Talks

The Hidden Footprint of AI — Miriam Aczel on carbon, water and land

56 min read

About this talk

Speaker: Miriam Aczel (guest); host not named in the transcript
Event: “Purposeful Empathy” podcast (season on empathy in the era of AI), episode with the lead author of the UNU-INWEH report on AI’s environmental cost
Date: ~July 2026 (inferred, not confirmed: she says the UN Secretary-General’s announcement at London Climate Action Week was “a couple of weeks ago” [34:13]; that speech was on 23 June 2026)
Length: 53:02

How to read this page. It has two kinds of content, always marked:

  • 🎙 In the talk. What Aczel (or the host) says, with [MM:SS] timestamps. Only these parts are her views.
  • 📚 Deeper theory. Background from the fields her work draws on, plus numbers and critiques from outside the talk. This explains why she says what she says and who disagrees. It is not her claim unless a timestamp says so.

Suggested route. Summary → Map → Key terms → Part 2: The intellectual map (read once, slowly) → Deep dives in order → “Putting it together”.

A note on sources. Facts about the report (numbers, authors, dates) come from the UNU press material linked below. Facts in Part 2 about older books and papers (names, dates, key ideas) are from memory unless a link is given; they are well-known works, but I did not re-check each one. Anything I could not confirm is marked “(not confirmed)“.

Speakers

LabelWhoRole
HostNot named in the transcript. The show Purposeful Empathy is hosted by Dr. Anita Nowak per its public listings (Podcast page, Publishers Weekly). That this is her episode is an inference (not confirmed)Interviewer. Says she has studied empathy for 15 years [02:45]; the series has 25 parts on AI and society
AczelDr. Miriam Aczel (captions: “Miriam Axel”). Introduced by the host [03:21]Researcher in sustainability transition and justice at UNU-INWEH. Lead author of the June 2026 report. Imperial College PhD on fracking; McGill undergraduate; Berkeley postdoc (UNU profile, CIEE news)
Pre-roll voicesUnnamed[00:00]–[02:01] is a trailer montage. The first clip [00:00] and part of the third [01:16] are Aczel (same words as [10:16] and [16:20]). The rest are other guests from the series (AI psychosis, Arabic-language LLMs, manipulation, decision autonomy). Do not attribute them to Aczel

Caption warnings. This is auto-captioned English speech. Corrections used on this page:

  • “Miriam Axel” = Aczel (confirmed by the UNU press release).
  • “Professor Kav Madani” = Kaveh Madani, director of UNU-INWEH and last author on the report.
  • “Inway” = UNU-INWEH (the Institute for Water, Environment and Health), pronounced “in-way” [05:29].
  • “Rouge” [48:43] = Khmer Rouge. “Pen” = Phnom Penh. “sle a lot of NDAs” [20:16] is probably “sign a lot of NDAs”. “part of Magnation” [20:16] is probably a captioning error (the host may have said “MAGA nation”; not confirmed).
  • “Karen” [45:35] is a supervisor’s first name; the surname is not given.

Slips in the talk (checked):

  1. Report title. The host calls it “Environmental Cost of Artificial Intelligence: Carbon, Water, and Land Footprints” [04:24]. The published title is “Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints”, released 3 June 2026 (DOI 10.53328/INR26RMA002). The word “Energy” matters: the report counts the footprint of electricity, not of the whole AI supply chain. Aczel makes this point herself [10:16].
  2. Age of the zero. Aczel says the Cambodian zero is “2,000 years old” [50:02]. The stone (K-127, from Sambor) is dated to year 605 of the Saka era, which is 683 CE, so about 1,340 years old (Scientific American, ScienceAlert). The French scholar who identified the zero was Georges Cœdès, and his paper is from 1931, not 1929. He is usually called an epigrapher or historian, not a mathematician.
  3. “Data centre cooling water was not in the report.” This is not a slip. She says it clearly [12:23]. It is the single most important scope limit; keep it in mind for every water number below.

Summary

The central question. What does AI really cost the planet, if you look past the “cloud” and past carbon alone?

The thesis (hers). AI is a physical system. It sits on minerals, electricity, water, land and a long chain of suppliers. Looking at one metric, carbon, can hide trade-offs. A low-carbon power source can have a large water footprint (hydropower) or a large land footprint (solar fields). So “clean” depends on which footprint you count and where you are. The right trade-off is local: a water-stressed place with spare land should choose differently from a place with the reverse [14:31].

