We propose Acceptability-Constrained Climate Policy Design (ACCPD) as a framework for integrating socio-cultural dynamics into policy design from the outset. This approach would: (a) use LLMs as cultural world models to map narratives and fairness concerns, (b) couple these with social dynamics via GABMs, linked to physical world simulators, and (c) monitor the system through continuous validation using real-world signals (refer to Fig. 1). The aim is to locate and ethically expand the Acceptability Frontier, the set of designs achieving climate impact while remaining socially tenable42,43. Importantly, “acceptability” involves not just the general population but also institutions, businesses, media, and particularly marginalized communities. Hence, the term ‘societal acceptance’ or ‘societal response’ encompasses all these actors.
The Observatory Layer (curved line) acts as an enveloping ‘Human-in-the-Loop’ interface where stakeholders define the Draft Policy and monitor the system. Inside the simulation, LLM-Agents assess the policy and diffuse opinions through a Social Network (GABM). These social outcomes interact with a Physical World Model (e.g., climate impacts), creating a feedback loop between social sentiment and physical reality. Finally, the Acceptability Frontier synthesizes these results to guide the Redesign (Optimization) step, where policymakers utilize the output to adjust policy parameters for higher political viability before re-assessment. We refer the readers to Supplementary Information Fig. S1 for the proposed comprehensive system architecture. This figure includes public domain icons from Wikimedia Commons.
The proposed ACCPD framework consists of the following components: A. LLMs, B. GABMs, C. Physical world models, D. Acceptability frontier, and E. Observatory layer. The following defines each of their roles and potential in augmenting climate policy design.
LLMs: Cultural World Modeling
Within the ACCPD framework, LLMs can be conceptualised as “cultural world models”, technological systems that can aggregate patterns from billions of lines of digitized human discourse18. Kozlowski et al.19 demonstrated this empirically: an LLM trained before COVID-19, when conditioned with political identities and exposed to pandemic facts, reproduced the partisan polarization that later emerged in reality. This suggests LLMs can “roll forward” societal processes within specified contexts.
When carefully prompted and calibrated, LLMs could potentially help anticipate how different publics might respond to policies. They can surface whose values are affirmed, what fairness concerns emerge, where identity sensitivities are triggered, and which counter-narratives gain salience. The goal is, however, not to substitute for human voices but to flag potential narrative risks and opportunities for subsequent testing with real-world data or for further in-depth citizen deliberation.
GABMs: From Narratives to Social Cascades
GABMs comprise agents imbued with the capabilities of LLMs, allowing them to interact with each other and their environment through natural language rather than a set of predefined rules. As a result, they allow us to create virtual societies with programmable environments38. This, in turn, enables us to study the emergence of collective communication behavior under various contexts. In our ACCPD framework, GABMs serve to extend socio-cultural policy assessments to social dynamics of policy support or rejection patterns. These models would simulate how initial responses spread through social networks, accounting for: (a) network effects: opinion clustering and spreading along social ties44; (b) threshold dynamics: tipping points where fence-sitters follow early adopters45; (c) counter-mobilization: opposition organizing and spreading competing narratives46; and (d) complex contagion: how societal responses spread differently than simple information47,48,49. While ABMs already provide these capabilities, GABM provides a natural and richer form of communication and behavior, allowing us to represent more complex social situations38. In addition, we can also incorporate power dynamics through network characteristics to account for the role of power actors driving policy acceptance or rejection.
Existing GABM frameworks like Concordia38,50 already show that such infrastructure can be developed, providing a practical pathway for its implementation. Other GABM frameworks focus only on interaction on social media, like OASIS51, capturing verbal interaction within social and algorithmic systems.
We acknowledge that GABM is not a singular solution for all modeling challenges. Traditional mathematical approaches, such as Equation-Based Modelling (EBM) or System Dynamics52, remain superior for simulating aggregate flows or systems with well-defined physical laws, offering transparency and lower computational costs compared to the high-dimensional parameter space of LLM-based agents. However, these methods lack the capacity for semantic processing, the ability to interpret and generate natural language justifications based on distinct cultural identities.
Physical world model
An integrated system that couples social simulation with physical world models, such as climate models, allows us to see the policy acceptance or rejection impacts on the physical world and the impact of a changing physical world on policy acceptance and rejection dynamics. While either can serve simply as inputs to the other (instead of a coupled system), this integration allows us to study policy adaptation patterns linked to measurable outcomes in the physical world (e.g. reduction of GHG emissions).
In simulating such a coupled framework, the societal response dynamics ultimately lead to either the acceptance or rejection of a policy. If a policy is accepted, it is implemented in the model, producing physical world implications such as reduced CO2 emissions. Conversely, if the policy is rejected, this too has physical world implications, such as continued GHG emissions or inadequate preparedness for extreme weather events. Integrated Assessment Models (IAMs) represent one possible class of models for this purpose, as they are designed to couple climate, energy, economic and land-use systems to project GHG emission pathways and evaluate mitigation and adaptation strategies53,54. However, traditional IAMs typically assume policy implementation as exogenous55,56, commonly overlooking the social dynamics that determine whether policies are adopted or abandoned—a gap that ACCPD explicitly addresses by making social acceptability endogenous and a key component of policy design.
