All walkthroughs · Paper briefs

Governing Artificial Intelligence: Public Preferences and Regulatory Options

Magnus Lundgren (University of Gothenburg) and Jonas Tallberg (Stockholm University)
arXiv 2607.14585 [econ.GN], 16 July 2026; also posted on SocArXiv and SSRN

Paper · Brief and rating · Dashboard entry

About 22 minutes · Download MP3

AI-generated walkthrough with synthesised voices: a host explains the paper and a discussant questions it. It covers claims, methods, key assumptions, citations and use so far, decision relevance, and what an evaluator should check. Every direct quotation was checked word for word against its source. Public sources only. Not an Unjournal evaluation or decision.

Transcript

Host
This is a walkthrough of Governing Artificial Intelligence: Public Preferences and Regulatory Options, by Magnus Lundgren of the University of Gothenburg and Jonas Tallberg of Stockholm University. It was posted to arXiv in July twenty twenty-six, with copies on SocArXiv and SSRN. This episode is an AI-generated walkthrough, not an Unjournal evaluation or decision. The paper reports a pre-registered conjoint survey experiment fielded in December twenty twenty-five in seven countries: Brazil, China, Germany, India, South Africa, the United Kingdom and the United States, with just over fourteen thousand respondents in total. The headline is that people in all seven countries support regulating AI, and that when they choose between regulatory proposals they prefer safety to innovation, government authority to industry self-regulation, and international to national rules. The safety preference is by far the largest of the three.
Discussant
Before the method, what is the paper claiming beyond those three preferences? A finding that people like safety is not surprising.
Host
There are three further claims. First, the safety preference is driven by risk perceptions rather than ideology. The authors write: "Evaluations of AI governance are not primarily structured by ideological orientation, but by how individuals perceive the risks, unpredictability, and personal relevance of AI." Second, the patterns hold across three application domains, AI in the workplace, in policing and in warfare, with some variation by domain and country. Third, and this is the claim most likely to travel, they argue that actual regulatory regimes are out of step with these preferences. They say the European Union's approach comes closest to the preferred profile, and write that "The US emphasis on innovation, market-led governance, and national flexibility is least consistent with public preferences, including among US respondents." They describe the politics of AI as shaped by what they call "a triple misalignment": between major powers, between governments and publics, and between short-run incentives to innovate and long-run concern about risk.
Discussant
Walk me through the design. What exactly did a respondent see?
Host
Each respondent completed three thematic blocks, one per domain, in random order. Each block opened with a short preamble on the benefits and risks of AI in that domain, then two questions on whether and how the respondent expects to be personally affected, then three paired comparisons. In each pair, two hypothetical regulatory proposals were shown side by side. Each proposal had three attributes with two levels each. The objective: strengthening people's safety against AI harms, or strengthening technological innovation in AI. The mode: rules developed by government authorities or by technology firms. And the level: national or international. The respondent picked the proposal they preferred, a forced choice, and also rated each one on a seven-point support scale. That is nine pairs per person, giving about two hundred fifty-six thousand evaluated profiles.
Discussant
Three binary attributes means only eight possible profiles. That is a very small conjoint.
Host
Yes, and the authors present it as a deliberate choice to keep the task manageable. In the limitations section they write that "the design necessarily abstracts away from additional complexities to maintain respondent comprehension." There is one more design wrinkle. The survey platform would not show two identical proposals, so within a pair each attribute matched about forty-three percent of the time, rather than half. The marginal distribution of each level stays balanced at fifty percent.
Discussant
What is the sample?
Host
Online panels run by Dynata, with at least two thousand adults per country, quota-stratified on age, gender and education, and then post-stratification weights on those same three variables. Country selection is meant to span region, regime type, income level and position in the AI industry. So the set runs from liberal democracies to China, which the V-Dem regime classification codes as a closed autocracy, and from high-income to lower-middle-income India.
Discussant
And the estimand?
Host
The main quantity is the marginal mean: the average probability that a profile is chosen when it carries a given attribute level, averaging over the other attributes. In the forced-choice task, a marginal mean above one half means that level tends to win. They also report average marginal component effects, the standard conjoint quantity from Hainmueller, Hopkins and Yamamoto, which is the change in choice probability from switching one attribute level while averaging over the rest. With binary attributes the two are closely related. In the pooled linear probability model, the baseline of innovation, government and national has a choice probability of about forty-four percent. Switching to safety raises it by about twelve and a half percentage points. Switching to technology firms lowers it by about three points, and switching to international raises it by about three points. So the safety effect is roughly four times the size of the other two. The rating outcome tells the same story, and the authors note that "even the less preferred options receive ratings above the midpoint of the scale".
Discussant
Before the conjoint, there is a descriptive result on general support for regulation. What does that show?
Host
The direct question is whether AI requires regulation, on a seven-point agree-disagree scale. The pooled weighted mean is about five point three. China is highest at about five point six, then the United Kingdom and the United States; India is lowest at about four point eight but still above the midpoint. A regression of that support on demographics and attitudes, with country fixed effects, finds support rising with age, education and income and with perceived AI risk and AI knowledge. Trust in technology firms, trust in government and left-right ideology show no significant association. That regression is exploratory.
