We’re all aware of the conversations around AI’s energy use and its environmental impact.
We can’t look away from it, especially as utilisation is expected to soar in the coming years. The race towards greater productivity should not come at the cost of lasting environmental and resource degradation.
How then can we as individuals responsibly use AI, should we decide to use it?
In thinking about this to guide my own technology use vis-a-vis resource use and avoiding waste, I keep coming back to a trailblazing example of automation we’re all familiar with: the washing machine. Even if we’re not paying per load, we intuitively know the machine uses a lot (and can see it in the bill at month’s end).
On average, an modern washing machine use around 40 liters of water and 300 watts per load (~1 hours use), and it is an everyday part of daily life. It’s also been around long enough that we’ve developed a common sense of how to use it right. So although it’s level of consumption is magnitudes more than an hour’s use of an LLM, the aggregate amount of people asking for an email summary adds up. Which brings me to the following:
“Laundry Rules for AI Use”
Choose: Before buying a machine, we look at its energy and efficiency ratings. We look for candidates that can handle our regular washing volume and grubbiness without using excessive water and power. Same goes for detergents, what kind of cleaning power do we need? Is it safe for sensitive skin? Is it environmentally toxic?
A couple of things to unpack here, and it begins with not accepting any service by default, even if its baked into your device. However, we don’t presently have energy rating stickers for AI services (or social media services like this one for that matter) that helps us compare resource use between services. It takes some effort to find and understand what their environmental impact may be. That’s just for one dimension of use; there’s also security, data privacy, local/cloud-hosted, limitations (or lack off) to consider, among others, for a service that meets the balance of utility and impact. It’s a lot of work even for IT professionals, and it’s definitely a challenge for regular people at this point. But – it’s early days yet and as more people consider beyond the long-press shortcuts on their phones, it will get easier.
Batch: Sure, we can use the washing machine everyday, but we don’t. Washing a pair of socks and gym clothes immediately after wearing them is clearly wasteful. We wait until there’s a sizeable amount of soiled clothes to wash.
Similarly, that great idea for a post in your head? Hold that thought and let it stew for a bit. There is no fire, usually. As commercial services, popular consumer AI applications are starting to feel eerishly similar to social media. Ask one thing, and it generates a response, then helpfully asks if you would like it to do something more. And it continues, that before you realize it you’re out of tokens from making detailed charts of the different types of migratory birds from a single prompt about a bird you saw earlier. We’ve learnt from social media and streaming that companies focus on developing user stickiness to maximise time spent on a service which expand monetization opportunities like ads and data collection. AI companies may be new but the commercial logic isn’t – if you’re not paying for something, you’re the product. Being intentional about what and when we use these services doesn’t just give us more utility from having clear objectives, but also controls how our attention is exploited for profits.
Adjust: As detergent ads constantly remind us, kids get their clothes really dirty and you need the strength of a full cap of detergent liquid. If clothes aren’t badly stained, muddied, or stank, you could use less and wash in eco mode to use less power and water. Oftentimes, that’s enough to get the job done.
This is about the ‘how’ of using AI services. The empty dialogue box is a design choice that veils the myriad of settings available. They are not hidden but this minimalist approach to user interaction requires individuals to read up on available features and settings. For example, better knowledge of model and effort level selection lets us adjust the method of processing to our needs which can save resources, by using simpler models for easy tasks, and also time, when we use complex ones for more difficult scenarios.
Online user reviews and service FAQs help lift the veil to comprehending their settings and features, just as an appliance’s manual does.
Filter: Some things are too bulky to wash manually. Most casual clothes can be machine-washed more efficiently than hand-washing them. Some fabrics need direct care and attention. Some stains need pre-cleaning. We need to know the difference to decide what to automate and what to ‘human’-ate, and when, to prevent irreversible damage and waste.
Let’s be honest, AI is a seductive shortcut. Minimal effort is needed to generate output that is probably good enough 80% of the time. In some cases, it may make sense: transcriptions, meeting summaries, microedits, shopping reminders, these tedious tasks that would otherwise take hours. For tasks that require genuine human attention and input, it probably isn’t, and maybe more harmful than helpful in the long run in ways we are only discovering now.
Pause before we prompt. Do we really benefit from a generated summary over reading a five page document? Should we just call and make the dinner reservation ourselves rather than leave it to an AI assistant? Can we at least sketch out an outline before asking a model to make something? Skills (including thinking) need practice: if we don’t use it, we lose it. More delegation doesn’t just use more resources, it also creates more reliance.
Pull the plug: Once the washing’s done, switch it off. There’s no need to run a post-wash rinse on an empty tub unless it needs cleaning, or to keep it standing by until the next laundry day.
As these services increase in capabilities, it will keep running 24/7. We are already seeing this: email services suggest phrases after typing in two words, factory-loaded camera apps add generated details automatically, door cameras constantly observing, music services curate multiple personal playlists, phones immediately screen calls; impressive since it’s only been four years since ChatGPT launched. Some automatic features currently require activating, but default creep is persistent. It’s not difficult to imagine a digital environment where an AI-free interaction is the exception rather than then norm.
Intention, again, is important, alongside awareness of the tools in use. Scanning through an application’s or devices features and settings provides an impression of how much it can/is doing even when we don’t truly understand what they are. More importantly, knowing allows us to learn about them and pick the ones we need, and turn off those we don’t. We can always turn them back on when they’re needed.
On a philosophical note, by being always-on, companies try to create a frictionless experience for their products. It’s life-optimization by default, but that level of enhancement can’t be good. Friction is a signal from our effort and experience that we can learn from: it’s feedback and guidance that’s not mediated by external interests.
The absence of friction, especially without us realizing, disrupts that feedback loop. We learn less about ourselves and our world when everything is automatically optimized, and are less prepared for dealing with friction ourselves.
We essentially trade competence for convenience in that world; would that be worth it?
~ Think that’s as far as I can take this metaphor without suds washing over. There are many dimensions to consider (and more beyond what is here), and it is certainly daunting to figure out individually. However, there is growing public consciousness which has begun to influence government and company policies, something we ought to welcome. Anything that lightens the load of making responsible choices by the individual will improve outcomes across the board.
So, now that I’ve listed the checklist@essay, do you think this was AI-written? If it wasn’t, should it have been?
Edited August 17 to illustrate principles beyond the metaphor.