Source-linked AI summary
Environmental Slow AI: Design Principles for Generative Systems
Vanessa Utz
TL;DR
Generative AI’s maximalist, frictionless designs embed cultural values while contributing to substantial environmental consequences, motivating a design alternative. The paper develops environmental Slow AI by connecting environmental humanities and Slow Technology to five interface-level principles. It concludes that these principles can restore human agency by reintroducing decisions removed by frictionless defaults, while warning that implementation can shift responsibility onto users or produce greenwashing.
Problem
Current generative-AI design treats maximal output and frictionless invocation as defaults, embedding cultural values while environmental consequences continue to grow.
Method
The paper develops environmental Slow AI through environmental-humanities analysis and specifies five principles operating materially and interpretively at the user interface.
Results
The paper specifies restraint, sufficiency, selectivity over retention, material visibility, and friction as affordance as an environmental design program for generative systems.
Takeaways & Limitations
The principles restore human agency by reintroducing decisions that frictionless defaults had silently removed and embedding interpretive reflection into design.
Takeaways & Limitations
A Slow AI focused primarily on individual users risks responsibilisation, greenwashing, guilt without agency, and disproportionate burdens while infrastructure remains unchanged.
Abstract
from arXiv · showhide
Generative AI (genAI) systems produce cultural artefacts at scale, but they also reflect embedded cultural values through their design. Once identified, these values become open to deliberate reshaping. This position paper examines the maximalist values of current generative AI through an environmental humanities tradition and proposes design principles in which environmental sustainability serves as the core value instead. The principles are developed under the umbrella of Slow AI, a term that already circulates across several distinct research and practice programs. Five design principles are articulated (restraint, sufficiency, selectivity over retention, material visibility, and friction as affordance), each of them illustrated against the current design of widely deployed systems. Each principle operates at two levels: a design implementation, and an interpretive layer at which users and developers are prompted toward reflective engagement with the system. Together these principles extend human agency by restoring decisions that frictionless defaults have silently removed and do so by building interpretive reflection into design.
1. Introduction
Current generative AI is designed around maximal output and frictionless invocation, embedding cultural values that increase use. The paper argues these defaults are not technically necessary and proposes environmental sustainability as a different organizing value.
- Current design values: Generative AI expresses maximalism and zero-friction through defaults such as multiple images, perpetual prompting, and one-click regeneration.Midjourney returns four images by default, while ChatGPT Image 2 returns up to eight in thinking mode.
- Current design values: These interfaces increase usage, with power users producing 2,000 to 5,000 images per week when text-to-image systems first became widely available.
- GenAI and culture: GenAI both produces cultural artefacts at scale and embeds cultural values through output, limitation, storage, and cost choices.
- Environmental stakes: The paper identifies substantial environmental consequences across inference energy, water demand, storage, hardware manufacturing, and use-phase climate impacts.GPT-4 training consumed 6 percent of West Des Moines’ city water in a single month.
- Proposed direction: Slow AI treats sustainability as a primary commitment and invocation as a meaningful design decision rather than a choice pre-answered by frictionless defaults.
2. Slow AI and Adjacent Programs
Environmental Slow AI connects Slow Technology, slow design, and Sustainable AI by treating environmental protection as the cultural value organizing generative-AI interfaces. It restores deliberation rather than simply reducing AI use.
- Lineages: Slow AI extends Slow Technology’s emphasis on reflection and slow design’s sustainability-oriented principles.
- Related programs: The term circulates across archaeological, educational, creativity, and decolonial design practices that restore deliberation against treating friction as a cost.
- Environmental specification: Its environmental case requires both sustainability commitments and attention to rebound effects.
- Environmental specification: The proposal sits at the intersection of these fields, using Slow Technology for restraint, sufficiency, and reflection and Sustainable AI for its environmental diagnosis.
3. Slow Violence and Intergenerational Equity
The paper uses slow violence to explain why generative AI’s environmental harms remain difficult to perceive and intergenerational equity to motivate action across generations. Slow AI therefore directs attention toward accumulated and deferred costs rather than simply slower use.