Her answer: information first. She wants a “nutrition label for AI” [19:23]: where the electricity came from, where the query was processed, the carbon, water and land footprints. It should tell people facts, and not judge them. Then choice follows: users can pick lighter models, shorter prompts, off-peak timing [23:44]–[41:03]. Companies should disclose and talk to communities early [38:40]. She welcomes the UN Secretary-General’s new AI Environmental Transparency Initiative [34:13].

Her stance. She says three times that she is “not anti-AI” [12:23], [13:08], [37:44]. AI helps climate science (wildfire, sea-level and flood modelling), and the benefits are large. But she sees a pattern from fracking: fast, extractive, benefits flow away, communities pay, NDAs [20:16]. She argues for fast-tracking only uses with clear public benefit [18:09].

The close. The last 8 minutes are the host’s standing “empathy” question. Aczel tells a personal story about support after her father’s death, and about his search for the oldest written zero [45:35]–[52:14]. It is a human coda; it does not carry the argument.

Note on the register. This is a friendly, public-audience interview. Aczel is careful and diplomatic, and she says several times what the report did not cover. The numbers in the report come from a press release and a short podcast; the talk itself gives very few numbers (2.5 billion prompts a day [35:32]; “a fraction of a watt hour” for a short answer [41:03]).

Map — talking points

#PointTime
1Trailer: clips from several guests, including Aczel on supply chains and “carbon alone is not enough”[00:00]–[02:01]
2Host’s series intro and guest introduction[02:45]–[04:24]
3Aczel’s path: fracking and justice, Berkeley community microgrid, then AI. UNU-INWEH as “the UN’s think tank on water”[05:29]–[08:45]
4Why “the cloud” misleads: AI depends on a long supply chain, from minerals to electricity[10:16]
5Carbon, water and land: low carbon is not low water or low land. Hydropower in Ontario and Quebec. Cooling water was not in the report[11:49]–[13:48]
6”The $64,000 question”: who decides which trade-off? Context specific. Life cycle from mine to e-waste[14:31]
7The report looked at countries, not communities. Local engagement, traditional knowledge, company disclosure[15:33]–[16:20]
8Lessons from fracking: don’t fast-track extractive technology; but some AI uses deserve a faster path[18:09]
9Climate justice defined as “who benefits, who pays”. The nutrition label for AI[19:23]
10Host on data-centre communities (brown tap water). Aczel: AI could entrench slow decarbonisation; concentration of control; NDAs; but real benefits (genomes, tornado forecasts)[20:16]–[23:02]
11Real-time disclosure: electricity maps, time-of-use, queue a task for 3 a.m.[23:44]–[26:14]
12AI for climate: modelling, wildfire, sea level; “This Climate Does Not Exist”; AI vs a human’s footprint[26:14]–[32:15]
13From guilt to systems: UN AI Environmental Transparency Initiative; prompt discipline; 2.5 billion prompts a day[34:13]–[37:44]
14Match the task to the tool (cookbook vs chatbot). Lighter models; “thinking” vs “instant”[37:44]–[41:03]
15One ask of companies: say where the power and the data centre are. Engage communities first[38:40]–[40:20]
16How citizens engage: disclosure first; local priority (water, noise, pollution)[42:39]–[45:00]
17Host’s closing question on empathy; Aczel’s father; the Cambodian zero[45:00]–[52:14]

