Love the 80s AI selfie trend? Here’s the environmental cost of those retro portraits

Every AI-generated image requires electricity and water, and researchers warn that the real environmental impact lies in the billions of images, videos and prompts generated at scale

People are feeding their selfies into ChatGPT, Gemini and other AI tools and turning modern photographs into retro portraits that look as though they have been pulled from an old family album. Big hair, pastel colours, dramatic studio lighting, denim and deliberately awkward poses have become a viral visual trend.

Creating one takes only a few seconds. But behind that apparently effortless transformation is a considerable amount of computing infrastructure — and therefore electricity, water and hardware.

An individual AI-generated image has a relatively small environmental footprint. The bigger concern is scale: people rarely generate just one image. They try different versions, change outfits and hairstyles, increase resolution and sometimes move from still images to AI-generated video.

Research from Carnegie Mellon University and Hugging Face, cited by Firstpost, estimated that generating an AI image can consume between 0.01 and 0.29 kilowatt-hours (kWh) of electricity, depending on the system and generation process.

The range is wide because different AI models and hardware have different energy requirements. The complexity of the request and the resolution of the final image can also affect consumption.

And image generation is only one part of AI’s growing energy demand.

According to the United Nations University, inference — the everyday process of using an AI model after it has been trained — accounts for roughly 80% to 90% of its total energy demand.

AI use has expanded rapidly from an occasional experiment into a daily activity for millions of people. A platform such as ChatGPT is estimated to process around 2.5 billion prompts a day, consuming hundreds of gigawatt-hours of electricity each year.

The energy required also varies sharply depending on what users ask AI to do.

A typical conversational AI query can use around 200 times more energy than basic text classification. Generating a single AI image can require around 1,450 times that baseline, while a short AI-generated video can consume as much electricity as approximately 200,000 spam-classification tasks, according to the UN.

That mns the seemingly simple retro selfie can require considerably more computing power than an ordinary text-based AI request.

The efficiency paradox

Several factors determine the environmental footprint of an AI generation, including the model being used, prompt length, output format and image resolution.

Most users never see these calculations. They upload a photograph, type a prompt and receive an image within seconds, while the computing takes place inside data centres.

As AI models become faster and cheaper to run, however, their efficiency could encourage people to use them more frequently. This is known as the rebound effect.

If generating one image becomes cheaper and more efficient, users may simply generate ten instead of one.

A viral trend can magnify that effect, turning millions of individual requests into a substantial demand for computing power.

AI’s hidden water footprint

Electricity is not the only resource involved.

AI data centres require water for cooling, while additional water can be consumed indirectly through electricity generation and the manufacture of semiconductors used in computing equipment.

A study published in Water Research, titled The water footprint of artificial intelligence: Emerging solutions and governance imperatives, estimates that the global water footprint of AI could reach between 4.2 billion and 6.6 billion cubic metres a year by 2027.

Where data centres are located also matters. The study says around two-thirds of data centres built after 2022 are in water-stressed regions.

The United Nations University estimates that the electricity required to generate a typical AI image is equivalent to powering a 10-watt LED bulb for around 17 minutes.

For a high-complexity AI-generated video, the equivalent energy could power the same bulb for roughly 42 hours.

The electricity-associated water footprint of one typical AI image is estimated at around 29 millilitres — roughly two tablespoons.

That does not mean 29 millilitres of water is directly poured into an AI system. The figure represents the water associated with producing the electricity required for the task.

For a complex AI-generated video, the estimated figure rises to around 4.1 litres.

Again, the environmental impact of a single image is relatively small. The concern is what happens when that consumption is multiplied across billions of requests.

The bigger data-centre problem

The environmental cost of AI becomes clearer when viewed through the infrastructure required to support billions of requests.

Data centres consumed around 448 terawatt-hours (TWh) of electricity globally last year, with AI accounting for roughly one-fifth of that consumption, according to figures cited by Reuters.

Those data centres also consumed around 4.5 trillion litres of water and generated approximately 189 million tonnes of carbon dioxide emissions.

The demand is expected to increase significantly.

By 2030, annual electricity consumption from data centres could almost double to around 945 TWh, with AI accounting for about 40% of the total. Water consumption could rise to 9.3 trillion litres, while associated CO₂ emissions could reach around 399 million tonnes, according to the figures cited by Reuters.

There is also a growing electronic-waste problem.

The United Nations University estimates that AI infrastructure could generate as much as 2.5 million tonnes of electronic waste every year by 2030.

The physical footprint of data centres is also expected to expand. Land used by data centres could increase from around 6,900 sq km to more than 14,500 sq km by 2030, according to the research.

The environmental story, therefore, is not really about one AI-generated picture. It is about the enormous infrastructure being built to make billions of images, conversations and videos possible.

India’s data-centre dilemma

The issue is already being debated in India.

Google’s planned $15 billion data-centre hub in Andhra Pradesh has faced opposition from environmentalists concerned about water supplies and wildlife, Reuters reported in August.

The project is being developed near a wildlife sanctuary, with activists raising concerns about its potential impact on Visakhapatnam’s already strained water resources.

The Andhra Pradesh government has denied allegations that the project was fast-tracked without properly assessing its potential impact on water supplies and wildlife. The project has also faced legal challenges relating to water use and its proximity to the sanctuary.

The dispute illustrates how the environmental costs of AI infrastructure are unevenly distributed.

A user sitting at home may see only a loading screen and a finished photograph. The electricity and water required to produce it may come from infrastructure hundreds or thousands of kilometres away — potentially in regions where water is already scarce.

Should you stop generating AI photos?

Not necessarily.

It is difficult to argue that a single AI-generated 80s portrait will have a major environmental impact. Its estimated footprint is small compared with the much larger demands of data centres, semiconductor manufacturing and other industries.

The bigger question is what happens when small amounts of resource consumption are repeated billions of times.

Users often generate several versions of the same image — changing the clothes, hairstyle, lighting or background — before settling on one. A viral trend can multiply that behaviour across millions of people.

AI companies are working to make models and data centres more efficient, but efficiency alone may not reduce total consumption if demand continues to rise.

For users, there is no need to feel guilty about every AI image. But it is worth recognising that the process is not environmentally weightless simply because it happens through a phone or laptop.

Behind that shiny retro portrait are servers, electricity, cooling systems, semiconductor chips, water and a rapidly expanding network of data centres.