AI ethics is not only about bias, safety, and privacy. It’s also about watts, water, wires and waste. If we ignore the physical footprint of “intelligent” systems, we risk building a smarter digital world on an unsustainable foundation.

Why this matters now

Over the last two years, generative and agentic AI have leapt from labs into daily life. That surge has a material cost, electricity to train and serve models, water to cool data centers, specialized chips to run them, and eventually electronic waste when hardware turns over. MIT researchers recently summed it up bluntly, we are improving AI faster than we are measuring the trade-offs, and our governance is struggling to catch up.

At the same time, major tech firms have disclosed that AI is complicating their climate pledges. Google’s emissions rose 13% in 2023 and are ~48% higher than 2019, largely because AI drove more data-center energy use exactly the opposite direction of its 2030 net-zero ambition.

The ethical question is not only “what did the model predict?” It’s also “what did the model consume to predict it?”

The AI energy story in plain terms

Training is a power-hungry marathon

Training frontier models requires vast compute clusters running for weeks. While exact numbers vary by setup, studies and disclosures converge on the same story, large models consume large amounts of energy and produce non-trivial emissions especially when grids are fossil-intensive. MIT’s explainer highlights rising electricity demand from both training and deployment, with significant uncertainty because measurement is still maturing. MIT News

Inference is a never-ending treadmill

Once a model is public, the real footprint begins: billions of queries mean billions of inference runs, 24/7, on fleets of accelerators. Even modest per-query energy can scale to enormous totals at global usage. Again, MIT notes that the operational phase (serving users) is a major share of generative AI’s overall impact and one organizations often underestimate.

Data centers are the new industrial sites

Data-center electricity use is climbing with AI. Reports and analyses around Google’s 2024 environmental report link the 13% YoY emissions rise and 48% five-year increase to AI-driven compute expansion. This aligns with wider concerns that data-center energy demand could double mid-decade. Data Center Dynamics

The water we do not see

Cooling high-density AI clusters takes water directly at sites using evaporative cooling and indirectly through water used in power generation.

One widely cited case, training GPT-4 on Microsoft’s Iowa supercomputing cluster coincided with a 34% jump in Microsoft’s global water consumption (2021→2022), and reporting described millions of gallons used for cooling during peak summer training. Local stories and the AP’s coverage made the “hidden water cost” legible to the public. Iowa Public Radio

For ethics teams, that reframes “responsible AI.” A system that treats users fairly but draws substantial water from stressed watersheds raises a different kind of harm one felt by surrounding communities and ecosystems. MIT’s two part series explicitly calls out water as a key impact vector that needs better measurement.

The hardware behind the hype and the e-waste ahead

AI’s footprint begins before the first line of code runs. Manufacturing advanced GPUs/accelerators is energy- and water-intensive and depends on minerals whose extraction can be environmentally damaging. Then, because AI evolves rapidly, expensive hardware turns over quickly feeding a rising e-waste stream.

Analyses from IEEE Spectrum and others warn that generative AI’s pace could add millions of tons of additional e-waste annually by the end of the decade if current refresh cycles persist. The waste includes not just chips but memory, boards, power systems, and batteries often containing hazardous substances. IEEE Spectrum

This is the part of “AI ethics” that almost never makes the slide deck but it should.

Reality check: what leading companies are reporting

• Google: Emissions +13% YoY in 2023; +48% vs. 2019. The company attributes much of the rise to AI driven data center energy and supply chain emissions, highlighting the difficulty of cutting carbon as compute intensity grows.
• Microsoft/OpenAI: As OpenAI’s cloud partner, Microsoft’s cooling water use and energy needs surged with GPT-4’s training. Reporting connected the Iowa build out to significant local water use during hot months. Microsoft says it’s pursuing cleaner energy, water-positive operations, and more efficient AI systems. AP News
• Industry wide: MIT researchers and the OECD both note that transparent, AI specific measurements are still limited, complicating independent verification and policy design.

From “Responsible AI” to “Responsible Infrastructure”

Ethics teams have matured on topics like fairness, explainability, and human oversight. The environmental dimension adds three more pillars to your governance stack:

1. Energy: What you consume and when and where you consume it matters. Emissions vary with grid mix and time of day.
2. Water: Cooling choices (evaporative vs. dry/immersion) and location (arid vs. water rich regions) change the real-world impact.
3. Materials/E-waste: Design for longevity, refurbish where possible, and build credible end-of-life pathways for gear.