Part 1 — Before you read

Key terms

  • Footprint. A measure of how much of a resource an activity uses, counted along a chain. A carbon footprint counts greenhouse gases. A water footprint counts freshwater. A land footprint counts area. Each can be counted differently, so numbers from different studies often do not match.
  • Water: withdrawal vs consumption. Withdrawal is water taken from a source, and much of it is returned (for example, river water passed through a power-plant cooler). Consumption is water that does not come back (evaporated, or built into a product). Many “AI uses X litres” claims mix the two. Always ask which.
  • Direct vs indirect water (scope). Direct (on-site) water is used at the data centre, mostly for cooling. Indirect (off-site) water is used to make the electricity. Embodied water is used to make the chips and buildings. Aczel’s report counts only the electricity (indirect) part [12:23].
  • Low carbon is not low impact. Her main idea. Hydropower: low carbon, but reservoirs evaporate water. Large solar: low carbon, but takes land. Biomass: can cut carbon but can raise water and land use a lot (the report gives roughly 70% less carbon but over 30 times the water and about 100 times the land when switching coal to bioenergy; from the press release).
  • Inference vs training. Training builds a model (once, very large). Inference is each time someone uses it (billions of times). The report says inference is 80–90% of AI energy. This is why she focuses on prompts.
  • Supply chain / life cycle. Mines → chips → data centres → electricity → use → e-waste. She says the report covered only electricity and that the rest is a “real lack of information” [10:16].
  • Climate justice. Her definition: “who benefits, who’s paying the cost” and where [19:23]. The harm and the benefit land in different places.
  • Extractive technology. One where value is taken from a place and flows elsewhere, and where harms are close to the site (fracking, mining). She asks whether AI infrastructure is the same [18:09].
  • Nutrition label (for AI). Her proposal: a standard label with sources and footprints. Key feature: it gives facts and does not tell you what to choose [19:23], [35:32].
  • Rebound effect (Jevons paradox). When something gets cheaper or more efficient, people use more of it, and total use can rise. She does not use the term in the talk. The report’s authors use the idea (the press release says efficiency gains may be offset, and calls for caps on tokens, resolution and output length). See Part 2.6.
  • Time-of-use pricing. Electricity costs more at peak hours (her example: 4–9 p.m. in California) [24:58]. It tells people when power is scarce.

People and works named (a reading list)

Person / workWhy it comes upStart with
Kaveh MadaniDirector of UNU-INWEH. Proposed the research topic [05:29]The press release and quotes in UNRIC
The UNU-INWEH report (Aczel, Chamanara, Matin, Farsi, Marwala, Madani, 2026)The basis of the talkPress release; the report itself (DOI 10.53328/INR26RMA002)
António GuterresLaunched the AI Environmental Transparency Initiative [34:13]UNU-INWEH note, Al-Monitor
Electricity MapsLive grid carbon data [23:44]Their public map (not checked for this page)
This Climate Does Not ExistAI tool that shows your address under climate impacts [29:30]Mila’s page
AOC (host’s mention)The host says she visited communities near data centres with water samples [20:16]See Part 3, dive 4 for what I could check
Oakland EcoblockAczel’s Berkeley project: community microgrid [05:29]Search the project name (not checked for this page)
Amir Aczel (not named in the talk)The father’s work matches his book Finding Zero (2015). That Miriam is his daughter is my inference from the surname and the story (not confirmed)Scientific American

Aczel’s own frame

Aczel trained as an energy and environmental justice researcher. Her PhD (Imperial College London) studied fracking in the UK, US, France, China and Algeria (profile). In the talk she describes it as “community impacts of energy systems … who pays the costs especially when an extractive technology like fracking [has] benefits flow elsewhere” [05:29]. Her postdoc at UC Berkeley was on community energy. That history explains why the talk keeps returning to community, voice and who pays, and why she treats the footprint numbers as a start, not an end. She calls the report a first pass screening overview [23:44] and says it is “not a case against AI” (also Madani’s line in the UNRIC release).

The current debate (2025–26)

