Responsible AI for Designers: Why I Built PromptLite
AI has become difficult to separate from modern creative software.
Generative tools now sit inside world-class design applications I use professionally whether I like it or not – much like the Gemini and CoPilot integrations – alongside the traditional tools of graphic design, photography, illustration and image editing. Some embedded design functions have been replaced by AI, and in those instance, the choice has been made for us.
Ignoring AI altogether no longer feels particularly realistic to me – but embracing it uncritically doesn’t feel right either.
I’m a designer with a long-standing interest in sustainability and ethical practice, and my position on AI is fairly simple: use it when there is a genuine reason to, understand what you’re asking it to do, and avoid generating simply because you can.
That became my motivation for creating PromptLite.
I’m not anti-AI. I am cautious about how I use it, particularly when it comes to environmental impact and responsible AI in creative practice.
My own use of AI aims to stay light and strategic. Most of the time I use text-based tools: helping interrogate information, structure thinking or troubleshoot something technical.
When a project needs imagery, my first choice is still existing material wherever possible – client assets, photography, illustration or properly licensed stock libraries such as Adobe Stock, Shutterstock and Dreamstime (historically built on real imagery captured by humans).
Image generation comes further down the list for me.
Sometimes it genuinely is useful. It is increasingly integrated into the software designers already use, and there are situations where a generated element or image can solve a problem efficiently. But when I do use it, I want the process to be deliberate rather than an endless cycle of generate, dislike, generate again.
Research gives good reason to pay attention to that distinction. A 2024 FAccT study benchmarking the inference energy of different machine-learning tasks found generative systems considerably more energy-intensive than many task-specific systems, with image generation towards the high end of the tasks tested. The figures varied substantially depending on model, hardware and output settings, which is also why simplistic “one image equals X carbon” claims should be treated cautiously. DOI
The issue that interested me wasn’t only what one generation consumes, it was behaviour.
How many images do we generate before arriving at one usable result? Could better thinking happen before pressing Generate?
Turning an ethical concern into a design problem.
In 2025 I joined Cosmic’s SIFA Level 3 AI Workplace Productivity Skills Bootcamp. For my capstone project I decided to explore the environmental tension I felt around generative image-making rather than sidestep it. The resulting project, PromptLite – Responsible Generation for Creatives, was awarded a Distinction.
The original question was:
Could a designer reduce unnecessary AI image generations while still producing useful creative results?
PromptLite was designed as a lightweight checkpoint before image generation. Rather than generating an image itself, it worked entirely through text. A designer entered an intended image prompt and PromptLite refined it, assessed it using an experimental Eco Index, suggested a lighter alternative where appropriate, and produced a transparency note.
Only after that decision-making process would the designer choose whether generating the image was actually worthwhile.
The aim was never to build an AI tool that encouraged more AI use. It was to build something that might interrupt the drive to generate – and track the impact.
The Eco Ledger: making an invisible process visible.
Alongside PromptLite I developed an Eco Ledger.
This was essentially a record of each approved prompt: the original idea, its refined version, the suggested alternative, the Eco Index and a transparency note.
There was also an intentional human step. The final approved result had to be copied into the ledger before automation took over. I deliberately chose not to log every draft automatically because that would have removed the pause for judgement I was trying to create in the first place.
The broader prototype connected the Custom GPT, Google Sheets and a simple Zapier workflow so that approved entries could be logged and reported.
Efficiency shouldn’t mean automating every possible decision – sometimes responsible design means putting friction and questioning back into a process deliberately.
An important caveat: PromptLite does not measure real carbon.
The Eco Index and the project’s “Simulated Carbon Units” were experimental indicators, not measurements of electricity consumption, carbon dioxide or GPU energy use.
The scoring system used visual and prompt characteristics to make relative complexity visible and to encourage comparison between approaches. My capstone explicitly treated the figures as simulated and unsuitable for scientific carbon reporting.
That makes PromptLite primarily a behavioural design tool, rather than an emissions calculator.
The clearest behavioural benefit within the prototype is simple: if better planning reduces ten speculative image generations to three, seven generations haven’t taken place.
A future version should go further. Rather than relying on simulated indicators, I would love to see this kind of creative interface connected to genuine model-level usage information or environmental-estimation systems.
