The case for responsible AI

Part 1: The state of AI

Part 2: The costs of AI

Part 3: The benefits of AI

Part 4: A framework for using AI responsibly

Part 5: What can we expect from here?

Part 1: The state of AI

There are people who think AI is going to save us. And there are people who think AI is going to ruin us. There are people who think AI will restructure nearly every part of our lives — that it's the most useful and disruptive technology in human history, and that, despite making millions of jobs obsolete, it will deliver a utopian arrival (at least for the privileged few). And there are people who think AI is an unmitigated environmental and social disaster whose impacts we can only begin to imagine as the technology becomes more and more ubiquitous. There are people who refuse to use it on moral and environmental grounds — and there are people whose social and professional dependence on the tool has already rewired their brains.

I believe there's a middle ground. A way forward that integrates AI without overrelying on it. That we can adopt the technology to make our jobs — and our lives — meaningfully better. That we can use it selectively and intentionally, and reach instead for a less impactful tool — or, radical as it sounds, our own agency, experience, and humaneness — when the situation calls for it.

What's real is that the AI age is here. ChatGPT has nearly a billion weekly active users. Nearly every industry is implementing AI to enhance data processing, automation, and efficiency. AI is inescapable, woven into modern life in ways that are increasingly hard to opt out of. Because of that, we all have choices to make, and those choices matter. This is particularly true for sustainability and impact professionals, whose commitment to responsible practice requires leadership in this moment.

In this five-part Deep Dive, we're going to explore what using AI responsibly actually looks like. We'll establish the current state of AI, look honestly at its costs and its genuine benefits, and build toward a practical framework for intentional use. We'll look at where this is all heading and what role we want to play in shaping it.

AI, of course, is an enormous catchall — encompassing everything from the large language models behind ChatGPT and Claude to application-specific models that detect methane leaks, map deforestation, and predict climate disasters. This series focuses primarily on the generative AI tools most of us encounter daily: the chatbots, image generators, and agentic tools we're increasingly reaching for at work.

A note on how this was written: I'm a writer and editor, and this series reflects my own evolving relationship with AI — personally and professionally. In the interest of full transparency, I used AI as a self-editing tool during the drafting process, to catch errors and pressure-test my own thinking. The words, the arguments, and the perspective are mine.

Part 2: The costs of AI

Right now, AI is incredibly cheap to use. If you only run a few queries a day, ChatGPT is free. Claude Pro costs $20 a month. Both companies are currently losing money, despite OpenAI carrying an estimated valuation of $852 billion and Anthropic reaching nearly $1 trillion following a $65 billion funding round in May 2026. We all know the tech model by now: scale users at all costs, worry about profit later.

But the subscription price only tells part of the story. The environmental costs are astronomical and certainly larger than the companies behind these tools want you to know.

Water is the underreported dimension of AI's footprint. Data centers rely heavily on water-based cooling systems to manage the heat generated by servers running 24/7. A hyperscale data center — the type that runs services like Gmail, Claude, or ChatGPT — uses an average of 550,000 gallons of water per day, roughly 200 million gallons per year. Smaller wholesale data centers still average 18,000 gallons per day. But that’s only the water used on-site. Expanding the view through the supply chain: the thermoelectric power plants that generate much of the electricity powering these data centers withdrew an estimated 48.5 trillion gallons of water for cooling in 2022 alone. Coal plants, which still account for roughly 20% of U.S. electricity generation, withdrew 18.3 trillion gallons of that total. Because AI data centers are disproportionately located in fossil-fuel-heavy grids — northern Virginia, home to more data centers than any other state, runs largely on natural gas — each query carries an indirect water cost layered on top of the direct cooling water the data center itself consumes.

Energy use varies significantly. Standard chat queries are relatively low-impact individually — a complex exchange uses roughly the energy equivalent of running a microwave for 10 to 30 seconds. The picture changes with agentic use, where AI makes multiple sequential calls to complete a task. Reasoning and deep research modes have been found to use up to 43 times more energy than a standard query. A single coding session involving multi-file edits and debugging may involve 20 to 50 API calls, multiplying the footprint proportionally. Video generation sits in a separate category entirely: a five-second AI-generated clip from a mid-tier open-source model requires roughly 3.4 million joules — about 700 times the energy of a high-quality generated image.

