For all its magic, AI has a dark side. Dark as in unpleasant: biased models perpetuate stereotypes, hallucinate lies and spread misinformation. These are things we know, things we can partially mitigate. But dark can also mean hidden. Every new AI model release is accompanied by boasts about what it can do. What we’re not told is the cost of how it performs its conjuring tricks – the cost to our planet, and our ability to maintain its precious resources. This needs to change.
AI improves efficiency, automating tasks performed by humans. Robots first replaced unskilled labour performing repetitive tasks but AI is now coming for knowledge workers, including those in the screen industries, potentially replacing skilled professionals with systems that imitate their human-crafted expressions. Paradoxically, while AI is expected to deliver significant efficiency gains, the way its newfangled generative form works is wildly inefficient and resource-intensive.
Vast amounts of computational power and energy go into AI training (the initial step in which models learn from the information they’re fed) and inference (the process by which trained models generate outputs based on the probability that they’re plausibly correct). Training OpenAI’s GPT-4 required an estimated 50 GWh of electricity – enough to supply nearly a quarter of a million UK households for one month. It’s further estimated that every image outputted by an AI model uses as much electricity as fully charging a smartphone. Using AI to answer queries requires 10 times the amount of electricity of classic search. Total electricity consumption by data centres – the near-workerless factories of the AI industrial revolution – is forecast to roughly equal Japan’s annual consumption next year.
Data centres don’t just consume huge amounts of electricity (nearly a third of Ireland’s total demand by 2026, according to studies), they also have a gigantic thirst since they evaporate water to keep cool. Creating a 100-word email using generative AI requires the equivalent of a 500ml bottle of water to prevent servers from overheating. Between now and 2030, data centres could emit 2.5 billion tonnes of greenhouse gases globally – three times the amount if generative AI hadn’t happened.
That’s a lot of stats. The problem is that almost every one of them is now wrong. They were calculated before the recent release of even bigger AI models; before a plethora of energy-sapping video generators went live; before search went generative, spitting out AI-written summaries whether we want them or not; before image generators were integrated within social networks; and before their use went mainstream.
In March, AI-generated images in the iconic style of Japan’s Studio Ghibli went viral thanks to OpenAI’s new image generator, allowing outputs resembling Ghibli’s distinctive hand-painted animations. OpenAI CEO Sam Altman said while it was “super fun” to see people loving Ghibli images, the craze was causing its graphical processing units (GPUs)—the workhorses of the generative revolution that perform all that power-hungry inference—to melt. Altman’s short-term fix was to cap the number of images users could generate to three per day. That at a time when OpenAI’s weekly active users soared to 800 million, doubling since February.
Longer-term fixes to the AI companies’ demand for more energy include calls for deep investment in nuclear power. That sounds like a clean solution, but hardly anyone talks about the spent fuel rods. Much of the radioactive waste is simply buried underground for future generations to deal with. Elon Musk’s ginormous AI data centre in Memphis is powered by portable methane gas turbines. In April, President Trump signed an executive order saying new coal-powered generators were needed to support America’s AI developers. Yes, coal. Sound sustainable to you? Me neither.
What can we do? Here are three things. Our industry should commit itself to only using AI when it’s absolutely necessary. Think of the number of discarded AI images and video sequences and the wasted energy that went into producing those warping, morphing, physics-defying clips that never make it to the edit. We should extend industry initiatives such as BAFTA’s albert from calculating the carbon footprint of TV productions to include AI, making clear how much AI material (including discarded content) was generated and how much energy it consumed. And we should lobby for AI companies to be totally transparent about their energy usage, adding consumption labels to models and outputs similar to those slapped on electrical appliances in the UK and across Europe.
In their punk era hit Something Better Change The Stranglers said we were “too blind” to see what’s “happening right now”. Decades later we’re largely oblivious to the environmental damage being caused by AI, and the devastation that unchecked AI is bound to leave future generations. “Ain’t got time to wait,” they sang. And we don’t. Something better change.
This article is from TVBEurope’s May/June issue, which can be downloaded for free here.