Developer tools are already moving in this direction. CodeCarbon can track computing on hardware you control, while EcoLogits provides environmental estimates for supported generative-AI API requests. docs.codecarbon.io
What interests me is bringing that kind of front-facing thinking into the everyday creative interface.
What happened when I tested it?
My capstone testing was deliberately small, so the findings should not be treated as a scientific study.
For example, across eight test prompts, the average simulated Eco Index fell from 10.8 to 4.1 after refinement — a 62% reduction in the index, not a measured 62% reduction in energy or carbon emissions.
More practically, the number of image iterations I was using per concept reduced from roughly ten to around three, and prompt-writing time reduced from around three-to-four minutes to one-to-two minutes. Every logged example also included a transparency note.
For me, that was the useful result. The experiment changed my behaviour.
Instead of treating prompting as trial and error, I began thinking more carefully about what I was trying to achieve before using the generation tool at all.
Where I think creative software could go next.
This is the part of PromptLite I would genuinely love someone to take further.
Imagine opening a generative tool inside your normal design software and seeing, alongside the Generate button, something like:
Estimated resource impact: Low / Medium / High
Perhaps the software could show how the chosen model, output dimensions, number of variations, generation mode and other measurable parameters affect likely resource use.
Perhaps it could also show:
- Images generated for this project: 3
- Estimated impact this session = X
- Try a lighter setting
- Use existing asset instead
- Add AI usage to project transparency record
Adobe already has pieces of the surrounding infrastructure. Firefly uses generative credits to represent usage, and Adobe applies Content Credentials to help record when generative AI has been used. Adobe.com.
What I would love to see added is the environmental layer – information at the point of decision.
Designers are used to making choices around paper stock, ink, print quantities, manufacturing methods and materials. Digital resource consumption could become another design constraint we learn to see and work with.
The same principle could be even more valuable in AI video production. Generative video can require substantially more energy than still-image generation, with demand varying according to the model, resolution, clip length, hardware and generation settings. A 2026 measurement study across 46 generative AI models found that video generation sometimes consumed more than 100 times the energy of image generation in the configurations tested. This makes video a logical next step for PromptLite: extending the system to consider factors such as clip duration, frame count, resolution and motion complexity before generation. arxiv.org.
Interestingly, when I developed PromptLite in 2025, I had already identified video as a future application, proposing that an extended Eco Index could consider frame count, motion complexity and render duration. Subsequent research into generative video energy use is examining closely related variables, including temporal length and resolution.
Responsible AI doesn’t mean rejecting technology.
One of the conclusions I reached through the project was that responsible AI use isn’t necessarily about refusing the tools.
It’s about understanding their limits and using them with intention.
My capstone described the tension as an “AI paradox”: no act of generation is impact-free, so the useful question becomes how consciously we use it – through awareness, restraint, fewer unnecessary iterations and honest disclosure.
That still describes my position well. I don’t want AI to dilute the craft I’ve spent more than twenty years developing. I don’t want it replacing human-made photography, illustration, existing stock or creative thought merely because generation is convenient.
But nor do I think designers gain much by pretending these tools aren’t becoming part of our working environment, and in some cases we have no choice to adapt to the trajectory of the very software that enables us to work.
For me, the middle ground is light, purposeful and transparent use – only where AI use is genuinely necessary.
- Explore and use what already exists before generating something new.
- Think before prompting.
- Use the lowest-impact method that solves the problem.
- Generate as little as necessary.
- And keep a human being responsible for the final decision.
PromptLite is a prototype and I’d love to see the idea grow.
PromptLite was created as a capstone prototype, not as finished commercial software. It deliberately used a small test set, simulated scoring and manual approval, and there is a great deal that would need to change before it could become a robust impact-measurement product. But I still believe the underlying design problem is worth pursuing and being made usable to all creatives.
How might we make the resource cost of generative creativity visible to the person making the creative decision?
If you work in responsible AI, sustainable computing, creative technology or software development – particularly within tools used by designers – I’d love to hear from you. PromptLite remains a prototype, but I believe the principle behind it has potential far beyond my own studio.
…
PromptLite was conceived and developed by Holly Stevens / Lead Creative in 2025. It formed the capstone project for Cosmic’s SIFA Level 3 AI Workplace Productivity Skills Bootcamp, where the project received a Distinction. The concept, naming, design system, workflow materials and supporting documentation are original work. Collaboration and licensing enquiries are welcome.