We worry about AI's environmental impact. What about the hours we spend scrolling?

AI deserves scrutiny, but it is not the only part of our online lives with a physical footprint.

People are beginning to ask sensible questions about the environmental cost of AI. Should we really use a powerful model for a throwaway question, and does every generated image need to exist? Behind the convenience of an instant answer sits a question people are only starting to ask.

Those questions matter. But they can make AI feel like the only part of our online lives with a physical footprint.

Meanwhile, millions of people spend hours on platforms designed to hold their attention for as long as possible. Something grabs your attention on TikTok, Instagram Reels, Facebook, LinkedIn or YouTube Shorts, and autoplay, personalised advertising and refresh after refresh keep it there. Search, streaming, social media and AI all rely on energy, data centres, networks, devices and cooling. The difference is not that one is clean and one is not. It is that we have started paying attention to one form of digital consumption while barely questioning the rest.

This is a case for being more intentional about the tools and platforms we use, and what they ask of us in return.

AI deserves the scrutiny

Training and running AI models uses real electricity, and the IEA expects data-centre power demand to more than double by 2030. That concern is justified: specialised chips, cooling and large data-centre infrastructure all draw on the grid. The International Energy Agency projects that global electricity generation to supply data centres will rise from 460 TWh in 2024 to more than 1,000 TWh by 2030 and 1,300 TWh by 2035. AI is a major part of that growth, even if it is not the whole story.¹

There is a useful distinction here. A long, complex prompt to a large reasoning model, a generated video, and a short factual search are not equivalent tasks. Their energy use will differ according to the model, the length of the request and answer, the hardware, the data centre and the electricity available at that moment.

That is precisely why blanket claims about "an AI prompt" can mislead. Most providers do not publish enough comparable data to calculate a reliable footprint for every task. We should demand better transparency, rather than filling the gap with a single viral figure that gets repeated long after it has stopped being useful.

Uncertainty is not an excuse to ignore the direction of travel. AI is expanding quickly, and the infrastructure behind it is very real.

The rest of the internet is physical too

Search, streaming and social media consume energy too, and video is the heaviest of the three. Those digital habits became ordinary long before anyone described them as an environmental issue.

Every search needs servers and network traffic. Every social feed is selected, stored and delivered somewhere. Video is especially data-heavy because it has to keep moving. A 2021 study from researchers at MIT, Purdue and Yale estimated that an hour of high-quality streaming on a service such as Netflix or Hulu could average around 441g of CO2 equivalent.² That is not a like-for-like measure of an hour on TikTok, Reels or Shorts, which differ in format, delivery pattern and the data centres behind them, but it is a useful reminder that video consumption is not weightless. The recommendation systems behind short-form platforms do not simply show what a person chose to look for. They continually test what might hold their attention next, then serve another clip, another advert and another set of data.

At individual level, it is tempting to ask which activity is "worse." One ChatGPT question or ten minutes on TikTok? A Google search or a Reel? There is no robust universal answer. The data is incomplete, tasks vary, and companies report their emissions at very different levels of detail.

At system level, though, the question is harder to avoid. Digital services are designed around scale and repetition. One person staying on a feed for an extra twenty minutes may seem trivial. Billions of such sessions are not.

This is where the conversation about consumption gets more useful. It is about recognising the difference between a tool that helps someone complete a task and a product designed to turn spare minutes into more consumption.

A better question than "should I use AI?"

For most people, opting out of digital life altogether is neither realistic nor especially helpful. The more practical question is: what is the lightest tool that can do this job well?

Sometimes that will be a conventional search, a source document or asking a person who already knows the answer. Sometimes a capable AI model will be the sensible choice because it saves a meaningful amount of time, turns a difficult document into something accessible, or helps solve a problem that would otherwise take several steps. And sometimes the answer is that we do not need to generate anything at all.

That is why Ecosia's expansion into AI search is interesting. It sits in a deliberate middle ground for the everyday questions that fall between a list of links and a full AI workflow, well short of solving AI's environmental cost and nowhere near replacing ChatGPT or Claude for the jobs that actually need them.