* Energy estimates are derived from open-source model measurements (MIT Technology Review / University of Michigan ML.Energy Leaderboard, 2025). Closed-source models like Claude and ChatGPT cannot be directly measured; actual figures may be higher. Long-context drafting sessions involve standard inference but scale with context size and number of queries per session.

The emissions picture is murkier than the industry would suggest — and deliberately so. To offset their reported carbon footprint, major tech companies purchase Renewable Energy Certificates, or RECs. A REC proves that renewable energy was generated somewhere on the grid. It doesn’t prove that renewable energy powered the data center processing your query. When you strip out the RECs and look at actual emissions from the local grid — what researchers call location-based accounting — the numbers look very different. Meta's officially reported Scope 2 emissions for 2022 were 273 metric tons of CO₂. The location-based figure: 3.8 million metric tons, a difference of roughly 19,000 times. Microsoft’s official data center emissions for the same year were 280,782 metric tons. Location-based: 6.1 million. Across Google, Microsoft, Meta, and Apple combined, real emissions from 2020 to 2022 were an estimated 662% higher than officially reported.

The environmental and social impact of data centers go well beyond energy and water use, too. Communities are also raising alarms about noise pollution, drastically increased energy costs, and biodiversity loss.

Part 3: The benefits of AI

To think about AI only in terms of doing the work we typically do, but faster, is to understate its potential. Framing AI as a productivity hack is true as far as it goes, but it misses the bigger picture.

At scale, AI excels at identifying complex patterns, crunching enormous data sets, and optimizing systems in ways that no human team could replicate at speed or volume. For sustainability professionals, that capability has direct and material applications. AI systems are already helping improve the efficiency of electrical grids, cut fuel use in shipping, and detect otherwise invisible leaks of methane. Researchers at the University of Leeds have trained AI to map Antarctic icebergs 10,000 times faster than a human could, tracking meltwater release into the ocean with a precision that was previously impossible. AI-powered systems are monitoring deforestation across more than a million hectares of land in over 30 countries. In São Paulo, companies are using AI to predict where and when climate disasters will occur, giving businesses and governments a window to prepare that they didn't have before.

For sustainability professionals specifically, the implications go beyond what AI can do for the climate broadly. The work of corporate sustainability is often constrained by two things: data quality and analytical capacity. AI improves both. Whether that's developing more accurate emissions inventories, synthesizing regulatory frameworks across jurisdictions, stress-testing climate risk scenarios, or helping a business scale its sustainability program with greater efficiency and ROI, these are meaningful improvements to the quality and rigor of work that the C-suite and customers are increasingly demanding.

None of this makes AI's environmental costs disappear, but the Economist's framing is worth sitting with: the doom-mongering may be misplaced. Data centers currently account for about 1.5% of global electricity consumption and AI, while a fast-growing share, is still a minority of that load. Streaming video, social media, and online shopping collectively drive a significant portion of data center energy demand.

So the question is really, when is AI’s footprint worth it?

Part 4: A framework for using AI responsibly

The framework I return to when addressing when to use AI is simple, if imperfect: AI earns its place when it saves meaningful time. Not a few minutes or seconds. Not as a balm for when I'm feeling lazy or uncreative. But an amount of time that allows me to put my attention and energy toward something more valuable, whether that's an evening with my family or a project that requires my expertise, insights, or empathy — all irreplaceable by AI.

The principle is easy enough to understand and harder to apply consistently, because AI tools are designed to be frictionless and beckon us whenever we open a laptop or pick up a phone, whether we're stuck or simply looking for the path of least resistance. Opening a chat window will always be easier than thinking, than sitting with a problem, or bouncing an idea off a colleague rather than a machine. That means my AI use is really more about discipline and intention than anything else.

