.
+0.15
correct answers out of 6, choice of language
Every estimate, every specification
- Unadjusted: −0.52 answers, 95% confidence interval −0.66 to −0.39.
- Demographics and country fixed effects: −0.52 answers, 95% confidence interval −0.65 to −0.39.
- Full controls: −0.51 answers, 95% confidence interval −0.64 to −0.39.
- Unadjusted: 0.12 answers, 95% confidence interval −0.01 to 0.24, which includes zero.
- Demographics and country fixed effects: 0.13 answers, 95% confidence interval 0.01 to 0.25.
- Full controls: 0.15 answers, 95% confidence interval 0.03 to 0.27.
Intention-to-treat effects of random assignment, pooled across the seven countries and estimated by OLS with HC1 robust standard errors, on the 4,630 multilingual respondents who could be assigned to any of the three versions (4,623 with full controls). Demographics are age, gender and education, with country fixed effects; full controls add whether the respondent speaks English and political interest. Control mean 3.71 correct answers. The gain under choice is statistically significant only once covariates are included. Across all 4,861 multilingual respondents the forced-language estimates are −0.49 to −0.52. The study was not pre-registered, so every specification estimated is shown.
Show the numbers as a table
| Outcome | Estimate (answers) | 95% CI | Control mean | n |
|---|---|---|---|---|
| Forced second language: Unadjusted | −0.52 | −0.66 to −0.39 | 3.71 | 4,630 |
| Forced second language: Demographics and country fixed effects | −0.52 | −0.65 to −0.39 | 3.71 | 4,630 |
| Forced second language: Full controls | −0.51 | −0.64 to −0.39 | 3.71 | 4,623 |
| Choice of language: Unadjusted | 0.12 | −0.01 to 0.24 | 3.71 | 4,630 |
| Choice of language: Demographics and country fixed effects | 0.13 | 0.01 to 0.25 | 3.71 | 4,630 |
| Choice of language: Full controls | 0.15 | 0.03 to 0.27 | 3.71 | 4,623 |
Effort, measured two ways
- Forced second language: 0.25 points, 95% confidence interval 0.18 to 0.33.
- Choice of language: 0.00 points, 95% confidence interval −0.08 to 0.07, which includes zero.
- Forced second language: 0.121 log points, 95% confidence interval 0.035 to 0.207.
- Choice of language: 0.092 log points, 95% confidence interval 0.010 to 0.174.
By country, the loss in understanding is clear in Austria, Belgium, France, Greece and the United Kingdom, and not clearly different from zero in Portugal and Spain. Each country sample is much smaller than the full sample, so the country results are only indicative.
Before acting on the answers, it matters what they leave open.
Chapter 4
What this cannot tell us
Online panel respondents tend to be younger, more digitally literate and more politically attentive than the general population, so effects in the wider population could be larger.
Most multilingual respondents speak English alongside their national language. Speakers of regional, minority and heritage languages, such as Catalan, Galician or Basque in Spain, could not be given the second-language version because no article existed in their language, so the study does not speak to them.
The test covers one short, factual article. Longer or more technical texts could behave differently.
Slightly fewer people finished the survey in the second-language group. Allowing for that leaves the penalty clearly in place.
An answer matters only if someone designs with it.
Chapter 5
From findings to recommendations
The three findings point one way: reading politics in a second language costs understanding and takes more effort, and a simple choice of language removes that cost. The study therefore recommends that the platform:
- 1Defaults to each user's preferred language, not to English
- 2Makes language choice simple, visible and reversible
- 3Labels translated or summarised content clearly
- 4Monitors comprehension and effort by language, so no group carries a heavier burden
These are recommendations to the consortium. How the platform takes them up is for the project to decide.
The paper
MultiPoD deliverable D1.1, revised 2 October 2026
Ethos Lab (MultiPoD Work Package 1)
MultiPoD is funded by the European Union's Horizon Europe programme.
Views and opinions expressed are those of the authors only and do not necessarily reflect those of the European Union.
Working on a similar question?
Tell us what you need to know and by when. We will reply within two working days with an honest view of what we could find out.