  • The report itself and its headline numbers. UNU-INWEH (published 3 June 2026) estimates data centres used 448 TWh in 2025 and could use 945 TWh by 2030. It puts the 2030 electricity-linked footprints at 399 million tonnes of CO₂, 9.3 trillion litres of water and over 14,500 km² of land (press release, UNRIC). A caveat from my read of the press release: the method and the water and land factors are not shown there, and I did not read the full report. Treat the equivalences (trees, football fields, people’s drinking water) as illustrations.
  • The demand forecast. The 945 TWh figure matches the IEA’s base case in Energy and AI (data centres at about 415 TWh in 2024, about 1.5% of world electricity, more than doubling by 2030; IEA). The IEA also says uncertainty widens after 2030.
  • How much does one prompt cost? Google’s August 2025 paper put a median Gemini text prompt at about 0.24 Wh and 0.26 mL of water, and said per-prompt energy fell about 33-fold in a year (SiliconANGLE, TechRepublic). That is a company’s own number, for one product, and it counts on-site water. Aczel’s “fraction of a watt hour” [41:03] is in the same range.
  • “It’s overblown” vs “it’s a crisis”. Andy Masley and others argue that a chatbot prompt is small next to everyday activities, and that US data-centre water is a small share of national use (his estimates; see the summary of his case). Others reply that totals, not per-prompt averages, are what strain local grids and aquifers, and that Masley’s framing is contested (AOL/Independent). Aczel sits in a middle position: per-use is modest, the aggregate and the location matter [14:31], [35:32].
  • Water estimates are wide. The best-known academic work, Shaolei Ren’s group (arXiv), separates on-site (scope 1) from off-site electricity (scope 2) water, leaves out chip-making, and projects 4.2–6.6 billion m³ of AI water withdrawal in 2027. UNU-INWEH’s 9.3 billion m³ is for a different year, a different measure (a footprint of electricity) and a different scope. Do not set these two numbers side by side.
  • Local fights are real. A 2025 New York Times investigation reported well-water problems near a Meta data centre in Georgia; Meta said its study found an effect “unlikely”; in 2026 families sued (PPC Land summary, AJC). Cause is disputed. It matches the kind of case the host describes at [20:16].
  • Transparency is now policy. Guterres announced the initiative at London Climate Action Week on 23 June 2026, asking every major AI company to measure and publish carbon, water and land impacts and to power data centres on renewables by 2030 (UNU-INWEH, Al-Monitor). Company responses were not found (not confirmed).
  • Efficiency vs rebound. A 2025 paper by Luccioni, Strubell and Crawford argues that efficiency gains in AI may increase total use (arXiv). The report’s own line, quoted by UNRIC, is Madani’s: “More efficient and affordable AI and energy mean more consumption of AI.”

Part 2 — The intellectual map

Aczel spoke to a general audience, so she named few theories. But her argument sits on eight bodies of work. Here they are laid out, including ones she does not name. (Dates and titles in this part are from memory unless a link is given.)

 FOOTPRINT ACCOUNTING ──────┐                 ┌──── WATER–ENERGY NEXUS
 LCA · water footprint      │                 │     thermoelectric · hydro
 GHG scopes · burden-shift  ▼                 ▼     withdrawal vs consumption
                    ┌─────────────────────────────────┐
 DIGITAL MATERIALITY│   ACZEL: carbon alone is not    │ ◄── ENVIRONMENTAL JUSTICE
 cloud = cables,    │   enough; look at the whole     │     Bullard · Nixon · Schlosberg
 mines, water       │   chain, locally, with people   │     UNDRIP · fracking studies
 ─────────────────► └─────────────────────────────────┘
                       ▲            ▲            ▲
   EFFICIENCY & REBOUND│   TRANSPARENCY GOVERNANCE   │ FLEXIBLE / CARBON-AWARE COMPUTE
   Jevons · Khazzoom   │   TRI · labels · defaults   │ time & place shifting, grid data
   Koomey · Masanet    │   Fung · Maniates           │ Radovanović · Dodge · WattTime
                       │                             │
                       └────── AI FOR CLIMATE (benefit side) ──────┘
                               Rolnick · Kaack · Ligozat

2.1 Footprint accounting — how we count a “cost”

The problem it addresses. Any product touches many places and many resources. If you count only one (carbon), you may move a harm instead of removing it. Footprint tools try to count the whole chain.

The chain of ideas

  1. Life-cycle assessment (LCA). Developed from the 1960s–70s (packaging and energy studies) and standardised as ISO 14040/14044 in the 2000s. It counts inputs and outputs “from cradle to grave” and across many impact categories (climate, water, land, toxicity). Key point for Aczel: LCA already says one category is not enough. Her look at the entire life cycle … to end of life [14:31] is plain LCA language.
  2. Burden shifting. The standard warning in LCA: fixing one impact can raise another. Her bioenergy example (less carbon, much more water and land) is textbook burden shifting. Her “trade-offs at every point” [00:00] says the same.
  3. Ecological footprint (Wackernagel and Rees, 1996). Converts human demand into a land area. It made “land as the common unit” a public idea and led to “how many Earths” claims. It is also much criticised (it hides detail), a cautionary note for any single-number footprint.
  4. Water footprint (Arjen Hoekstra, from 2002; Water Footprint Network, 2008). Splits water into blue (surface and groundwater), green (rain used by plants) and grey (water needed to dilute pollution). It also brought in virtual water (Tony Allan, 1990s): the water embedded in traded goods. Key point: when she says the electricity has a water footprint [12:23], she is using this idea. The water is used far from the data centre, at a dam or a power plant.
  5. GHG Protocol scopes (1, 2, 3). Scope 1 = direct emissions; scope 2 = purchased electricity; scope 3 = everything else in the value chain (chips, buildings, use). Ren’s water papers borrow the same scheme (on-site = scope 1, electricity = scope 2). The UNU report covers the scope-2 type of footprint only, which is why she calls the data on supply chains weak [10:16].
  6. The accounting trap: location vs market basis. A company can buy renewable certificates and report low scope-2 carbon even when its local grid runs on gas. The location-based number shows what physically happens. This is why she wants to know “where that renewable electricity is coming from” [23:44].