- Slow violence: Slow violence describes gradual, dispersed harm that escapes ordinary aesthetic and political recognition.
- Slow violence: At scale, genAI’s inference harms are distributed across interfaces, data centres, cooling infrastructure, and semiconductor fabrication.
- Slow time: Nixon’s framework identifies how fast-time framings suppress the legibility of accumulation, inheritance, and deferred cost.
- Slow time: Slow AI uses slow time as a mode of attention to phenomena obscured by fast-time framings, not as a proposal to use generative systems more slowly.
- Intergenerational equity: Intergenerational equity provides the ethical motivation by treating the planet as held in trust across generations.
4. Design Principles
The paper proposes five design principles that change system defaults while prompting reflective engagement by users or developers. Together, they restore decisions that frictionless interfaces have removed without requiring every system to implement every principle.
- Design Principles: Each principle operates materially by changing system behavior or defaults and interpretively by demanding reflective engagement.
- Restraint: Restraint makes non-use a first-class option by surfacing whether generative assistance suits the task.It responds to documented gaps of impatience and overreliance in users’ decisions about AI assistance.
- Sufficiency: Sufficiency sizes outputs to the task rather than rewarding maximal production or exploration.
- Selectivity over Retention: Selectivity over retention makes preservation opt-in, reducing storage load and returning curation to the user.
- Material Visibility: Material Visibility surfaces inference costs such as energy and water at the point of interaction instead of leaving them unseen.
- Friction as Affordance: Friction as Affordance retains pauses, confirmations, and prompt-revision guidance where an operation’s cost or consequence demands reflection.It interrupts action, while Material Visibility presents its hidden costs.
- Scope: The five principles do not exhaust the design space, and no single system needs to instantiate all of them.They provide a target for subsequent empirical building and evaluation.
5. Evaluation Implications
The paper argues that evaluation practices should add environmental restraint and material cost to existing capability measures, rather than replacing them.
- Current benchmarks and leaderboards reward maximalist definitions of progress, including output count, visual fidelity, and prompt alignment.
- A proposed Restraint Rate measures how often systems decline, redirect, or recommend non-use for prompts with comparable conventional solutions.
- Environmental restraint can extend refusal-rate evaluation as a parallel quality dimension for generative systems.
- Existing evaluation practices should be supplemented, so comparable capability at substantially lower material cost counts as better performance.
6. Limitations
The proposal acknowledges risks of shifting environmental responsibility onto users and of rebound effects that increase aggregate demand despite local moderation.
- A user-focused Slow AI could reproduce consumer responsibilisation by burdening individuals while leaving production infrastructure unchanged.
- The proposal locates responsibility in system design, assigning commitments to builders rather than requiring environmental virtue from users.
- Rebound may increase aggregate demand as generative AI diffuses into new populations and use contexts, even when individual interactions become more moderate.
7. Conclusion
The conclusion presents environmental sustainability as a cultural value for generative AI design and specifies Slow AI as a user-interface and inference-layer framework.
- Environmental Slow AI organizes generative AI design around sustainability, slow violence, intergenerational equity, and the slow technology lineage.
- Its five principles restore human agency by reintroducing decisions that frictionless defaults had silently removed.
- Each principle operates materially and interpretively, embedding reflection on environmental harm directly into design.
- The proposal treats generative AI use as a meaningful act that should be accountable through evaluation and concretely specified at the interface.
Impact Statement
The impact statement warns that environmental Slow AI could cause harm when its vocabulary is adopted without substantive agency, equitable implementation, or attention to infrastructure.
- Greenwashing could result if the proposal’s vocabulary is adopted without implementing its substantive design commitments.
- Displaying environmental costs without giving users agency to act could produce guilt rather than informed engagement.
- Restraint could disproportionately burden users lacking non-AI alternatives while leaving infrastructural decisions untouched.
- Implementations should attend to who bears the proposal’s costs and who designs its systems.