The table below can help make that intention more concrete. It covers the most common AI use cases, their estimated energy costs, and a frank assessment of when the trade-off makes sense and when it doesn't.

Before you read it, a few things worth knowing.

The energy estimates are based on open-source model proxies — the major closed-source providers (Claude, ChatGPT, Gemini) don't disclose this data, which is itself a problem. The numbers are approximate, but the order of magnitude is reliable: standard text queries are low-impact, agentic and reasoning modes multiply that footprint significantly, and video generation is in a category of its own.

Think of the "guidance" column not as a policy but a prompt — the question worth asking yourself before you reach for the tool. The goal isn't perfection. We're all learning as we go. It's the habit of choosing deliberately, every time, in the same way you'd choose any other professional resource.

One caveat: Responsible AI use extends well beyond what we cover here. Data privacy, algorithmic bias, disclosure obligations, and governance frameworks are all fundamental to that work. This series focuses specifically on the environmental dimension and how to make better decisions with that impact in mind.

Okay one more thing: The most energy-efficient AI interaction is the one you don't have. Not because AI is bad, but because your judgment, expertise, and creativity are the point. AI works best in service of those things and not as a substitute for them.

Part 5: Where will AI go from here?

There is so much we don’t know. It’s worth remembering that ChatGPT hasn't even been around for five years. These technologies — and how we use them personally and professionally — are evolving faster than anyone, including the people building them, can fully track. Predictions made today will look naive within a year.

But a few things seem likely.

We are going to have to pay for it. Right now, AI is artificially cheap — subsidized by investor capital in the familiar tech pattern of buying users first and worrying about margins later. As these companies move toward profitability, pricing structures will need to reflect the actual cost of running these models. Most industry observers expect future pricing to scale with query complexity, which in turn reflects the energy required to complete it. That means the table in the previous chapter won’t just be a guide to reducing your environmental footprint — it will increasingly map to your company’s AI budget.

That creates an interesting alignment. Corporate sustainability has long made the case that reducing water, waste, and energy consumption isn’t just an environmental imperative but a business one. AI may follow the same arc: the organizations that learn to use it efficiently now will be better positioned when the costs become visible on a balance sheet.

And there’s a disclosure dimension worth watching. Regulatory frameworks like the EU’s CSRD are already pushing companies toward greater transparency on digital infrastructure and technology-related emissions. It’s not difficult to imagine a near future in which AI use is a standard line item in sustainability reporting. Sustainability professionals who understand this landscape now, before it becomes a compliance requirement, will be better positioned to both report on AI's footprint and operationally reduce it.

It's also a leadership opportunity that's becoming more urgent: the American people are growing increasingly skeptical of AI, even as their adoption of the tech grows. A May 2026 Pew Research survey found that about half of U.S. adults now use AI chatbots — up from a third in 2024 — but that majorities believe AI is advancing too quickly, predict it will have a negative impact on their lives, and worry it will put their personal information at risk. Adoption and skepticism are rising in parallel. Meanwhile, community opposition to data center construction hit a record in the first quarter of 2026, with at least 75 projects worth $130 billion blocked or delayed nationwide — more than in all of 2025 combined. Active opposition groups more than doubled to 833 across 49 states. Proposed moratoriums on data center construction appeared in 14 states, with federal legislation introduced by Bernie Sanders and Alexandria Ocasio-Cortez.

This is the environment in which AI will mature. Public trust is not guaranteed. The social license to operate — a concept sustainability professionals know intimately — applies here too. The skills this moment requires aren't primarily technical. They're the skills of someone who knows how to read a materiality assessment, counsel a client through a disclosure they'd rather not make, make the case that transparency is a competitive advantage rather than a liability, and understands that the gap between what a company reports and what it actually does is where reputational risk lives.

Few professional communities are better equipped for this moment than sustainability. The work has always been about making costs visible, holding institutions accountable to what they actually do rather than what they report, and making the case for long-term thinking in organizations built for short-term results. AI doesn't change that work. It expands it.