Large Language Models For Social Policy: A Deployment Ladder Framework For Sustainability-Aware AI Governance

1. Introduction: LLMs as a Social Policy Issue

As public institutions increasingly experiment with artificial intelligence, particularly large language models (LLMs) in policy analysis, questions of governance, responsibility, and long-term impact become unavoidable. While fundamentally agreeing with Feng and Chandra’s assessment of LLMs as epistemic technologies (2025) in policymaking, we argue for an integrated framework under which to view LLMs’ policymaking role that “advances” the social aspects of policy in line with United Nations’ Sustainable Development Goals (SDGs). Despite the widespread scepticism among academics regarding various governance aspects of artificial intelligence in general, and large language models in particular (Robles and Mallinson, 2025), there remains limited work on relating these concerns cohesively with UN’s sustainability goals of quality education (SDG4), reduced inequalities (SDG 10), responsible consumption and production (SDG12), climate action (SDG13), and peace justice and strong institutions (SDG16). Such cohesion also requires an interdisciplinary effort whereby practitioners realise the significance of framing LLMs as socio-technical infrastructure rather than neutral tools.

We propose an approach to such realisation that situates AI governance within a broader global sustainability and justice agenda through a framework we refer to as a “deployment ladder”. The ladder treats LLM usage in social policy contexts as a technical process built on top of foundational governance constraints, whereby the various methods through which LLMs can be incorporated into decision-making, i.e., prompting, retrieval augmented generation (RAG), finetuning, and reinforcement learning from human feedback (RLHF) are not mere choices but rather a scale for computational cost and environmental impact, governance burden, and public value. Our LLM deployment ladder is unique in that it includes a call to action for various social and technical stakeholders, particularly to address aspects of the UN’s Sustainable Development Goals.

2. The LLM Deployment Ladder (Prompting → RAG → Fine-Tuning → RLHF)

LLM deployment in social policy should be understood as a proportional governance architecture in which regulatory oversight, documentation obligations, and accountability mechanisms intensify in step with increasing technical complexity, environmental footprint, and public impact. Rather than a linear path toward technical sophistication, the Deployment Ladder—Prompting, Retrieval-Augmented Generation (RAG), Fine-Tuning, and RLHF—represents escalating governance commitments. As illustrated in Figure 1, each upward step increases environmental cost and policy risk and must therefore be justified by commensurate public value.

At the base, Prompting relies on pre-trained models without modifying weights. This avoids additional training-related energy expenditure, which in large frontier systems has reached thousands of megawatt-hours during training cycles (Strubell et al., 2019). Governance exposure is comparatively limited because institutional data are not embedded into model parameters, reducing long-term compliance risk under evolving data protection interpretations (European Data Protection Board, Opinion 28/2024). Prompting is proportionate for low-stakes tasks such as drafting policy briefs, summarising consultation responses, or supporting internal research synthesis, where public value lies in efficiency gains rather than in automated decision-making.

Retrieval augmented generation (RAG) introduces document-grounded inference by retrieving curated institutional sources at runtime. While inference costs increase due to embedding and vector search operations, RAG avoids the substantial environmental and retraining burden associated with a process known as weight modification which is a process through which different degrees of importance are attached to different institutional sources (Amore and Gentili, 2025). Importantly, it enhances traceability: outputs can be linked to verifiable policy documents, strengthening contestability and aligning with OECD principles of transparency and accountability in public-sector AI (OECD, 2022). Governance burden rises through source curation, audit logging, and documentation obligations. Public value correspondingly increases in contexts such as citizen-facing information systems or legislative interpretation support.

Fine-Tuning modifies model parameters using domain-specific datasets, increasing GPU[i] utilisation and carbon intensity. Embedding institutional knowledge within model weights complicates provenance tracking, deletion rights, and accountability under data protection frameworks (European Data Protection Board, Opinion 28/2024). Governance obligations expand to include data minimisation protocols, impact assessments, and enhanced oversight mechanisms (Section 4). Fine-tuning may deliver higher public value in domain-sensitive contexts, such as benefits in eligibility pre-screening. Still, such escalation must demonstrate measurable societal benefit relative to its environmental and regulatory cost (Section 3).

At the highest rung, Reinforcement Learning with Human Feedback (RLHF) encodes normative alignment signals through reinforcement optimisation. This configuration requires iterative human annotation cycles and additional compute-intensive training phases, contributing to the broader rise in AI-related data-centre energy consumption documented across Europe (Lin et al., 2024). Governance intensity peaks, as normative judgements become embedded within opaque optimisation processes, raising democratic accountability concerns consistent with ethical AI frameworks (Paraman and Anamalah, 2023). While RLHF may enhance consistency in large-scale automated service delivery, its environmental footprint and oversight demands necessitate explicit justification through demonstrable public value.

The Deployment Ladder therefore, operationalises proportionality. Technical escalation is legitimate only where increasing societal benefit outweighs environmental impact, governance burden, and accountability risk. In social policy contexts—where sustainability, rights protection, and public trust are foundational—innovation must remain disciplined by transparency, regulatory responsibility, and long-term societal impact (Sections 3–6).