Crucially, this layer is not limited to climate systems. It is designed as a modular interface to incorporate diverse domain-specific simulators, ranging from bottom-up physical system models (e.g., transportation, electricity grid) to economic models (e.g., energy pricing model), depending on the specific policy context. To demonstrate the framework’s capacity to integrate with established engineering tools across different scales, we reference two distinct implementation classes: micro-grid simulations for local energy policy can be done using GridLAB-D57) and national-scale climate risk and adaptation assessment to inform resilience planning can be performed by OpenCLIM58. However, developing suitable interfaces between social and physical models remains a significant challenge (e.g., time compatibility where one system operates in a different time scale than the other).
The Acceptability Frontier
In the Acceptability Frontier (AF), the term “Acceptability” refers to the concept of social acceptance. While conventionally defined as meeting a minimum threshold of public support necessary for a policy to remain viable59,60,61, the ACCPD framework proposes to treat acceptability not as a fixed binary variable, but as a continuous objective.
As previously noted (see Introduction), the determinants of climate policy acceptance range from social legitimacy and fairness concerns1 to the influence of entrenched powerful actors (e.g., fossil fuel industry employing disinformation tactics to weaken public support for climate policies such as renewable energy deployment5). Consequently, the specific threshold for ‘viability’ is a political factor, set and adjusted by human stakeholders, often informed by decision-support frameworks such as Multi-Criteria Decision Analysis (MCDA)62 or PESTEL (Political, Economic, Social, Technological, Environmental, and Legal)63 analysis.
The proposed Acceptability Frontier itself is the technical component facilitating this decision. By treating social acceptance as one objective and other policy factors, such as CO2 emissions and implementation costs, as additional objectives, we can identify an optimal policy through multi-objective optimization. The boundary of these optimal trade-offs is what we call the “Acceptability Frontier.” This is akin to a Pareto Frontier, which identifies a set of equally optimal policy options where different compromises are possible (refer to Supplementary Information file for further details).
We note that the multiple objectives are not necessarily equivalent in the context of optimization. For instance, if the overall goal is to limit warming to well below 2 °C, atmospheric CO2 cannot exceed a specific threshold; this is a fixed constraint. This limits the flexibility of the emission reduction objective within a multi-objective optimization process. In contrast, acceptability is not fixed; it can be influenced and adjusted. Therefore, optimizing for acceptability necessarily involves an iterative process of adjusting a policy or its communication to increase support while maintaining emission reduction goals. The ACCPD framework can be used to test these potential adjustments.
The frontier can be dynamic and expanded through strategic choices. For example, if a proposed climate infrastructure is projected to face resistance within a community, resulting in a lack of acceptable choices, we can look at different measures that allow us to expand this acceptability frontier. This includes (a) benefit redistribution that ensures affected communities capture value42, (b) stakeholder engagement that ensures meaningful participations in decisions64, (c) phased implementation such as running initial pilot studies where positive benefit experience by the community drives support beyond early adopters13, and (d) narrative reframing, for instance emphasizing co-benefits like jobs64 or resilience or appealing to our moral obligation of preventing harm and protecting others65. For narrative reframing, LLMs’ persuasive capabilities can be leveraged. The ACCPD framework can be used to systematically test how different policy design choices shape public responses.
The Observatory Layer: Monitoring, Validation, and Transparency
To keep ACCPD grounded and reproducible, we include an “observatory” layer that functions as a monitoring and audit module. Its role is to (i) compare model outputs with empirical signals (e.g., surveys, planning documentation, and other appropriate public indicators of response), (ii) maintain versioned records of datasets, prompts, parameters, and model configurations to enable reproducibility, and (iii) support structured recalibration when simulated trajectories diverge from observed signals.
The observatory can function as a hybrid-intelligence interface where communities, experts, policy makers, and other stakeholders examine results, question assumptions, and collectively adjust parameters. This layer does not prescribe a new governance institution; rather, it specifies a set of operational functions that can be carried out by existing oversight arrangements (e.g., research governance, regulatory review, ethics processes, or independent audits). In practice, stakeholders can (e.g., policy makers, civil service, city councils, citizen climate assemblies, etc.) act as the users of the system, utilizing the Acceptability Frontier to identify viable design constraints. Our focus on democratic application is deliberate, as recent literature demonstrates that participatory approaches are highly effective mechanisms for ensuring the long-term viability of climate infrastructure52,66. Therefore, our objective is to complement these proven participatory systems rather than deliver a framework that is agnostic to political governance. Yet, ACCPD may also find usage in other governance systems that may seek public buy-in for climate mitigation and adaptation projects67.
However, implementing such a system faces enormous challenges in data accessibility (e.g., tightening API policies of news and social media platforms), data integration, computational resources, and institutional coordination. It is worth noting that more data may not correspond to a more robust system. The choice of real-world signals should also depend on the required level of fidelity in the model’s key indicators, determined by the problem context.