How it feeds the talk. The whole “$64,000 question” [14:31] is a multi-criteria question: LCA says no single metric wins, so someone has to weigh them. LCA itself cannot say who. That moves the issue from science to politics, which is where she goes next (2.4).

2.2 The water–energy nexus — why electricity “uses” water

The problem it addresses. Producing power needs water, and moving and cleaning water needs power. The two systems are tied. Water numbers for AI depend mostly on how the power is made.

The chain of ideas

  1. Peter Gleick, “Water and energy” (1994). An early systematic account of the link. It put the water–energy nexus on the policy agenda.
  2. Withdrawal vs consumption. Thermal plants (coal, gas, nuclear) withdraw huge amounts for cooling and return most of it, but they consume some by evaporation. Wind and solar PV use very little operational water. Numbers vary widely by cooling type (once-through, recirculating, dry). The NREL review by Macknick and colleagues (2012) is the usual source for these factors.
  3. Hydropower is the hard case. A reservoir exposes a large water surface that evaporates. Estimates run from near zero (run-of-river) to very high (tropical reservoirs). The US Argonne GREET note gives about 9.85 gallons per kWh on average for hydro, with most of that from reservoir evaporation (ANL). A review warns that the common method counts gross evaporation and ignores what the land would have evaporated before the dam, so it can overstate the net change (IAHS review). Key point for Aczel: she is right that hydro is low carbon and has a large attributed water footprint [12:23]. But the size is a method choice, and the number is disputed. Mekonnen and Hoekstra (2012) is the best-known paper on the blue water footprint of hydropower.
  4. Direct cooling water. Data centres cool with chillers, evaporative towers or dry coolers. The Green Grid’s WUE metric (litres per kWh of IT energy) is the standard measure. Evaporative cooling saves electricity and spends water; dry cooling does the reverse. This is the same trade-off at the building scale as the hydro-vs-solar one at the grid scale. She names cooling water as important but out of her report’s scope [12:23].
  5. Water stress is local. The same litre means little in a wet basin and a lot in a dry one. Tools like the WRI Aqueduct atlas map baseline stress. Key point: this is her “if a community is water stressed but has abundant solar … better to have a larger land footprint” [14:31].

How it feeds the talk. It explains her paradox: the greenest grid (hydro) can have the thirstiest AI. And it explains why she says country-level numbers are not enough [15:33]: water is a basin-level problem.

2.3 Digital materiality — the cloud is a factory

The problem it addresses. Words like “cloud”, “virtual” and “digital” hide that the internet is cables, buildings, chips, mines and power plants. A research tradition has worked to put the matter back.

The chain of ideas

  1. Media archaeology and “geology of media” (Jussi Parikka, 2015). Media run on minerals and end as waste. Digital culture has a deep-time material base.
  2. Infrastructure studies. Nicole Starosielski, The Undersea Network (2015): the internet rides on cables laid on the sea floor. Tung-Hui Hu, A Prehistory of the Cloud (2015): the cloud metaphor hides older networks and power. Vincent Mosco, To the Cloud (2014): a critical political economy of cloud computing.
  3. Water and data centres. Mél Hogan’s “Data flows and water woes” (2015) on the Utah data centre was an early study of data-centre water politics.
  4. “Anatomy of an AI System” (Kate Crawford and Vladan Joler, 2018) and Atlas of AI (Crawford, 2021). Maps one smart speaker back to lithium mines, shipping, labour and landfill. The book frames AI as an extractive industry, with chapters on minerals, labour and data. Key point: Aczel’s opening, “the cloud sounds so ethereal … a very long and diverse supply chain from critical minerals … halfway across the globe” [10:16], is this tradition in one sentence. Her “extractive” framing [18:09] matches Crawford’s.
  5. Karen Hao, Empire of AI (2025). A journalist’s account of OpenAI and its data-centre expansion, including local water and power conflicts. It adds a political-economy layer: who controls compute.
  6. The measurement turn in AI research. Emma Strubell et al. (2019) put a carbon number on training language models. The headline figure was an extreme case and was later criticised, but it started the field. Schwartz et al., “Green AI” (2019–20), argued for reporting efficiency, not only accuracy. Patterson et al. (2021, Google) argued that careful choices of model, hardware and location can cut training emissions a lot. Luccioni and colleagues (e.g., the BLOOM footprint study, 2022–23, and “Power Hungry Processing”, 2023) measured inference across tasks: generating an image takes far more energy than text generation or classification (secondary summary, not checked at source). Alex de Vries (Joule, 2023) projected the sector-wide electricity demand. Key point: the report’s “200× and 1,450× the energy of basic text classification” comparison [press release] is in this line of work.

