Sustainable Prompting: Inspiration for UX Teams

AI makes many things easier. Faster analyses, better prompts, more output in less time. But somewhere between the third tool test this week and the twentieth prompt attempt for the same briefing, a question arises that most people quickly brush aside: What does this actually cost?

Not in euros. But ecologically.

This article is not a call for asceticism, nor is it a lecture. It is food for thought because the topic is still rarely discussed in the UX community, even though we work with these tools every day.

📌 Key takeaways

• AI infrastructure consumes resources on three levels: training, daily use, and hardware production

• Individual behavior is part of the puzzle, but not a free pass for thoughtlessness

• The dilemma is real: Those who don’t participate fall behind. Those who blindly participate consume resources without reflection

• Long chats are rarely a good idea, neither for quality nor for resource consumption

• Clear prompts, conscious tool selection, and a critical eye toward image and video generation are concrete starting points

• Perfection is not the goal. Awareness is.

  
  

What actually makes AI usage resource-intensive?

AI usage is not a homogeneous entity. Resource consumption arises at three different points, which are often lumped together.

The training of a large language model is the most obvious cost driver. Weeks or months on thousands of specialized chips, in data centers that operate on very different energy mixes depending on their location. This happens once, or again with new model versions, and is completely beyond our control as users.

Daily use—that is, every single request to a model—costs significantly less than a training process. But it happens billions of times a day. Cumulatively, this is significant, even if a single prompt seems small in comparison.

Hardware production is the least visible part. GPUs and TPUs require rare earth metals and energy-intensive manufacturing. Refresh cycles are short because the model landscape evolves so rapidly. The environmental footprint of the hardware is attached to every tool we use before we’ve even written a single prompt.

Added to this is a factor that is often forgotten: many data centers cool their servers with water. Water consumption is a more serious problem than electricity demand in some regions, but it is rarely discussed in public debate.

It is difficult to say what all of this means in concrete numbers. The major AI companies do not publish consistent lifecycle data. That is a problem in itself, because it makes informed decisions difficult.

The dilemma no one likes to talk about

This is where it gets uncomfortable.

Anyone working in UX who wants to use AI seriously can’t avoid one thing: you have to know the tools. New models, new capabilities, new limitations. Those who don’t will give worse advice, write worse prompts, and lose touch with a development that isn’t slowing down.

This applies on a personal level. It applies even more strongly at the corporate level.

A team that consciously opts out of the next model hype risks competitors becoming faster, cheaper, or more convincing. That’s not paranoia; it’s market logic. And this logic creates a structural pull toward more usage, more testing, more iterations. Regardless of whether that makes sense in individual cases.

The honest answer to the question “Can I work with AI in a sustainable and competitive way?” is: Yes, but only if awareness is actively cultivated. It doesn’t happen on its own. The default is consumption, not reflection.

The dilemma cannot be talked away. But it can be navigated more consciously.

What we can actually influence

No area where individual decisions solve the structural problems. But there is also no area where individual decisions play no role. Both are true at the same time.

Here is what actually lies within our own daily work.

Long chats are rarely a good idea

Anyone who works regularly with LLMs is familiar with the phenomenon: At a certain point in the conversation, the quality deteriorates. The model loses its train of thought, repeats itself, or produces output that has less and less to do with the original goal. The context becomes too long, too diffuse.

The pragmatic solution is a new chat. Fresh context, clearer focus, better results.

What many don’t consider: This fresh start isn’t just methodologically sound. A long, tangled chat that processes more and more tokens without delivering usable output is also more resource-intensive than two short, focused conversations with a clear goal. Quality and efficiency go hand in hand here.

Clear Prompts Instead of Trial and Error

Poor prompting is expensive, in every sense. If you ask an unclear question, you get an unclear answer, and then you have to iterate five times to reach your goal. If you ask precisely, you need fewer attempts.

This is not a new argument for good prompting. But it takes on an additional dimension when you factor in resource consumption.

Methodologically sound work and more sustainable work are one and the same here.

A helpful check before sending: Would I ask this question to a human expert in the same way? If the answer is no, it’s worth refining the prompt further.

Choose tools, don’t collect them

Every new model that’s announced triggers a reflex: test, compare, categorize. That’s understandable and professionally necessary. But there’s a difference between informed testing and reflexive curiosity.

Conscious tool selection doesn’t mean ignoring new developments. It means asking yourself: Do I need this for my specific work? Does it solve a problem I actually have? Or am I testing it just because everyone else is?

A smaller, well-known selection of tools usually leads to better output than constantly switching between models, because you know the strengths and limitations of your own tools.

Consider image and video generation separately

Text prompts and image generation are not the same thing. Image and especially video generation is significantly more resource-intensive than text processing. That’s not an argument against using these capabilities, but an argument for using them more consciously.

The question “Do I really need an AI-generated image here, or would a stock photo or a sketch suffice?” is one that comes up more often in everyday UX work than you might think.

Can we work sustainably and competitively?

No easy answer. But an honest one.

The structural dilemma (competitive pressure versus ecological responsibility) cannot be resolved through better prompts or shorter chats. That would be naive. It requires decisions at the corporate and industry levels, more transparent data from the AI companies themselves, and a societal discourse that has barely begun.

What we as UX professionals can influence is our own way of working, and the way we bring this topic into our teams and organizations.

Awareness is not the same as a solution. But without awareness, there is no solution. And right now, the vast majority of AI users are working without even beginning to consider this awareness.

That is the starting point.

Conclusion

Sustainable prompting is not a finished concept. It is an attitude that manifests itself in small decisions: clearer prompts, shorter chats, more conscious tool selection, a second thought before the next image generation.

This doesn’t solve the structural dilemma. Competitive pressure is real, development doesn’t wait, and the environmental costs of AI infrastructure are largely beyond individual control.

But awareness changes behavior. And changed behaviors within a community that works with these tools every day are not an irrelevant factor.

The first step is to stop pushing the question aside.

FAQ

Does it make a difference which model I use? 

Yes, but the data on this is scarce. Smaller, specialized models are generally more resource-efficient than large frontier models when they perform the task just as well. The problem: companies do not publish consistent comparative data. The pragmatic recommendation is to use the smallest possible model that reliably performs the task.

Is it more sustainable to run models locally? 

It depends. Local models shift energy consumption to your own device and your own electricity mix. Those who use renewable energy have a real advantage here. Those who use conventional electricity may not. Furthermore, locally executable models are currently still significantly smaller and less powerful than cloud models, which makes comparison difficult.

How do I deal with competitive pressure without blindly jumping on every trend?

A helpful distinction: Staying informed is a must; implementing every trend immediately is not. A structured monthly review of new developments, combined with a clear assessment of the concrete benefits for your own work, helps separate knee-jerk reactions from thoughtful decision-making.

Is there such a thing as a “green” AI provider? 

Some providers actively communicate their use of renewable energy in data centers. Whether these promises are fully kept and how the entire supply chain is assessed is difficult to verify from the outside. It is a criterion that can be factored into tool selection, but it is not a free pass.