Project Overview

Reflective Resistance is a research through design project completed with the UC Berkeley CoDesign Lab under Prof. Kosa Goucher-Lambert, in collaboration with Nicole Goridkov (lead) and Diana Bolaños. The work was published at the 2026 ASME International Design Engineering Technical Conferences (IDETC), and explores how "design for sustainable behavior" (DfSB) strategies can be embedded into generative AI (GenAI) interactions to make their otherwise invisible environmental costs tangible.

We designed and built two artifacts that apply distinct DfSB strategies: a browser-based eco-feedback Chrome extension that visualizes real-time energy consumption during ChatGPT use, and a physical, eco-steering keyboard that dynamically modulates typing resistance in real time based on GenAI token usage. My primary contribution was leading the mechanical design and fabrication of the physical keyboard prototype, including the pneumatic bladder mechanism and the open-source heat-sealing fabrication method used to produce it.

The two artifacts were deployed together in a within-subjects user study (n=30), examining how physically-augmented versus purely digital feedback shapes GenAI usage patterns, cognitive workload, gaze attention, and reflective engagement with the environmental cost of AI.

Design Framing — Eco-Feedback & Eco-Steering

Our approach drew on two DfSB intervention strategies. Eco-feedback informs users clearly about what they are doing through real-time data, while eco-steering goes a step further by structuring the interaction so unsustainable behavior becomes physically harder to perform, while preserving user agency.

We used Pierce's counterfunctional design schema to frame the keyboard's physical intervention: normally, a user can engage in high-frequency, low-deliberation prompting; our device ensures a user can not sustain that behavior without expending physical effort; as a result, the user is pushed to engage with GenAI more deliberately and reflectively.

Digital Prototype — Eco-Feedback Extension

We built a Chrome extension that overlays real-time energy metrics directly onto the ChatGPT interface. Once a prompt is submitted, the overlay updates with session time, a color-graded resistance status bar, and energy-use equivalencies (iPhone battery charges, Google searches, minutes of LED light, and estimated cost), alongside token counts per query and per session.

Tokens were estimated using the standard heuristic of one token per four characters, with a 60x multiplier applied to prompts that generated images, reflecting the substantially higher computational cost of image inference relative to text.

Physical Prototype — Eco-Steering Keyboard

The physical prototype is a mechanical tactile keyboard augmented with an inflatable Thermoplastic Polyurethane (TPU) bladder positioned beneath the key switches. As GenAI token usage accumulates, the bladder inflates to exert variable opposing resistance against the user's keystrokes — to our knowledge, the first mechanism to dynamically vary typing resistance mid-keystroke, rather than relying on static keycap substitution.

Bladder inflation is driven by a motor-controlled air supply and Arduino microcontroller, which receives live token usage data from the Chrome extension and modulates air pressure via pulse-width-modulated (PWM) motor control.

Bladder fabrication. The bladder was fabricated from double-layered 0.1mm TPU film, heat-sealed to conform to the negative geometry of the keyboard's underside. Operated without filament, the printer's heated nozzle served as a controlled thermal bonding tool guided along predefined seam paths designed in CAD, exported to DXF, and processed through the custom toolpath pipeline. This method produced far more repeatable, consistent seam patterns than manual heat sealing, and the programmable workflow allowed systematic variation of temperature, travel speed, and pressure depth. Due to print-bed size constraints, the bladder was fabricated in three parts and subsequently joined. The full method is documented and shared publicly for future researchers working with soft pneumatic interfaces.

Assembly. The fabricated bladder was trimmed to accommodate the keycaps and switches, then connected to an air supply via flexible silicone tubing and 3D-printed valve splitters, sealed with rubber bands and secured with 3D-printed retaining rings. The air supply, motor driver, and Arduino microcontroller are housed in a 3D-printed enclosure, powered via a wall-connected supply, with both the Arduino and keyboard interfacing to the host laptop over USB.

Control Pipeline — Token to Physical Feedback

A three-stage pipeline translates GenAI token usage into physical keyboard resistance:
1. Token calculation — the extension's content script estimates token count per prompt (1 token ≈ 4 characters), with multipliers applied for image generation, accumulating into a running session total.
2. Resistance mapping — accumulated tokens are mapped linearly to a 0–100% resistance value (capped at 90 tokens), decaying 1.25% every 5 seconds during idle periods to mimic a computational "cooldown."
3. Serial transmission — the resistance value is relayed through a local Node.js/Express bridge server to the Arduino over USB, which maps it to PWM duty cycle (60–100%) with a ramped step size for smooth motor acceleration and a 31kHz switching frequency to minimize audible motor whine.

User Study — Method & Protocol

We ran an IRB-approved, within-subjects study with 30 participants to observe how each modality shapes AI use, perceived value of AI assistance, and moments of reflection. Participants completed a pre-survey, eye-tracking calibration, and a tutorial, then completed a ChatGPT-based image storyboarding task under two counterbalanced conditions: digital feedback only (browser extension), and physical feedback (browser extension + resistive keyboard). Participants thought aloud throughout, and completed post-task NASA TLX workload surveys, reflection questionnaires, and a closing interview.

Screen capture, audio, and gaze data (via the Beam Eye Tracker) were collected throughout, enabling both quantitative analysis of prompting behavior and a Locus-of-Attention Index (LAI) tracking how often participants' gaze fell on the feedback display versus the rest of the screen.

Attention Over Time

Gaze tracking revealed that attention in the purely digital condition remained stable throughout the task, directed consistently away from the browser extension. In the physical condition, attention was similarly stable through the first half of the task but declined significantly in the second half. This shift may reflect participants checking the energy tracker more as resistance increased and they began to feel its effects directly at the keyboard.

Perceived Workload

Post-task NASA TLX surveys showed the physical condition imposed significantly higher Mental Demand, Physical Demand, Effort, and Frustration than the digital condition, while Time Pressure and Performance were statistically indistinguishable between conditions. Despite this added burden, participants in the physical condition reported stronger agreement that the feedback prompted deliberate reflection and helped them maintain an overview of their GenAI use.

Discussion & Takeaways

While aggregate prompting behavior changed only modestly across conditions, interviews revealed that many participants meaningfully shifted how they thought about their GenAI use through negotiating real tradeoffs between output quality and environmental cost, and describing new feelings of ownership over their consumption. Augmenting digital eco-feedback with physical, embodied friction surfaced distinct patterns of reflection not observed with digital feedback alone, suggesting the two modalities are complementary rather than interchangeable. At the same time, several participants noted that current AI interfaces offer little scaffolding to act on this new awareness, pointing to an open gap between environmental awareness and actionable tools in sustainable AI interface design.

This project was my first deep dive into soft pneumatic actuation and closed-loop physical/digital systems, and pushed my fabrication skills into new territory.