How it feeds the talk. This tradition gives the frame (AI is physical). The talk adds a specific proposal: put the physical facts on a label (2.5).

2.4 Environmental and climate justice — who pays

The problem it addresses. Harms from industry are not spread evenly. They fall on poor, minority and rural communities, while benefits go elsewhere. The field asks how to see and change that.

The chain of ideas

  1. The origins (US, 1980s–90s). The 1982 Warren County (North Carolina) protests against a toxic-waste landfill; the United Church of Christ report Toxic Wastes and Race (1987); Robert Bullard, Dumping in Dixie (1990). The finding: waste sites follow race and poverty, not only land cost.
  2. Three (later four) dimensions of justice (David Schlosberg, 2007). Distribution (who gets benefits and burdens), recognition (whose knowledge and identity count), and participation (who decides). Later work adds capabilities. Key point: her talk walks through all three. Distribution: “who’s paying the cost” [19:23]. Recognition: “traditional knowledge, community experience” [16:20]. Participation: “engaging early and often” [16:20].
  3. Slow violence (Rob Nixon, 2011). Harm that is gradual, spread out and unseen (pollution, climate change) gets less attention than sudden harm. This is why the host’s brown tap water [20:16] is vivid but the “noise … bird migration … pollination” chain [44:25] is easy to miss.
  4. Climate justice. From the Bali Principles (2002) and the movement that followed: countries and groups that did least to cause climate change often suffer most, and have least say. Applied to AI: the compute is in two countries (the report says over 90% of AI-specialised data-centre capacity is in the US and China), and over 150 countries have little or none (press release).
  5. Extractivism critiques (Latin America and elsewhere). Eduardo Gudynas, Alberto Acosta and others describe economies built on exporting raw materials, with local costs and weak local benefit. Naomi Klein’s term “sacrifice zone” (from earlier US usage) describes places treated as expendable.
  6. Free, prior and informed consent (FPIC). UNDRIP (2007) gives Indigenous peoples a right to consent to projects on their land. Her “traditional knowledge” [16:20] and meaningful engagement … not just the tick in the box [05:29] connect to this. FPIC is a legal standard for Indigenous peoples; it does not apply to every community near a data centre.
  7. Fracking as a case study. Her own field. Shale gas arrived fast, with local water, air and noise effects and contested evidence, and companies often settled disputes privately with non-disclosure agreements. She uses it as a pattern [20:16]. Whether AI follows the same pattern is a claim to test, not a given (see “Where others push back”).

How it feeds the talk. It supplies her values: the question after “how much?” is “for whom?” [19:23]. It also explains why she asks for early community engagement and not only disclosure.

2.5 Transparency as governance — labels, disclosure, defaults

The problem it addresses. If markets and regulators lack information, bad outcomes persist. Can publishing facts change behaviour?

