Issue-Level Variation in AI Persuasion Effects (Studies 2-3)

This report provides a comprehensive breakdown of treatment effects across individual political issues in Studies 2 and 3 of Hackenberg et al. (2025). Where the original paper reports only overall and model-level ATEs, we estimate issue-specific effects for 598 individual policy stances with sufficient sample size (N >= 20 treated, >= 3 control per issue, pooling across Studies 2-3).

Full data: data_and_analysis_code/output/heterogeneity/issue_level_ates.csv


1. Overall Distribution of Issue-Level ATEs

Statistic Value
N issues analyzed 598
Mean ATE 7.44 pp
Median ATE 7.34 pp
SD of ATEs 6.17 pp
Range [-11.7, +37.6] pp
IQR [3.7, 10.6] pp
Issues with negative ATE 50 (8.4%)
Issues with ATE > 15pp 54 (9.0%)

The SD of issue-level ATEs (6.17pp) is nearly as large as the mean (7.44pp). Issue-level effects span nearly 50 percentage points, from -11.7pp (the AI conversation moved people away from the assigned stance) to +37.6pp. Roughly 1 in 12 issues shows a negative ATE — the intervention backfired on average. Another 1 in 11 shows effects above 15pp — roughly double the overall ATE.


2. Issue-Level ATEs by Issue Area

Issue Area N issues Mean ATE SD of ATEs Min Max Mean backlash rate
Technology and Digital 34 8.60 7.75 -10.8 34.7 19.2%
Taxes and Gov’t Spending 51 7.88 7.60 -11.7 37.6 21.5%
Immigration 26 8.45 7.36 -1.7 32.3 22.5%
Energy and Utilities 41 8.37 6.82 -9.6 24.0 18.0%
Democratic Institutions 37 7.90 6.70 -8.0 19.4 20.2%
Economy and Jobs 41 7.71 6.44 -7.2 23.2 17.7%
Housing and Planning 27 8.32 6.15 -11.3 23.6 19.0%
Civil Rights 44 6.64 6.09 -9.9 26.6 19.9%
Climate Change 57 8.06 5.73 -6.9 33.1 19.1%
Criminal Justice 40 6.39 5.43 -4.4 20.0 19.9%
Foreign Policy 37 7.50 5.40 -3.7 24.8 19.8%
Healthcare 52 5.83 5.29 -8.5 22.3 19.1%
National Security 35 7.21 5.21 -4.2 17.7 21.7%
Transport 30 7.28 5.15 -3.6 17.2 19.5%
Education 46 6.52 5.10 -9.2 16.0 18.3%

Sorted by within-area SD of ATEs. Notable patterns: - Technology, Taxes, and Immigration show the most between-issue variation (SD = 7.4-7.8pp). These domains contain a mix of relatively straightforward policy proposals and deeply contentious values questions. - Education, Transport, and Healthcare show the least between-issue variation (SD = 5.1-5.3pp). These are domains where most stances involve relatively standard policy trade-offs. - Immigration has the highest average backlash rate (22.5%), despite having one of the higher mean ATEs (8.45pp) — consistent with high variance in individual response. - Healthcare has the lowest mean ATE (5.83pp). Health stances may be anchored in more entrenched personal experiences and values.


3. Issue-Level ATEs by Stance Partisanship

Partisanship N issues Mean ATE SD % negative % > 15pp Mean backlash rate Mean pre-attitude
Conservative 186 8.24 6.62 8.1% 11.3% 21.5% 55.8
Labour 294 6.78 5.77 9.2% 7.1% 18.7% 63.2
Neutral/Bipartisan 118 7.83 6.24 6.8% 10.2% 19.0% 61.3

Conservative stances are 1.5pp more persuadable on average than Labour stances. This is partly mechanical: the average pre-treatment attitude for Labour stances is 63.2 (participants already lean toward the stance), leaving less room for persuasion, while Conservative stances start at 55.8. Conservative stances also have higher backlash rates (21.5% vs 18.7%), consistent with more people actively resisting conservative arguments, resulting in a higher-variance distribution.


