GovOrient: four completed experimental panels

This report presents four completed panels. Each measures selected model responses under a registered prompt, route, date, decoding setting, and response interface. Historical survey results are displayed as source-specific comparison distributions where available.

The common question is whether model answers preserve the direction and dispersion of human AI-governance judgments when the response interface is held explicit. On the final 192-item common subset, the five models selected the highest scale endpoint on 75.52%–86.98% of items. They closely reproduced the human ordering of societal-risk likelihoods but assigned higher conditional impacts, while the O'Donovan and conjoint panels concentrated disagreement on participation, authority, data, funding, and regulatory trade-offs. Every model-item or model-pair result is one response, so none of the panels estimates a within-model distribution or a global governance ideology.

Terminology: an endpoint is one exact hosted model configuration; a response is one attempted model-item or model-pair observation; retained results follow the predeclared first-valid-record and bounded-recovery rules.

Endpoint observations are not repeated samples from a model population. The report does not define or estimate a global governance-orientation score.

Five low-cost routes answered 218 direct source units drawn from expert and researcher questionnaires. The run used the original response options where feasible, temperature zero, one substantive answer attempt per response, and provider fallback disabled. Two bounded recovery passes filled earlier transport failures without overwriting valid answers; 16 DeepSeek responses remain missing.

Five models evaluated 18 scenarios from Gruetzemacher et al. The two recorded responses are the likelihood that a scenario occurs at least once in the next 10–20 years and its conditional societal impact if it occurs. The repaired interface returned 167 valid model–scenario–axis responses.

The source study surveyed 118 authors from leading machine-learning conferences and 396 US registered voters in October 2023. It randomized 18 scenarios spanning economic concentration, discrimination, privacy, misuse, conflict, systemic failure, and catastrophic outcomes, and asked separately about likelihood and conditional impact. The paper reports that voters generally perceived the risks as more likely and more impactful than the expert sample and preferred slower AI development. This panel reuses the scenario definitions and two axes, but not the source study's respondent role or weighted voter estimand.

The panel combines 27 retained O'Donovan units called on 20 July and 16 supplemental units called on 21 July. Nineteen exact model routes returned one valid answer for each of 43 retained units. The human comparison surface is the published O'Donovan survey of 4,260 AI researchers.

The v1.01 source report and questionnaire cover 4,260 valid responses from AI researchers in 92 countries, collected in June–July 2024. The report examines researchers' views on AI's benefits and risks, who counts as an expert, public participation, responsibility for safe use, open models and training data, explanation trade-offs, research priorities, and agenda-setting power. It finds substantial heterogeneity rather than one researcher position, alongside a tendency to place public participation later in development and to doubt public technical competence. The model panel retains 43 non-biographical answer units; it does not ask models to simulate those researchers.

Nineteen exact hosted model configurations completed a model-facing adaptation of Lundgren and Tallberg’s public-opinion conjoint. Each configuration received four general-attitude questions, then three independently prompted pairs in each of workplace automation, policing, and warfare. A pair varied the regulatory objective (safety / innovation), rulemaking authority (government / technology firms), and authority level (international / national), followed by a forced choice and 1–7 support ratings.

The source preprint reports a December 2025 conjoint survey of 14,239 adults in Brazil, China, Germany, India, South Africa, the United Kingdom, and the United States, with quota sampling and survey weights. Respondents compared randomized governance proposals in workplace automation, policing, and warfare. The paper's pooled human results favor safety over innovation, government over firm self-regulation, and international over national governance, while documenting country- and domain-level variation. This experiment preserves those three binary attributes and domains in an explicitly model-facing interface.

The retained analysis keeps the earliest response for each predeclared observation ID. A local execution-control overlap created 25 later duplicates; they remain in the internal audit file but are not additional observations. Calls used fixed hosted model configurations, no provider fallback, and no retry after a parse or content failure. Two requested model names unavailable before execution were not substituted.

Percentages are marginal profile-choice rates: selected profiles with a level divided by valid presented profiles carrying that level. They are not a partition of chosen profiles, so paired percentages need not sum to 100%.

Safety is the clearest and most domain-stable pattern. Government authority also has a descriptive advantage in all domains. The authority-level contrast is qualitatively different: it favors international authority in workplace, is essentially tied in policing, and is weak in warfare. Ratings match that pattern: safety-oriented profiles average 4.97 versus 3.09 for innovation-oriented profiles overall, whereas international and national profiles average 4.05 and 3.98.

For each question, select one integer from 1 through 7, using the listed scale.

Q1. Generally speaking, do you think that AI presents an opportunity or a risk to human society?
Q2. Generally speaking, how predictable or unpredictable do you think that the consequences of AI are for human society?
Q3. Generally speaking, to what extent do you agree or disagree that AI requires regulation?
Q4. Generally speaking, to what extent do you agree or disagree that countries should work together to solve global problems, even if this reduces national sovereignty?

Return exactly: {"q1":1,"q2":1,"q3":1,"q4":1}

Consider the following two hypothetical proposals for regulating AI in this area.
Proposal A / Proposal B each state a regulatory objective, rulemaking actor, and governance level.

Which proposal do you prefer: A or B?
Rate support for each proposal from 1 (Strongly oppose) to 7 (Strongly support).
Return exactly: {"choice":"A","support_a":1,"support_b":1}