Where Ecosia may fit

Ecosia now offers AI Overviews in search results and an AI Chat feature in the UK. The chat is currently powered by Mistral Small 4, while its AI Overviews use Mistral Small 3.2. The company says it has chosen smaller models to balance usefulness with energy and water efficiency, and avoids some particularly energy-intensive features, such as video generation.³

That makes it a plausible option for lighter work:

• Basic questions where a short answer and sources are enough
• Everyday browsing and finding things online
• Simple comparisons, explanations and planning
• A first pass on sustainability-related research, followed by checking the original sources
• People who want a search default that reflects their values more closely

Ecosia also allows users to switch off AI features and use an AI-free search experience, currently available on desktop with mobile support planned. That option matters. A responsible product should not treat more AI as automatically better, or make opting out unnecessarily difficult.

For people trying to reduce needless use of larger models, this is the useful proposition: use a smaller, more bounded tool when that is all the task needs.

Where the fit breaks down

A smaller model like Ecosia's is not built for deep reasoning, complex strategy or extended research. There is no benefit in pretending a smaller search chat tool is best at everything.

For work at that level, advanced writing, specialist analysis, coding, or substantial document review, ChatGPT, Claude and other specialist tools may still be stronger. Their capability can be justified when the work genuinely needs it.

The goal here is genuine fit between tool and task. Using a weaker tool three times, then switching anyway, is unlikely to be an environmental or practical victory, and replacing well-considered work with an inadequate answer simply because it looks lighter on paper serves nobody.

Good digital judgement means matching the tool to the job. It also means being honest about the value of the job itself.

Be careful with "green AI"

Ecosia's environmental claims are accurate, but they leave out the emissions from training the model underneath. It says it generates more renewable energy from its solar and wind projects than its AI features consume, and that it directs its profits towards climate action.³

That last qualification is important. Renewable generation can support the wider transition away from fossil fuels, but it does not mean each individual AI response runs directly on clean power. It does not remove the environmental cost of chips, data-centre construction, water use or the model's training. It also does not resolve the wider problem of demand growing faster than efficiency gains.

This standard applies to every technology company making climate claims, Ecosia included.

The responsible language is therefore "lower impact by design" or "a more considered option," where the evidence supports it. Avoid "green AI," "carbon-free" or any suggestion that a product can make digital consumption consequence-free.

Better defaults

The burden should not sit entirely with individuals. Platforms decide whether video autoplays, whether a feed ends, whether notifications pull people back in, and how much environmental information they disclose. Technology companies need to publish clearer data on electricity, water, hardware and emissions, in forms that make meaningful comparison possible.

Better defaults from platforms would help more than any individual habit change. Until they arrive, though, there is still plenty within reach.

Five places to start

None of this requires purity. It requires noticing the defaults you have never questioned, and deciding whether they still serve you.

1. Do you need to sit on TikTok or Instagram Reels for two hours, or would twenty minutes still scratch the itch?
2. Switch off autoplay on YouTube, Netflix and Instagram, so the next thing has to earn your attention rather than arrive automatically.
3. Turn off the notifications that pull you back to a feed you would not otherwise have chosen to open.
4. Try a search or the source document first, and save a powerful AI model for the question that actually needs it.
5. Before you generate an image or a video, ask whether you actually need it to exist, rather than joining the latest viral trend simply because everyone else has, whether that is turning your photo into an action-figure doll or chasing whatever stylised filter is having its moment.

Ecosia points towards a more useful habit: choosing technology that is proportionate to the task, and refusing the idea that more consumption is always progress.

---

A note on how this was made: research and fact-checking used web search against primary sources (IEA, Ecosia's help centre, MIT Energy Initiative), and Claude for drafting support. The framing, the audience, the editorial judgement, and every significant decision along the way were mine.

Footnotes

1. International Energy Agency, Energy supply for AI, 2025. Figure refers to electricity generation to supply data centres; the IEA also publishes a separate consumption-based figure (415 TWh in 2024, rising to 945 TWh by 2030) elsewhere in the same report. Forecasts should be checked against the latest IEA release before publication. ↩

2. Purdue University, Yale University and MIT, cited via MIT Energy Initiative, 2021. Figure is specific to high-quality Netflix/Hulu-style streaming and should not be treated as a direct proxy for short-form video platforms. ↩

3. Ecosia, Generative AI on Ecosia Search, last updated 21 June 2026. Product features and sustainability claims should be rechecked immediately before publication. ↩ ↩²