Dealing with the environmental impact of AI on advertising production
How should adland grapple with the environmental toll of AI in advertising production? We asked AdGreen’s Tom O’Brien
For quite some time I was convinced my ChatGPT-obsessed dad was personally draining Olympic-sized swimming pools every time he opened ChatGPT.
While there is some truth in this - I’m convinced he is in the top 1% of global users - the reality is more complicated. There is still no universally agreed figure, but estimates suggest a single text prompt uses somewhere between a few drops and a few teaspoons of water.
That does not mean discourse about environmental concerns like mine are overblown. Far from it. The challenge is that the conversation is often focused on the wrong scale.
For advertising production specifically, the bigger environmental concerns are not really about somebody occasionally asking ChatGPT to summarise a meeting transcript.
They sit in what happens when AI becomes embedded across production workflows at industrial scale, from image generation and synthetic video creation to hyperpersonalised versioning, automated asset production and the sheer volume of content these tools now make possible.
Adland has spent years trying to reduce the environmental impact of production, from cutting flights and reducing waste on shoots to measuring supply chains and building sustainability frameworks around creative work.
Generative AI has now entered that conversation too, bringing both opportunities and new complexities.
A recent survey from Spark AI found 13% of agency respondents raised environmental concerns about AI entirely unprompted, asking questions around energy consumption, water usage, data centre infrastructure and whether AI usage aligns with their company’s sustainability goals.
At the same time, the research also attempted to inject some perspective into a conversation that can often become highly polarised. According to Spark AI, using ChatGPT 50 times a day for the rest of your life would still emit less CO2 than a single one-hour commercial flight, while eight prompts a day for an entire year uses roughly the same amount of energy as running a space heater for two hours total.
The findings point towards something the industry is increasingly trying to understand: the environmental story around AI production is far more nuanced than the internet discourse sometimes suggests.
On an individual level, the impact of a single prompt can be relatively small. The larger issue sits at infrastructure and production scale. Spark AI estimates consumer-facing chatbots account for around 3% of AI’s total energy consumption, with the remaining 97% embedded inside platforms and services people already use every day, including digital advertising systems, recommendation engines, streaming platforms, social media infrastructure and content personalisation tools.
In other words, the environmental impact of AI production is not simply about individuals generating a few images in Midjourney. It is tied into a much larger production ecosystem that advertising increasingly operates within.
That wider complexity is something organisations like AdGreen are now trying to navigate.
AdGreen has spent the past several years helping the advertising production industry measure and reduce production-related emissions associated with producing content. Its carbon calculator allows production teams to input data across 11 activity areas including transport, accommodation, catering, data storage, post-production and now AI usage.
Tom O'Brien
Carbon Calculator Executive
AdGreen
Tom O’Brien, Carbon Calculator Executive at AdGreen, says the goal is not to sensationalise AI’s impact, but to start building a clearer picture of how the technology is actually being used across production workflows.
“AdGreen has been going for five years now, and we support the advertising production industry in measuring and reducing production-related emissions.”
“My role is supporting the development and updating of the calculator. Users input data relating to their production and then we deliver back a carbon figure.”
The organisation introduced its AI Usage section around six months ago. Since launch, users have input information about 32,000 text responses, 80,000 images and nearly 400,000 AI-generated videos into the calculator.
“What we’ve discovered is that, when it comes to the impact per individual AI content generation, the CO2e is actually much lower than people might expect,” O’Brien says.
But that does not necessarily mean AI production’s overall environmental impact is small. The scale of usage and the wider infrastructure picture still matter enormously.
That said, video generation stands out as the biggest contributor by far.
"Since launch, 56 tonnes of CO2 have been added to the calculator through video generation,” he says. “For perspective, a business return flight from London to Cape Town is 7.953 tonnes [BEIS 2025].”
Even so, O’Brien is careful not to overstate what the data currently shows. One of the biggest themes throughout AdGreen’s research is that the industry is still in the very early stages of understanding how to properly measure AI usage.
“The one thing we have to be quite hesitant with when it comes to our data is that it’s still early days,” he says. “We can only report on what we know.”
Right now, the biggest limitation is simply visibility.
“The main limitation for us with AI is data collection: people knowing how much AI content they’re actually creating. People are working really hard to implement systems to measure this, but it’s still really difficult to capture.”
That challenge is partly because AI is increasingly blending into existing workflows rather than sitting separately from them. Agencies are also not necessarily replacing traditional production entirely with AI-generated work. More often, the technology is being used selectively alongside existing production methods.
O’Brien says this “hybrid” approach is where AI may genuinely help reduce emissions in certain scenarios.
"If an advert is primarily being shot where the creative team are based, which we always suggest for reducing emissions because shooting locally is generally the best way to keep travel low, but there’s one shot that would add significant creative value and would otherwise require a flight, then doing that part in AI could significantly reduce the total emissions of the production,” he explains.
“So it’s about using it selectively and using it with an understanding of how other elements impact your carbon emissions.”