OECD’s recent work on the “AI footprint” urges governments and companies to standardize measurement, improve transparency, and look beyond just operational electricity to lifecycle impacts (manufacturing through disposal). That’s the blueprint to turn good intentions into comparable numbers and eventually, accountability. OECD

What policy is (and isn’t) doing yet

EU AI Act: a start, not the finish line

The EU AI Act is the first comprehensive AI law. Its core focus is risk to people, but the final text and subsequent guidance are beginning to pull in sustainability especially for foundation models and general-purpose AI, where transparency around resource use is emerging. Observers still call the Act a missed opportunity on environment, but the door is open via codes of conduct and delegated acts to strengthen energy and transparency provisions. Clifford Chance

UNESCO: environment is an ethical principle

UNESCO’s 2021 Recommendation on the Ethics of AI…. adopted by 193 member states explicitly elevates environmental and ecosystem well-being as a core value alongside human rights. While non-binding, it gives countries a common language to integrate sustainability into national AI strategies and procurement. UNESCO

OECD: measure first, govern better

The OECD’s 2025 work program on the AI footprint pushes for standardized metrics, broader data collection, and AI-specific impact tracking across energy, water, and materials so policies can target AI as AI, not just as generic “ICT.” OECD AI

Bottom line: policy is moving but measurement and disclosure are prerequisites. Without them, legislating effectively is guesswork.

The overlooked risks: chemicals and fugitive gases

As scrutiny grows, advocates are flagging PFAS (“forever chemicals”) in cooling systems/electronics and f-gases used in HVAC or chipmaking. These persistent substances pose health and environmental risks if leaked or poorly handled, adding another layer to the AI-infrastructure footprint. Expect transparency and phase-down debates to accelerate with the AI data-center boom. The Guardian

A practical playbook for “Green AI” in your organization

You don’t need to run a hyperscaler to act. Here’s a pragmatic checklist you can adopt (and signal publicly):

1) Measure like you mean it

• Track at the workload level. Start attributing energy and emissions to specific training runs and high-traffic inference services. If your cloud lacks granular meters, use best-available estimators and push vendors for better telemetry. OECD’s guidance offers a measurement scaffold.
• Include water. Log site-level cooling water draw and (where possible) power-sector water intensity, not just electricity. MIT’s experts emphasize water as a first-class impact.
• Account for hardware. Add embodied carbon of accelerators/servers into your lifecycle inventory so upgrade decisions reflect the true cost.

2) Design for efficiency by default

• Right-size models. Consider distilled, pruned, or specialized models for most workloads; reserve largest models for use-cases where they clearly add value. (Your users feel latency not parameter count.)
• Carbon-aware scheduling. Where latency allows, shift non-urgent training/jobs to hours and regions with cleaner grids. Many cloud regions publish carbon-intensity signals.
• Optimize inference. Use quantization, caching, prompt engineering, and batching to cut per-request compute.

3) Cool smarter, site smarter

• Cooling tech. Explore dry cooling or liquid immersion to reduce water draw; upgrade controls (AI-assisted optimization has yielded big PUE/WUE gains).
• Location strategy. Avoid placing water-intensive facilities in stressed basins. if you must, pair with meaningful offsets and community agreements, and publish the numbers.

4) Close the loop on hardware

• Extend lifetimes. Prioritize refurbish/redeploy pathways over early retirement; design procurement to require take-back and certified recycling. Analyses warn of a coming AI driven e-waste spike if we don’t.
• Secure-by-design reuse. Use robust data-sanitization to unlock reuse without security trade-offs.

5) Disclose and commit (publicly)

• Publish an AI footprint report (energy, emissions, water, hardware) at least annually, even if estimates are imperfect. Transparency builds trust and momentum for better data.
• Align to UNESCO/OECD principles and the EU AI Act’s evolving sustainability expectations; consider joining voluntary data-center pacts where applicable.

The trade-offs to navigate honestly

• Carbon vs. water: Air cooling can save water but increase power draw; immersion cooling can save power but add complexity. Context beats one-size-fits-all.
• Performance vs. efficiency: The market rewards accuracy and capability, not joules saved. Leaders will make efficiency part of product DNA (and storytelling).
• Local jobs vs. local resources: Data-center investments bring tax base and work but also strain water and grids. Community-level transparency and benefit-sharing are key.

What to watch in 2025–2026

• EU delegated acts & codes of conduct that may harden AI energy/transparency expectations especially for foundation models. White & Case
• OECD measurement pilots and tooling that standardize how firms report AI energy, water, and hardware impacts.
• Corporate sustainability updates from hyperscalers as their AI build-outs collide with 2030 net-zero and water-positive pledges (expect more difficult conversations in annual reports).
• Chemicals & f-gases scrutiny tied to data-center cooling and semiconductor manufacturing.

Bringing it home: ethics beyond algorithms

If your responsible-AI program ends at model cards and bias audits, it’s incomplete. The environmental dimension is now table stakes:

• For leaders: Set a visible target (e.g., “50% reduction in energy per inference by 2026”) and report quarterly progress.
• For builders: Treat efficiency as a feature. Celebrate a 30% energy cut like a 3-point accuracy gain.
• For policy teams: Push for AI-specific disclosure standards so leaders are not punished for being transparent while laggards hide in averages.

AI can help solve climate problems from grid optimization to materials discovery. But the means should match the ends. When we make how AI lives on the planet as important as what AI does for people, we move from “responsible AI” in theory to responsible AI infrastructure in practice.

What would you add to this playbook? If your org is measuring (or struggling to measure) AI’s footprint, I’d love to hear what’s worked and what hasn’t.