Figure 1: Proportional LLM Governance Architecture. The staircase model illustrates how escalating technical sophistication in LLM deployment increases environmental impact and governance burden while potentially enhancing public value, underscoring proportionality and sustainability as core principles for responsible public-sector use

3. Green AI and Sustainability as Governance Principles

The proportional governance architecture introduced in Section 3 establishes that technical escalation must be matched by escalating oversight and accountability. Sustainability is therefore not an auxiliary consideration but a structural governance constraint. Data centres already account for a growing share of electricity consumption across Europe, and AI-driven workloads are projected to intensify this demand in the coming decade (Strubell et al., 2019). Because Large Language Models (LLMs) require substantial computational resources—particularly during fine-tuning and reinforcement-based optimisation—environmental impact becomes inseparable from deployment design.

The Green AI paradigm calls for energy consumption and carbon intensity to be treated as first-order evaluation criteria alongside performance gains. Within a proportional deployment model, upward movement along the ladder—from prompting to RAG, fine-tuning, and RLHF—entails increasing computational demand and associated carbon intensity (Amore and Gentili, 2025). These increases are not incidental; they are structural consequences of architectural choice. As such, environmental escalation must be justified by corresponding gains in public value.

Figure 1 visualises this stepwise escalation, illustrating how environmental footprint and governance burden increase together. Figure 2 complements this logic by introducing the foundational governance pillars that constrain all deployment decisions. Sustainability, trust and accountability, GDPR-aligned oversight, and public legitimacy operate as cross-cutting obligations at every rung of the ladder—not merely at higher levels of technical sophistication.

This framing aligns with OECD principles for responsible AI in the public sector (OECD, 22) and broader European commitments to trustworthy and sustainable AI development (Regulation (EU) 2024/1689). Under this model, sustainability functions as a legitimacy condition. Public institutions must make environmental trade-offs explicit, measurable, and contestable—particularly in domains such as welfare administration, health communication, and education services, where systems operate at scale.

Green AI in social policy is therefore not about minimising energy use in isolation. It requires disciplined proportionality: more computationally intensive configurations are permissible only where demonstrable public value outweighs environmental and regulatory cost. Embedding sustainability within deployment governance reinforces climate responsibility, strengthens public accountability, and ensures that technological ambition remains aligned with long-term societal objectives.

Figure 2: Proportional LLM Governance Framework. LLM deployment options entail increasing governance burden, environmental impact, and explainability risk, highlighting proportionality and sustainability as guiding principles for public-sector use.

A proportional deployment framework supports Green AI objectives by encouraging institutions to adopt the least resource-intensive approach capable of achieving a clearly defined policy goal. In this sense, Green AI is not about minimising energy use at all costs, but about ensuring that environmental impact is justified, visible, and accountable. More complex or computationally demanding approaches may be appropriate in high-stakes contexts, but they should be explicitly motivated rather than treated as a default.

Sustainability considerations also intersect with public accountability. AI systems used in social policy are typically funded through public expenditure, meaning their environmental costs are ultimately borne by society through public expenditure and infrastructure use. Making sustainability an explicit governance criterion helps ensure that these trade-offs are recognised and debated, rather than remaining implicit.

Embedding sustainability into LLM governance therefore reinforces both environmental responsibility and public trust. By treating Green AI as a core policy concern, policymakers can ensure that innovation in social policy remains effective, legitimate, and aligned with long-term societal goals.

4. Rights-Based Safeguards: GDPR and Human Oversight

Beyond formal compliance, rights-based safeguards require translating legal principles into operational practices. Human oversight, documentation, and traceability are not merely regulatory obligations but governance mechanisms that shape how LLMs are designed and deployed. In social policy contexts, where automated outputs may influence access to services or information, ensuring contestability and meaningful human intervention is essential. This perspective shifts GDPR from a compliance checklist to an active design constraint that informs proportional deployment and reinforces public trust.

5. Trust, Explainability, and Accountability

Trust is a foundational requirement for the use of AI systems in social policy. When Large Language Models (LLMs) are used to support decision-making, public communication, or service delivery, citizens and stakeholders must be able to understand, at least at a high level, how and why outputs are generated. Without sufficient transparency, even technically accurate systems risk undermining institutional legitimacy and public confidence.

Explainable Artificial Intelligence (XAI) provides a framework for addressing this challenge by emphasising intelligibility, accountability, and contestability. In policy contexts, explainability does not require exposing complex model internals or technical details. Rather, it involves ensuring that system behaviour, limitations, and sources of information can be meaningfully communicated to decision-makers, auditors, and affected individuals. This form of explanation supports oversight, enables informed judgement, and allows decisions to be questioned or corrected when necessary.

Importantly, explainability is not solely a technical feature but a socio-technical property shaped by deployment choices and institutional processes. Systems that rely on verifiable sources, documented workflows, or human oversight are inherently easier to justify and audit than opaque, heavily fine-tuned models whose behaviour may be difficult to trace or interpret. As a result, choices about how LLMs are deployed have direct implications for accountability and trust.

