The footprint of AI tools and digital content
What the software a writer uses and the digital content a writer produces cost in energy, water and emissions, and how to use and design both with less, is the question this leaf answers. The rest of a writer's own footprint is in cutting the footprint of your own working practice, and the wider ethics of the tools in ethics of artificial intelligence in research and medical writing.
Count the running cost of the tools as well as the travel
The energy and water a writer's software consumes is a newer line in a writer's footprint account, and it was put sharply at EMWA's May 2024 symposium in Valencia. Mike, a researcher in the public policy programme at the Alan Turing Institute, was demonstrating a general chat assistant failing a simple reasoning question and named what the failure was costing while it ran: a system of that kind consumes a city's worth of electricity, and a city's worth of water to cool its servers. His argument is not that the tools should not be used but that this is one of the AI industry's ecosystem effects and has to be weighed alongside the labour conditions of the people who train and annotate these systems and the concentration of the most capable systems in a handful of firms — and weighed each time, rather than settled once by noting that the application in front of you is useful. What a professional judgement has to take in is the benefits, the costs, and how both are distributed. The record is Vimeo's uncorrected automatic captions. Ethics of artificial intelligence in research and medical writing carries the rest of that argument.
Azuka Iwobi, a medical writer and disclosure manager, and Rienne Schinner, a library services specialist who sits on an AI evaluation task force, took the question up in the Crofter column of EMWA's December 2025 issue, prompted by the sustainability group's expert seminars on healthcare's carbon footprint. Generative AI's energy use is hard to quantify: ChatGPT processes about 2.5 billion prompts a day, and a generative search and response may use up to ten times the energy of a standard web search. The demand comes from computation, which needs a steady electricity supply; from cooling, since data centres need drinking-quality water, straining scarce reservoirs in some areas; and from hardware, a 2-kilogram computer needing an estimated 800 kilograms of raw materials and AI's microchips needing rare earth elements often mined destructively. Their own exchange with ChatGPT, printed as a figure, has the model agreeing that its footprint is a serious question. The December 2025 issue of Medical Writing (volume 34, number 4, on safety writing) drawn on here was admitted from an existing Markdown conversion without content repair; its known limitations are broken ligatures, mangled multi-column tables, interleaved page headers and footers, and figure text represented as marked text.
The easy answer — keep AI for complex queries and use a plain web search for simple ones — had been undercut by search engines that now return AI overviews automatically and by assistants embedded in applications such as WhatsApp, so they turn to a four-part approach to green AI that Antoine Gaget, a technology leader at Capgemini, proposed in a podcast: quantification, reduction, awareness and standardisation. Quantification lacks standard methods and should cover the whole lifecycle, from mining the materials for a chip through training to retirement; researchers estimating the emissions of 79 well-known systems between 2020 and 2024 from the energy use of their graphics processing units found the top 20 consumed as much energy as a small country such as Iceland or North Korea, and training duration and data-centre location matter too, since centres are often built where non-renewable, on-demand power such as coal is available and so cluster in carbon-intensive regions. Water adds about 2 litres for cooling per kilowatt-hour a data centre consumes, and small tasks add up: an MIT Technology Review estimate puts 28 prompts for a fundraiser — 15 for background, 10 for an image and 3 for a video — at 2.9 kWh, about three and a half hours of a microwave. OpenAI's GPT-5 was reported to use up to 40 watt-hours for a response of about 1,000 tokens, which against 2.5 billion requests a day could reach the daily electricity demand of 1.5 million US homes, though exact comparisons between models were not yet possible.
Use generative AI only where it adds value
Awareness and standards are the slower levers: calls for transparency from providers, including a 2024 US Senate bill asking the Environmental Protection Agency to study AI's environmental impacts and create a reporting system; work at the MIT Lincoln Laboratory on power-capping graphics processors, on workload scheduling and on a climate-aware computer vision tool that switches to more efficient models when carbon intensity peaks; and proposals for a carbon footprint label like those on flights, industry norms and certification, energy audits that lead to action, and renewable power and efficient equipment for data centres. Reduction includes renewable power, modern servers and workstations, and models hosted on an organisation's internal server rather than an external large language model.
Their advice to writers is concrete. Use a standard search engine for simple queries, such as finding background for an article, and a language model for reviewing a draft; use AI only when it adds value, and edit by hand rather than re-running prompts for perfect wording on minor edits, which also keeps editing skills sharp; keep a prompt library of tested prompts for reuse across projects and teams; consolidate requests into one well-framed prompt that specifies tone, audience, structure and length rather than prompt-spamming follow-ups; run tools locally where possible for spelling, readability and similar checks, such as Word plug-ins; choose platforms with sustainability commitments, and check them rather than taking their word; speak up and propose sustainability by design in editorial workflows; and model responsible use by saying when AI was used and avoiding AI-generated content where critical judgement is required. Their closing line is that every prompt counts.
Count the carbon in the content itself
Alisa Bonsignore of Clarifying Complex Ideas made the content a writer produces the ethical question in the Winter 2022 AMWA Journal's Everyday Ethics column. Climate change is expected to harm the health of all humans — more intense heatwaves, flooding, damaging storms and changing patterns of infectious disease — with the most severe consequences falling first on women, people experiencing poverty, and Black, Indigenous and people of colour; the 2022 IPCC summary for policymakers says that near-term action limiting warming to close to 1.5°C would substantially reduce losses and damages. Since, in Gerry McGovern's phrase, digital is physical, and every bit and byte is energy that for the foreseeable future emits greenhouse gases, she argues that writers have an ethical responsibility to weigh those emissions in content strategy, content design and content governance. The Winter 2022 AMWA Journal issue (volume 37, number 4) was converted from its PDF by layout-aware Poppler XML conversion and admitted without content repair, in which figure labels can be inlined into prose, cover and contents columns can interleave, and reference URLs can be malformed or truncated. She points the effort at work rather than home: a 2017 report by CDP attributes more than half of all greenhouse gas emissions since the industrial revolution, and 71% since 1998, to 100 companies, so the most leverage a writer has is through the organisations they work for.
Her arithmetic starts from a report by the American Council for an Energy-Efficient Economy that each gigabyte of data requires 5.12 kWh of energy, which the US Environmental Protection Agency's greenhouse gas equivalencies calculator turns into about 5 pounds of emissions per gigabyte — a figure that moves as the calculator is updated, the United States having generated about 20% of its energy from renewable sources in 2021. Her three examples are the transferable part. A website: the average home page weighed close to 2 MB in 2020 and the ninetieth percentile more than 7 MB, and every one of eight health organisations she sampled with the Pingdom page-weight calculator was well above the average; roughly half a modern page's weight is imagery, often stock art that adds emissions without informing, which she sets against Balmford's view of plain language as the whole document — language, structure and design. Replacing a 4.9 MB home page's unnecessary stock images of providers in scrubs and masks with informative product photographs halved its weight and, by her calculation, avoided nearly 17 tons of emissions, the equivalent of taking more than three passenger cars off the road for a year. Video against audio: a half-hour recording with slides and a talking head at 24 frames per second comes to about 1.73 GB, the same content as a monochannel talk-only podcast about 23 MB, so decide case by case whether a video is worth 75 times the energy. Meetings: with everyone's video on, a call uses roughly 800 to 900 MB per person; her own month of calls averaged five participants and about 4 GB a call with video on, against about 190 MB with screen sharing and only the speaker's thumbnail visible, which she puts at 20 times the energy in her text and 21 times in her figure. Her compromise is video on for the first five minutes of greeting and off during screen sharing. The tools she names are the Pingdom website speed test and a video file-size calculator.