4. The 20 Most Persuadable Issues

Rank ATE (pp) N treat N ctrl Area Partisanship Issue Stance (truncated)
1 37.6 50 5 Taxes Conservative Raise the inheritance tax threshold, even if this means reduced government revenue
2 34.7 61 3 Technology Neutral Implement mandatory digital literacy courses in all secondary schools
3 33.1 66 3 Climate Conservative Permit the use of gene editing in agriculture to enhance crop resilience
4 32.3 53 6 Immigration Labour Implement stricter criteria for investor visas
5 26.6 54 5 Civil Rights Labour Implement mandatory gender diversity reporting for all companies
6 24.8 63 3 Foreign Policy Conservative Impose stricter regulations on Chinese technology companies
7 24.2 73 4 Taxes Labour Provide increased subsidies for live music venues
8 24.0 59 6 Energy Conservative Increase competition in the energy market by reducing regulations
9 23.6 46 5 Housing Conservative Streamline the planning approval process to reduce bureaucracy
10 23.2 73 3 Economy Labour Mandate equal pay audits for large companies
11 22.3 56 3 Healthcare Neutral Legalize psychedelic drugs for medical research purposes
12 21.6 54 3 Energy Neutral Invest in expanding nuclear energy capacity
13 20.2 58 3 Civil Rights Labour Implement mandatory gender pay gap reporting for all companies
14 20.0 52 5 Criminal Justice Conservative Increase funding for police forces to enhance community safety
15 19.4 55 3 Democratic Inst. Labour Enhance the powers of devolved governments
16 19.3 57 3 Immigration Conservative Strengthen immigration controls to prioritize skilled workers
17 19.2 59 5 Healthcare Labour Increase funding for mental health services
18 19.1 56 3 Criminal Justice Conservative Increase funding for local police forces
19 19.1 75 9 Housing Conservative Streamline building regulations to encourage faster housing development
20 18.9 56 5 Economy Conservative Implement a cap on public sector pay increases

Qualitative observations: Many of the most persuadable issues involve how-type disagreements — questions about the best means to achieve broadly shared goals (streamlining planning, increasing police funding, expanding nuclear energy). Others involve is-type claims that participants may have been uncertain about (gene editing in agriculture, psychedelic drugs for research, digital literacy). Relatively few are primarily ought-anchored, where disagreement hinges on deep values.


5. The 20 Most Resistant Issues (Negative or Near-Zero ATEs)

Rank ATE (pp) N treat N ctrl Area Partisanship Issue Stance (truncated)
1 -11.7 71 4 Taxes Neutral Maintain the triple lock on state pensions
2 -11.3 54 4 Housing Neutral Implement stricter affordability checks for mortgage applications
3 -10.8 60 3 Technology Neutral Increase funding for local radio stations to preserve regional media
4 -9.9 59 5 Civil Rights Conservative Ensure that faith-based adoption agencies can operate according to their beliefs
5 -9.6 61 3 Energy Conservative Prioritize domestic natural gas production to enhance energy security
6 -9.2 60 3 Education Conservative Maintain university tuition fees at current levels
7 -8.7 60 3 Energy Labour Implement stricter regulations on energy companies to protect consumers
8 -8.5 55 4 Healthcare Conservative Allow more private healthcare providers to operate within the NHS
9 -8.0 55 5 Democratic Inst. Labour Allow alternative forms of identification for voting
10 -7.2 61 3 Taxes Labour Implement policies to ensure equitable pension contributions
11 -7.2 57 4 Economy Labour Enhance funding for education and training programs in underserved regions
12 -6.9 64 6 Climate Labour Impose stricter regulations on outdoor advertising to reduce visual pollution
13 -6.5 69 6 Economy Conservative Maintain the flexibility of zero-hours contracts
14 -4.9 59 6 Democratic Inst. Conservative Maintain the British Monarchy as a constitutional institution
15 -4.4 60 5 Criminal Justice Labour Implement stricter penalties for homophobic hate crimes
16 -4.2 73 3 National Security Conservative Increase defence budget to ensure national security
17 -4.0 56 4 Education Labour Implement policies to increase social mobility through education
18 -3.7 47 5 Foreign Policy Labour Prioritize human rights considerations in trade negotiations
19 -3.6 44 3 Transport Labour Subsidize public transport fares to make them more affordable
20 -2.9 57 3 National Security Labour Prioritize diversity and inclusion in military recruitment

On these issues, the AI conversation moved people away from the assigned stance on average. Qualitative patterns: Many involve deep ought-type commitments (faith-based adoption, the Monarchy, zero-hours contracts, defence spending, the NHS) or issues where people likely have strong prior personal experience (pensions, mortgages, tuition). The AI’s information-heavy approach may have triggered reactance on these value-laden topics, or participants may have found the arguments unpersuasive precisely because their attitudes are anchored in normative commitments rather than factual uncertainty.


