Moving from Perceptible to Palpable Environmental Costs of Generative AI through a Design Fiction Workshop · ACM Full Paper ↗
“What if every AI prompt came with a physical price — Jumping jacks, Snacks, and Water you had to consume?”
Generative AI has brought unprecedented convenience to everyday life, yet this technological leap comes at a significant environmental cost — powering every digital request demands a massive yet largely invisible energy infrastructure.
To explore ways of raising user awareness and prompting reflection on this issue, we conducted a design fiction workshop where 13 participants physically enacted computational labor, translating GPU latency into physical exertion and energy consumption into caloric intake.
This somatic engagement successfully rendered the abstract weight of consumption palpable, serving as a critical provocation against the frictionless interaction for generating AI content. We invite the community to further explore new forms of human-AI interaction that creatively yet reflectively engage with AI's environmental impact.
Global data center electricity consumption reached approximately 415 terawatt-hours in 2024 — roughly 1.5% of total global electricity consumption — and is projected to double by 2030, with AI identified as the primary driver.
Inference-stage computation accounts for the vast majority of AI's operational energy use, meaning everyday usage patterns directly shape AI's environmental footprint. Recent behavioral analysis reveals these systems are frequently used for non-essential interactions.
Despite growing data on AI's environmental impact, interfaces present computing power as an infinite resource, masking heavy infrastructure and energy costs behind every keystroke. This seamlessness fosters mindless trial-and-error, causing even environmentally conscious users to waste resources unintentionally.
We mapped digital inputs to tangible metabolic penalties so teams could see cost accumulate in real time.
We invited 13 graduate students from the Department of Design (ages 23–35) to physically embody the machine: divided into three teams, wearing chip-style badges assigning roles as Memory, CPU, Algorithm, or GPU, with wrists tethered by wire to a Black Box symbolising hidden algorithms.
Participants crawled through a frame representing a computer window to enter — from that moment, they stopped being human users and became components within the system. The workshop centered on a three-round “Reverse Prompt Engineering” competition: every digital action triggered a physical cost — the GPU participant performed jumping jacks until the image fully loaded, while metabolic penalties (snacks and water) accumulated and were distributed after each round.
Example prompts included “A cat with wings” and “Standing on a book.” While the GPU jumped, the room watched loading time become bodily fatigue — and snacks piled up as unavoidable evidence of cost.
Participants moved from merely seeing the cost to deeply feeling it. As metabolic debt cluttered the workspace, abstract numbers became unavoidable visual evidence.
P2: “Seeing snacks and drinks accumulating on the table, made the burden appear.” · P1: “Watching him jump until exhausted — I suddenly realized this cost was something we had created.”
Team System 001 made six attempts in Round 1 with heavy penalties and no win. They pivoted to a strict single-prompt strategy for Rounds 2 & 3 — minimising burden while earning first place both times.
When cost becomes palpable, users abandon mindless iteration for quality-focused decision-making.
System 003 evolved from “An octopus in deep blue sea” in Round 1 to an exhaustive blueprint in Round 3. Physical constraints compelled users to articulate complex details immediately, replacing iterative guesswork with precise definition.
Current AI design prioritises frictionless interaction, inevitably obscuring its environmental reality. Our work demonstrates that strategic discomfort — bodily fatigue and caloric accumulation — serves not as a usability barrier but as a critical provocation, transforming passive consumers into active infrastructure components.
We argue this points toward a Slow AI paradigm, where users sensitised by physical memory consciously weigh the necessity of each prompt against its ecological footprint.
Future directions include tabletop games simulating computational labor, classroom AI literacy kits, and public installations where passersby perform jumping jacks to generate AI images in real time — making the hidden cost of every prompt viscerally felt at community scale.