The chain of ideas

  1. Targeted transparency (Archon Fung, Mary Graham, David Weil, Full Disclosure, 2007). Governments require firms to publish specific facts in a standard form (food labels, car safety ratings, the US Toxics Release Inventory from 1986). It works when the data is simple, comparable and used by someone with power: consumers, investors, regulators, journalists. It often fails when the facts are complex or no one has a reason to act. Key point: Aczel’s “nutrition label … no judgments” [19:23] is targeted transparency. The Toxics Release Inventory is the closest precedent for AI footprints, because it made local emissions public and helped communities push back.
  2. Labels that worked. Energy-label schemes (EU appliance labels, Energy Star). They worked partly because they were simple (one A–G scale) and partly because producers competed on them.
  3. Defaults and choice architecture (Thaler and Sunstein, Nudge, 2008; Johnson and Goldstein’s organ-donor study, 2003). People rarely change defaults. So whoever sets the default sets the outcome. Key point: her point that lighter defaults are generally better [41:03] is a default-design claim. A label informs; a default decides.
  4. The case against individual responsibility (Michael Maniates, 2001, “Individualization: Plant a Tree, Buy a Bike, Save the World?”). Putting the burden on consumers can distract from structural change. Key point: Aczel tries to avoid this: “it’s not a question of guilting people” [36:53]. But she still gives individual tips [35:32], so she walks this line.
  5. Existing and emerging rules. The EU AI Act requires providers of general-purpose models to document training and compute and the model’s known or estimated energy use (Annex XI; from memory, not checked). The EU’s data-centre reporting rules under the Energy Efficiency Directive (2023) ask operators for energy and water data (from memory, not checked). Hugging Face’s “AI Energy Score” (2025) rates models on energy for set tasks (from memory, not checked). The UN initiative of June 2026 asks for disclosure of carbon, water and land [34:13].
  6. Why disclosure is hard here. She says it herself: “It was really hard … finding this kind of granular information on … where a query … is sent” [38:40]. Firms route queries across regions, and the same model runs in many places. Real-time labelling needs data that firms do not currently publish.

How it feeds the talk. It is the core of her policy answer. Without this tradition, “nutrition label” sounds like a slogan. With it, you can ask the precise questions: Who reads the label? What can they do? Does it change the default?

2.6 Efficiency, rebound and growth — why “more efficient” may not mean “less”

The problem it addresses. Technology gets more efficient. Does total use fall?

The chain of ideas

  1. William Stanley Jevons, The Coal Question (1865). Better steam engines did not reduce coal use; they made coal use cheaper, so it grew. This is the Jevons paradox.
  2. Khazzoom–Brookes postulate (1980s) and the rebound literature. Economists (Daniel Khazzoom, Leonard Brookes) argued that efficiency gains in energy use lead to more use. Steve Sorrell’s UK Energy Research Centre review (2007) found rebound is real but varies: sometimes partial, sometimes more than 100% (“backfire”).
  3. The optimistic data-centre story. Jonathan Koomey’s work on computing efficiency (compute per joule doubling every 1.5 years or so, “Koomey’s law”) and Eric Masanet et al. in Science (2020) found that from 2010 to 2018 global data-centre computing grew several-fold while energy use rose only a few percent, thanks to efficiency. Key point: this is the optimist’s case. It held for a period when most growth came from shifting small servers into efficient hyperscale buildings, a one-time gain.
  4. The pessimist’s case now. AI-specific demand is of a different kind: dense GPU racks, new buildings. The IEA expects AI-optimised data-centre demand to more than quadruple by 2030 (IEA). Luccioni, Strubell and Crawford (2025) argue Jevons-type effects are central to AI (arXiv). The report’s authors say efficiency “alone won’t contain growth” and call for caps on tokens, resolution and output length (press release).
  5. Why this matters for her tips. Her per-prompt advice [35:32] is about intensity. Rebound is about volume. If shorter prompts make AI cheaper, more people may use it more. She says the aggregate matters (“2.5 billion new prompts per day … really adds up”) but does not discuss rebound in the talk.

How it feeds the talk. It sets the limit on what individual choices can do. It is also why the report pairs disclosure with demand-side caps.

2.7 Flexible and carbon-aware computing — when and where, not only how much

The problem it addresses. Electricity is not the same at every hour or place. A kWh at 3 a.m. in a windy region is cleaner than a kWh at 6 p.m. in a gas-heavy one. Can computing follow the clean power?

The chain of ideas

  1. Time-of-use tariffs and demand response. Utilities have long moved demand with price. Her California 4–9 p.m. example [24:58] is a standard peak-pricing window.
  2. Carbon-aware computing. Google’s team (Radovanović et al., 2022, “Carbon-Aware Computing for Datacenters”) shifted flexible jobs to hours with cleaner power. Dodge et al. (2022, “Measuring the Carbon Intensity of AI in Cloud Instances”) showed that the region and time of a training run change its carbon a lot. Services such as Electricity Maps and WattTime publish grid carbon data. Key point: her “queue it and get your answer tomorrow morning … 3:00 a.m.” [24:58] is carbon-aware batch computing offered to users.
  3. Grid flexibility as a resource. Researchers (for example, a 2025 Duke study led by Tyler Norris; from memory, not checked) argue that if large new loads accept small amounts of curtailment, the grid can host them without as much new capacity. This turns “flexible AI” into a grid-planning tool, not only an eco-tip.
  4. Limits. Time-shifting does not help much when the grid is fossil-dominated at all hours, or when the load is inference with a human waiting. Much AI demand is real-time. Also, moving work to the cleanest place can move water and land costs to that place (2.1, burden shifting).
  5. Siting. Where to build matters: cold climates cut cooling needs; sunny dry places give solar but may lack water. This is her water-stress vs land trade-off [14:31] in practice.

