Research

This work establishes how much energy diffusion-based image generation consumes once it reaches consumer scale. The work was among the earliest peer-reviewed treatments of inference-phase cost in consumer generative AI, published when the field was still measuring training almost exclusively, and, as far as I have been able to establish, the first to frame mass adoption itself as the primary environmental risk. Producing the estimates meant assembling fragmentary public figures on active users, hardware, and output volume, because the operators of these systems do not publish what would be needed to calculate the answer directly. That absence is a finding in its own right, and it recurs across the rest of this research.

Bytes held in a data centre draw power whether or not anyone looks at them again, and consumer generative AI produces them at a rate nobody is counting. Lifecycle assessments of these systems account for training and inference but treat storage as background load, so the content they generate has no place in the environmental ledger and no measures attached to it. This strand builds both. It names the behavioural pattern driving the volume, distinguishes the output a user commits to from the intermediate generations discarded on the way there, and converts the resulting burden into workflow-level metrics that a disclosure regime could require. Enterprise figures put storage energy intensity at roughly 46 kWh per terabyte per year, with around half of stored enterprise data never accessed after collection. Consumer-generated content carries none of the organizational incentives that keep enterprise data under review, so its accumulation rate, and its burden, may exceed that baseline considerably.

The maximalist design of consumer generative AI is a cultural commitment rather than a technical constraint. Midjourney returns four images per prompt by default, and ChatGPT presents a perpetually available prompt field alongside a single-click regenerate button. Neither choice follows from anything about how the underlying models work. Technologies embed the values of the people who build them, so a different set of values would produce a different system, and this strand specifies what that alternative would commit to. Five principles collected under the heading of Slow AI each operate at two levels, the design change that implements the principle and the interpretive shift it asks of the user. The lineage runs back to Slow Technology, which argued in 2001 that systems built for reflection rather than efficient throughput serve their users differently. What each principle does is return a decision that frictionless defaults currently make on the user's behalf without surfacing it.

U.ness is a locally-run, open-source generative AI system that displays what each prompt costs. It runs distilled language models through Ollama behind a browser interface built with Gradio, and its Energy Use Indicator reports per-prompt watt-hours, cumulative session energy, and a rolling A-to-E efficiency grade after every query. Nothing leaves the machine it runs on. The system addresses four dimensions of user agency that commercial platforms resolve on the user's behalf, namely access, data ownership, environmental cost, and operational control.

The system exists to test what published emissions data cannot settle, which is whether people change what they do when the cost is visible at the moment they act rather than reported in aggregate afterwards. A between-subjects study with fourteen participants found that they do. Participants who saw the indicator consumed roughly half the energy of the control group across an identical three-task workflow and submitted around eight prompts per session against the control group's fourteen, with no significant difference in how long the sessions ran. The sample is small and the tasks were short, so this establishes a behavioural signal worth pursuing rather than a population-scale effect.

The interviews qualified the finding. Participants valued the visibility but described the watt-hour figure as a measurement they had no way to act on, and asked for guidance rather than more numbers. That gap set the agenda for the work on prompting strategy as an energy variable, which measures what a user can actually change.

Two papers reporting this work are currently under submission. Restoring User Agency in Generative AI: Inference-Phase Energy Visibility introduces the system and reports the study described above. Rethinking Prompt Engineering Through an Energy Lens: An Empirical Study of LLM Inference Consumption measures per-query energy across seven prompting strategies, five task scenarios, and two inference modes, and supplies the evidence base for the guidance layer the next iteration of U.ness will carry.

Developed at the iViz Lab under the supervision of Steve DiPaola, with Rafael Arias Gonzalez.

Visit the project site →

Before moving to sustainability, my work examined how the visual properties of a stimulus determine a viewer's response, and how computational systems can be used to control those properties experimentally. Deep learning systems can produce images and videos that observers do not reliably distinguish from human-made artworks, which means a stimulus set can be built to an experimental specification rather than assembled from whatever existing works happen to approximate it. Control over the generator becomes control over the independent variable. The aesthetics studies used that leverage to test how the two processing streams of the visual system interact with aesthetic judgment, returning initial evidence that movement within a percept engages reflexive attention and lowers the judgment that follows.

A parallel line asked what the visual system extracts from images that never reach awareness, testing semantic priming from pictures rendered invisible by backward masking and by continuous flash suppression. The backward-masking condition did not replicate the priming effect reported in the earlier literature, a result worth stating plainly given how much of that literature rests on effects of this kind. A third line applied the same knowledge to design rather than measurement. Faces carry identity and emotional expression at the same time, so conventional anonymization removes both, and this work tested AI stylization as a way to obscure the first while preserving the second in research interview footage.

This research secured the multi-year funding that supported my doctorate. Its methods transfer directly to the work above, since an interface is also a stimulus whose properties shape behaviour, and the same experimental design, controlled stimulus construction, and behavioural measurement underwrite the U.ness study.