Embedding explainability into deployment decisions allows policymakers to balance innovation with responsibility. By aligning system design with expectations of transparency and oversight, public institutions can strengthen trust while retaining flexibility in how LLMs are used. In social policy contexts, where decisions often affect vulnerable populations, such alignment is essential for maintaining both democratic legitimacy and public confidence.

6. Evaluation Beyond Accuracy: A Policy Scorecard

Section 2 established that LLM deployment in social policy operates within a proportional governance architecture: technical escalation increases environmental footprint, regulatory exposure, and oversight obligations, and must therefore be justified by demonstrable public value. Evaluation is the mechanism through which this proportionality principle becomes enforceable.

Conventional AI assessment frameworks prioritise technical metrics such as accuracy, latency, or efficiency. While relevant, these indicators are insufficient in public-sector contexts where legitimacy, sustainability, and rights protection are co-equal governance objectives. A highly accurate system may nonetheless be disproportionate if it embeds unnecessary computational intensity, introduces opaque optimisation processes, or generates heightened compliance burdens. Conversely, a less complex deployment may deliver adequate performance while better aligning with environmental responsibility and democratic accountability.

A policy-aligned evaluation framework must therefore extend beyond performance metrics to assess whether architectural choice is justified relative to societal benefit. This scorecard approach directly operationalises the Deployment Ladder (Section 2) and the sustainability constraints articulated in Section 4, while reinforcing rights-based safeguards (Section 4) and explainability requirements (Section 5).

Table 1: Proportional LLM Policy Evaluation Scorecard

This scorecard translates proportionality into decision criteria. Movement upward on the Deployment Ladder must demonstrate measurable societal benefit sufficient to outweigh increased environmental cost and governance burden. By making trade-offs explicit, the framework ensures that sustainability, accountability, and public legitimacy remain central to LLM adoption in social policy environments, as illustrated in Figure 3.

Figure 3: Proportional LLM Policy Evaluation Matrix. The matrix operationalises proportional governance by mapping LLM deployment modes—Prompting, RAG, Fine-Tuning, and RLHF—against public value justification, governance burden, sustainability impact, data protection and rights risk, explainability, and governance escalation requirement.

Evaluation plays a central role in determining whether AI systems are fit for use in social policy contexts. However, prevailing assessment practices often prioritise technical performance metrics, such as accuracy or efficiency, while overlooking broader societal and governance considerations. For LLMs deployed in public services or policy processes, such a narrow focus is insufficient.

In public-sector contexts, accuracy alone provides limited insight into whether an LLM deployment is appropriate, legitimate, or sustainable, as these qualities depend on proportional deployment choices, governance burden, and societal impact rather than technical performance alone.

A policy-oriented evaluation framework should therefore move beyond accuracy to consider a wider set of criteria relevant to public interest and accountability. These include transparency and explainability, governance burden, sustainability and environmental impact, data protection risks, and the degree of human oversight required. Together, these dimensions provide a more holistic basis for judging whether an LLM deployment is appropriate for a given social policy task.

This broader perspective enables policymakers to make informed trade-offs explicit. A highly accurate system may still be unsuitable if it imposes excessive governance costs, lacks transparency, or entails disproportionate environmental impact. Conversely, a less complex system may offer sufficient performance while better aligning with sustainability goals and public trust requirements.

Framing evaluation as a policy scorecard, rather than a purely technical benchmark, supports responsible decision-making and encourages proportional deployment. It also creates space for interdisciplinary collaboration, allowing social scientists, policymakers, and technologists to contribute jointly to defining what “success” means in the context of public-sector AI use.

References

Feng, N., and Chandra, Y. (2026). The Epistemic Impact of Large Language Models on Policymaking. Policy Studies Journal.

Robles, P., & Mallinson, D. J. (2025). Artificial intelligence technology, public trust, and effective governance. Review of Policy Research, 42, 11–28.

Strubell, E., Ganesh, A. and McCallum, A. (2019). Energy and policy considerations for deep learning in NLP. In Proceedings of the 57th annual meeting of the association for computational linguistics (pp. 3645-3650).

European Data Protection Board. (2024). Opinion 28/2024 on Certain Data Protection Aspects Related to the Processing of Personal Data in the Context of AI Models

Amore, G. and Gentili, A. (2025). AI and climate: an ethical sustainability framework for balancing risks and responsibilities. AI & SOCIETY, pp.1-11.

Organisation for Economic Co-operation and Development (OECD). 2022. OECD Recommendation on Artificial Intelligence. Paris, France: OECD Publishing.

Lin, Y., Lin, H., Xiong, W., Diao, S., Liu, J., Zhang, J., Pan, R., Wang, H., Hu, W., Zhang, H. and Dong, H. (2024). Mitigating the alignment tax of RLHF. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (pp. 580-606).

Paraman, P. and Anamalah, S. (2023). Ethical artificial intelligence framework for a good AI society: principles, opportunities and perils. AI & SOCIETY, 38(2), pp.595-611.

Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) (2024) Official Journal of the European Union, L 2024/1689. Available at: http://data.europa.eu/eli/reg/2024/1689/oj

[i] Graphics Processing Unit:  specialized computer chip designed to handle massive amounts of data simultaneously, making it incredibly fast at processing.