There are already examples of this emerging in practice. O’Brien points to brands experimenting with hybrid production approaches where AI is used selectively rather than replacing production entirely.
One example discussed within AdGreen’s wider sustainability conversations was Italian food brand Terre di Sanvito, which recreated a live-action commercial frame-for-frame using AI. According to the figures presented, the original production generated 2.18 tonnes of CO2 compared to 0.13 tonnes for the AI recreation.
For O’Brien, however, examples like this also highlight the industry’s bigger dilemma. If AI dramatically lowers the cost and friction involved in producing content, agencies may also be tempted to create significantly more of it.
That measured approach appears to reflect how much of the industry is currently approaching AI adoption. Speaking recently with production companies and agency leaders, O’Brien says many businesses are already developing dedicated AI production capabilities while still maintaining traditional production models alongside them.
“We’re seeing agencies using it complementarily,” he says. “Some clients want fully AI-generated work, some want hybrid approaches. In some cases, production companies are actually asking: ‘Why do you want this done in AI? Is that actually the best use case?’”
Rather than framing AI as inherently positive or negative environmentally, AdGreen’s position is largely centred around responsible implementation and transparency.
Part of the nuance here comes from the fact that AdGreen’s calculator only measures a specific part of the AI process.
“We focus on inference, the energy used at the moment a prompt generates an answer,” O’Brien says.
The organisation does not currently include emissions linked to model training or wider infrastructure construction within individual prompt calculations. However, this is a deliberate methodological choice designed to keep AI measurement consistent with the rest of the calculator.
“We measure the fuel used in a car journey, but we don’t measure the emissions involved in building the car itself,” O’Brien explains. “Similarly, with AI, we don’t include model training emissions as part of the emissions associated with a prompt.”
Even within that narrower scope, AdGreen is aware (and has also highlighted in a recent report) that AI’s environmental impact extends beyond carbon alone.
The company is aware of the need to explore additional pressures linked to data centre infrastructure including water demand, energy usage and air pollution, alongside broader ethical and creative questions surrounding AI adoption within advertising such as copyright and job retention.
“Because AI is such an emerging and multifaceted technology, we felt it was important to provide a broader, more holistic understanding of the different impacts involved,” he says.
Still, one of the most interesting points emerging from AdGreen’s work is that the environmental concern may not necessarily be individual prompts themselves, but the possibility of excessive content generation at scale.
“We really want to emphasise the idea of responsible prompting and responsible usage,” O’Brien says. “From an environmental perspective, the real danger lies in digital waste, especially video waste.”
Because AI tools dramatically reduce the friction involved in generating content, there is naturally a risk of producing far more material than is ultimately needed.
“These tools make it very easy to generate huge amounts of content. You can end up leaving loads on the cutting room floor. As things become faster and more efficient, there’s a real possibility of overproduction.”
For agencies, that does not necessarily mean avoiding AI altogether. In many cases, O’Brien believes the answer is actually becoming more intentional about how these tools are used.
“The more detailed and intentional the prompt is, the more likely it is to deliver the result you actually want,” he says. “When it comes to video and image generation, text generation itself has a very minimal CO2 impact. So you’re often better off using AI text tools to help craft prompts first, then applying those to image or video generation.”
That process, he argues, can reduce unnecessary generation cycles and improve efficiency overall.
Alongside prompting practices, AdGreen also believes agencies should start formalising governance structures around AI usage more broadly.
“I think agencies should be looking at building proper AI policies,” O’Brien says. “That should include upskilling employees on how to use AI effectively, as well as putting processes in place for recording AI usage.”
Many large agency groups are already building proprietary systems and bespoke AI tools internally. The challenge now is ensuring those systems are designed in ways that allow the amount of AI-generation used in creating the work to be tracked properly.
“What’s important is that those systems are designed in a way that creates a digital trace for every asset created, so usage can actually be tracked,” he says.
This is not purely about sustainability reporting either. O’Brien believes agencies that establish clear tracking systems now will likely be in a stronger position later as regulation and reporting expectations continue to evolve.
“I think it will ultimately get wrapped into scope 3 emissions reporting,” he says. “Mandatory reporting has been coming for a while, and AI will eventually become part of that conversation.”
Without those systems in place, agencies could eventually face the difficult task of retrospectively untangling years of AI-generated production work.
Despite the ongoing challenges around measurement, O’Brien remains relatively optimistic about how the industry is approaching the issue overall.
“I definitely don’t think it will undo all the hard work around reducing emissions,” he says. “AI will become another tool in the arsenal.”
Ultimately, AdGreen’s view is not that AI should be feared or avoided, but that it should be approached thoughtfully, transparently and with clear creative purpose.
“Being human-first and creatively led - seeing AI as an effective tool, understanding when to use it and when not to - will separate the agencies that succeed with AI from those that don’t,” O’Brien says.
So no, my dad probably is not personally wasting Olympic-swimming-pool amounts of water every time he asks ChatGPT to predict a West Ham lineup.
Built with Shorthand