6. Issues with the Highest Within-Issue Heterogeneity

The issues producing the most variable responses among treated participants (highest SD of attitude change):

Issue ATE SD (treated) Backlash rate N treat N ctrl
Not require voter ID for elections 18.8 34.8 30.0% 60 4
Raise inheritance tax threshold 37.6 33.8 22.0% 50 5
Incentivize private healthcare providers 9.3 25.9 30.6% 62 6
Cap on public sector pay increases 18.9 25.7 32.1% 56 5
Privatize more aspects of rail infrastructure 9.3 25.6 29.1% 55 3
Decriminalize the selling of sex 16.2 25.3 19.7% 66 10
Allow market forces to determine energy prices 14.8 25.3 17.9% 67 4
Stricter criteria for investor visas 32.3 25.1 17.0% 53 6
Regulate zero-hours contracts 0.9 24.9 13.5% 74 3

These issues combine substantial average effects with enormous individual variation. Voter ID (SD = 34.8, backlash = 30%) is the standout: the conversation powerfully persuades some people but produces strong resistance in others. This is textbook IOH territory — voter ID involves factual claims (about fraud), values commitments (about access vs. integrity), and strategic judgments (about the efficacy and legitimacy of the policy means). People will vary in which component anchors their position.


7. Issues with the Highest Backlash Rates

Issue ATE Backlash rate N treat N ctrl
Maintain current income tax thresholds -0.3 38.0% 50 3
Privatise more public services to increase efficiency 6.7 37.9% 66 5
Ensure private healthcare services are more accessible 0.7 37.7% 53 6
Protect green belt areas from development -1.9 37.5% 48 6
Limit immigration to reduce pressure on public services 0.6 37.5% 64 6
Limit expansion of clean air zones to protect small business 7.1 37.5% 56 4
Maintain the British Monarchy -4.9 37.3% 59 6
Maintain economic sanctions against Russia 7.1 37.0% 54 4
Prioritize funding for STEM subjects in schools 10.8 36.2% 47 4
Prioritize access to public services for citizens over recent arrivals -1.7 35.7% 70 5

On these issues, more than a third of treated participants moved against the assigned stance. Several are quintessentially ought-anchored: the Monarchy, immigration limits, privatization of public services, protecting green belt land. The AI conversation appears to activate resistance in a large minority, even where it persuades others — producing the mixed-effect pattern IOH predicts.

Note the STEM funding issue: it has a positive mean ATE (10.8pp) but still 36.2% backlash. This is a striking illustration of how the ATE can badly misrepresent the individual experience. The “average” person was persuaded by 11 points, but more than a third of individuals moved in the opposite direction.


8. Correlates of Issue-Level Persuadability

Correlation r N issues
Mean pre-treatment attitude vs. ATE -0.248 598
Within-issue SD vs. ATE +0.339 598
Backlash rate vs. ATE -0.208 598
  • Pre-treatment attitude is negatively correlated with ATE (r = -0.25): issues where participants already lean toward the assigned stance are harder to move (ceiling effect).
  • Within-issue SD is positively correlated with ATE (r = +0.34): the most persuadable issues are also the most variable in their effects. This is not what you’d expect under a uniform-effect model — if an issue were “easy to persuade on” for everyone, you’d expect high mean and low variance. Instead, high-ATE issues tend to produce a wide spread, suggesting the high ATE comes from a subset of people being moved a lot, not everyone being moved a little.
  • Backlash rate is negatively correlated with ATE (r = -0.21): more backlash means lower average persuasion, as expected. But backlash rates remain substantial (15-38%) even for high-ATE issues.

9. Between-Issue vs. Within-Issue Variation

A crucial comparison for the IOH argument:

Level of analysis SD of effects
Between issue areas (15 groups) 0.83 pp
Between individual issues (598 issues) 6.17 pp
Within individual issues (individual-level) ~16-17 pp (median within-issue SD)

The ratio matters: within-issue individual variation (~16pp) is roughly 2.5x the between-issue variation (6.2pp) and 20x the between-area variation (0.8pp). The specific issue someone discussed matters — but who they are matters far more. This is the gap that IOH configurations are predicted to fill.


Summary

This issue-level analysis reveals several patterns that bear on the IOH framework:

  1. Issue-level ATEs vary enormously (-11.7 to +37.6pp), but this variation is dwarfed by individual-level variation within issues. The issue matters; the person matters more.

  2. 8.4% of issues show negative average treatment effects — the AI conversation backfired. These tend to be value-laden topics (the Monarchy, immigration, faith-based adoption, NHS privatization) where informational persuasion plausibly misses the load-bearing component of most people’s attitudes.

  3. The most persuadable issues tend to involve how-type disagreements (policy means, implementation details) or is-type uncertainty (empirical questions people hadn’t thought deeply about). These are precisely the types of attitudes that informational AI conversations should be well-suited to shift.

  4. High-ATE issues are also high-variance issues (r = +0.34), meaning that the “most persuadable” issues aren’t uniformly easy — they produce both the largest positive shifts and substantial backlash. This is consistent with different people holding the same position for different IOH reasons.

  5. Backlash rates of 15-38% persist across virtually all issues, reinforcing that persuasion interventions systematically produce opposing effects in different subpopulations — the core heterogeneity finding.