Discussant
Now domains and countries. How uniform is this?
Host
The safety preference is the most uniform part. It appears in every domain and in every country, strongest in the United Kingdom, Germany and the United States, weaker in Brazil and South Africa, more modest in India, and in China it points the same way but is not significant at the five percent level in the weighted analysis. The other two dimensions are more uneven. The preference for government over firms is driven by China, Germany, the United Kingdom and the United States. Brazil shows a slight preference for firm self-regulation, and India and South Africa show none. In the workplace domain, respondents overall show no clear preference between government and firms. On level, five countries prefer international rules, while China prefers national rules and the United States shows no clear preference. The international preference is strongest for military AI.
Discussant
So the pooled preference for government and for international rules is a seven-country average that conceals real disagreement. The two countries that build most of the frontier AI systems are the two that do not want international rules.
Host
That is a fair reading, and the authors say as much, attributing it to national competitive position. The heterogeneity analysis is where the mechanism claim comes from. They split respondents on pre-treatment measures. Those who see AI as a risk rather than an opportunity, those who see its consequences as unpredictable, and those who expect to be negatively affected personally all show a larger safety premium. Internationalists, defined by agreeing that their country should cooperate on global problems even at some cost to sovereignty, show a larger international premium. Ideology does not move the government-versus-firms preference, which contradicts their pre-registered hypothesis eight that the left would favour government regulation more. There is also an exploratory elite comparison, where elite means high education, high income and a professional or managerial job. Elites show a safety premium about eight and a half percentage points smaller than non-elites, and a government premium about eight and a half points larger.
Discussant
The risk-perception result needs care. Perceived risk is measured, not randomized. Respondents who see AI as risky differ in many ways, so this is a correlation between a trait and the size of an experimental effect. It shows who holds the preference, not what causes it.
Host
Right. The attribute effects are causal in the conjoint sense, because the attributes are randomized. The moderation results are descriptive. The paper's language of an underlying mechanism goes a little further than the design supports, though it does say the pattern is consistent with earlier work.
Discussant
What robustness checks are there?
Host
Several, in the supplement. Dropping the survey weights leaves the direction the same and the estimates similar or slightly larger. The weights matter mostly in India and South Africa, where a small number of low-education respondents get large weights. Without weights, the safety preference in China and India becomes significant. The attention check was demanding: respondents had to select exactly radio and TikTok from a list of eleven news sources. About sixty-one percent passed. Restricting to that group gives similar or slightly sharper estimates; the authors conclude that "inattentiveness introduces noise rather than systematic bias." The first-listed proposal was chosen about fifty-six percent of the time, a primacy effect, which washes out because levels are balanced across positions. The two-way attribute interactions are not significant, though in a three-way model safety and government show a small complementarity. On the rating outcome, Chinese respondents rate everything higher, which the authors read as "This likely reflects cross-national differences in response styles rather than substantive differences in preferences", and which is why they lead with the forced choice.
Discussant
There is one robustness result I would flag. The block-order test.
Host
Yes. The safety effect does not depend on whether a block came first, second or third. But the preference for government over firms is significantly stronger when a block is shown second or third, and the international preference is stronger in the third block. So two of the three headline preferences grow as respondents move through the survey. That could be learning, priming from the preambles, or respondents settling into a heuristic.
Discussant
That reinforces the point that the mode and level results are fragile and the safety result is the one that carries the paper. Let me press on what the safety attribute means. Respondents choose between a proposal whose purpose is to strengthen people's safety against AI harms and one whose purpose is to strengthen technological innovation in AI. Nothing in the profile says what is given up. Real regulatory choices trade compliance costs, slower deployment, or lost competitiveness against risk reduction. Here safety comes at no stated price.
Host
The authors are aware of a version of this. They cite the Dafoe, Zhang and Caughey work on information equivalence and say they wrote the attribute labels in parallel form and neutral wording, and that each block preamble presents benefits and risks of comparable length, partly to address the worry that innovation reads as more abstract than safety.
Discussant
Let us look at the preambles, then. What did the workplace block say?
Host
The workplace preamble gives as examples, quoting the instrument, "AI systems replacing human workers in call centers, warehouses, and areas like accounting, translation, and computer programming." It then says some see a path to productivity and growth, while "others are concerned that AI could replace many jobs and disrupt labor markets." The warfare preamble mentions attack drones and missile systems operating with limited or no human oversight, then battlefield advantage on one side and arms races and the laws of war on the other.
Discussant
So the examples themselves are the harms. The benefit side is stated abstractly, as productivity or advantage, while the concrete images are job loss and weapons without human control. An evaluator should ask whether that framing raises the safety premium. It probably does not create it, but its size, twelve and a half points, is the number the paper leans on, and the size is what the framing could inflate.
Host
One indirect check is that the safety effect does not vary with block order, so it is not only a product of accumulating preambles. But every block has a preamble, so the design cannot separate the safety preference from the framing.
Discussant
A second construct question is the mode attribute. Technology firms writing the rules is an extreme version of private governance. It is not industry standards with government backstops, or co-regulation, which is how most real regimes work, including much of the European AI Act's reliance on harmonised standards.
Host
And the authors' own mapping onto real regimes depends on that. In the supplement they compute predicted support for all eight profiles. Safety, government and international ranks first, and they say it aligns with the European AI Act and the OECD AI Principles. Innovation, firms and national ranks last, and they say it approximates the current US approach. The G7 Hiroshima process is described as matching on safety and international scope but relying on voluntary self-regulation.
Discussant
That mapping is the step I would scrutinize most. The US federal approach has not been pure firm self-regulation: there are agency rules, state laws, export controls, and procurement standards. Coding a whole national regime as one of eight profiles is the authors' judgment, not a measurement, and the misalignment claim depends on it. The predicted-support analysis is also listed as exploratory.
Host
And within the United States, the survey shows a safety preference and a government preference, but no clear preference on national versus international. So the US-respondent part of the misalignment claim rests mainly on safety and mode.
Discussant
What about stated preference validity more generally?
Host
The authors cite the Hainmueller, Hangartner and Yamamoto paper validating conjoints against real Swiss naturalisation votes. But these are low-salience, hypothetical policy choices. There is no cost to the respondent and no competing priority, such as growth, jobs or national security against a rival. Whether stated preference for safety translates into votes or political pressure is the open link between this survey and the legitimacy argument. The authors argue that regulation aligned with citizen preferences commands more compliance and trust, but the paper does not test that; it cites literature for it.
Discussant
China and India deserve their own caution. Online panels in India over-represent urban, educated respondents, which is why the weights are large there. In China, a closed autocracy, questions about government authority and international rules may be answered with an eye to the official line. The finding that Chinese respondents most prefer government authority at the national level is exactly what you would expect from either sincere preference or socially desirable answering.
Host
The paper does not address preference falsification directly. It does note that Chinese respondents rate everything higher, and attributes that to response style.
Discussant
Now, how has the paper been used and cited so far?
Host
Very little, which is unsurprising for a paper less than three months old. OpenAlex lists several records for it, on arXiv, SocArXiv and SSRN, and none has a recorded citation. We found one news write-up, in The Cryptonomist on eighteen July twenty twenty-six, which summarizes the findings and the misalignment framing without outside comment. The Altmetric record shows a handful of posts on X and Bluesky. We did not find it in policy documents, government consultations or lab policy statements. A search miss is not proof it has not been read; it means we could not confirm uptake.
Discussant
Is there related work an evaluator should compare against?
Host
Yes. Ehret's twenty twenty-two comparative study in the Journal of European Public Policy is the closest earlier cross-national work, and the authors cite it. Separately, a team at the University of Amsterdam reported a six-country survey in twenty twenty-five, covering Brazil, Denmark, Japan, the Netherlands, South Africa and the United States, asking whether citizens agree with easing AI regulation for innovation. It is a different design, a direct-question survey rather than a conjoint, and it does not cite this paper, but its overlap in countries and question makes it a useful comparison.
Discussant
Who might act on this, and why would an Unjournal evaluation be useful?
Host
The obvious users are people arguing about the legitimacy of different governance models. That includes European officials defending the AI Act's approach, advocates for binding international agreements, and critics of voluntary commitments by AI labs. The authors themselves end with a policy recommendation: "Governments wishing to secure durable public support for AI governance may thus need to rebalance toward safety-oriented regulation", along with public oversight and international coordination. Funders interested in AI governance might cite it as evidence of public demand. Because the topic is politically charged, the paper is the kind of result that gets quoted for its headline without its caveats. An evaluation that says clearly how much weight the twelve-point safety premium and the misalignment claim can bear would help those readers.
Discussant
And it is checkable. It is pre-registered on the Open Science Framework, the instrument is fully reproduced in the supplement, and the estimation is standard. The main obstacle is that the data and code do not appear to be posted yet.
Host
So, to summarize what an evaluator should check. First, the pre-registration against the paper: hypotheses one through eight are pre-registered, and the general-support regression, the elite comparison and the predicted-support profiles are exploratory. Second, the construct validity of the safety attribute: whether a costless safety label and harm-heavy preambles inflate the effect size. Third, the mode attribute as an extreme version of private governance, and how that shapes the mapping to real regimes. Fourth, the block-order results for mode and level. Fifth, sample quality in India and China, including weight concentration and the possibility of socially desirable answers. Sixth, whether the mapping from eight abstract profiles onto the European, US and Chinese regimes supports the word misalignment.
Discussant
I would add one more. The paper's strongest result is the cross-country consistency of the safety preference. Its weakest is the inference from that preference to which real regulatory regimes are legitimate. An evaluator should keep those two separate, and ask the authors to temper the policy conclusion to what an eight-profile, costless, hypothetical choice can show.
Host
That is the walkthrough. The paper is by Lundgren and Tallberg, on arXiv since July twenty twenty-six. Again, this was an AI-generated walkthrough, not an Unjournal evaluation or decision.

Sources