How it feeds the talk. It makes “choose when and where” a practical lever beside “choose how much”. She supports it [23:44] but notes the company has to publish the data.

2.8 AI as a climate tool — the benefit side

The problem it addresses. AI also helps climate work. How do we count the benefit against the cost?

The chain of ideas

  1. “Tackling Climate Change with Machine Learning” (Rolnick et al., 2019). A broad survey of where ML can help: power grids, buildings, transport, farming, forecasting, and climate science. It started a field.
  2. Three kinds of impact (Kaack et al., 2022, Nature Climate Change, “Aligning artificial intelligence with climate change mitigation”). (1) Compute-related (energy, hardware); (2) immediate application effects (e.g., a model that makes oil exploration cheaper or a grid cleaner); (3) system-level effects (rebound, lock-in, behaviour change). Key point: this is the best frame for the “cost vs benefit” question the host asks [26:14]. The benefit side and the cost side are not on one scale.
  3. Weather and flood forecasting. DeepMind’s GraphCast (Lam et al., Science, 2023) and later systems matched or beat traditional forecasts at far lower compute per forecast. Google’s flood forecasting (Nearing et al., Nature, 2024) reached ungauged basins. These are examples of the “tornado forecasting … things that on human computing scales we can’t do” that she cites [20:16].
  4. Visualising climate (Mila’s “This Climate Does Not Exist”, 2021). Built by a team under Yoshua Bengio, with Sasha Luccioni among the lead researchers; it uses generative adversarial networks to render flood, smog and wildfire effects on any Street View address (Mila). The aim was to make climate risk personal. Key point: Aczel presents it as a positive use [29:30]. Note that the creator, Luccioni, is also a leading researcher on AI’s footprint, which fits the “both/and” stance.
  5. The “AI vs a human” comparison. A 2024 Scientific Reports paper (Tomlinson et al.) estimated that AI writing and illustration emit far less CO₂ per page or image than a human doing the same task. The method has drawn criticism (a person’s footprint exists whether or not they write that page, so the comparison needs care about what is substituted). Ligozat et al. (2022) warn that studies of “AI for the environment” often skip the AI’s own life-cycle cost. Key point: Aczel makes a version of this comparison [27:19]; it is a per-task comparison, and the total effect depends on whether AI replaces work or adds to it (rebound again).

How it feeds the talk. It is her balance. She does not want the cost story to bury the benefit story [29:30]. The tradition tells you what a fair accounting would need: net effects, not gross.

Part 3 — Deep dives

1. “Carbon alone is not enough” — the three footprints [10:16]–[16:20]

🎙 In the talk. She starts with the word “cloud”: it “sounds so ethereal” and hides “a very long and diverse supply chain” from minerals “halfway across the globe” to electricity [10:16]. Her central line: “low carbon is not necessarily low water or low electricity … low land” [10:16]. Her examples:

  • Hydropower (Ontario, Quebec) is low carbon but has “a large water footprint” [12:23].
  • A large solar field has “a land impact” [13:08].
  • Fossil electricity drives the carbon footprint up as demand grows [14:31].

She says people think first of carbon, which is “really important”, and that the shift to renewables is “very important” [11:49]. But land and water “haven’t been capturing quite as much attention” [12:23]. She repeats the point at [16:20]: “looking at one metric, carbon alone is not enough.” She also says what the report did not do: it looked only at electricity [10:16], not data-centre cooling water [12:23], and not the rest of the supply chain, where “we don’t really have good data” [10:16].

The host sums it up: “clean electricity isn’t automatically clean across the board” [13:48]. Aczel does not say AI is not green. She says you cannot tell from carbon alone.

📚 Deeper theory

  • This is burden shifting (2.1). Moving from one footprint to another is the oldest warning in life-cycle assessment. The report’s bioenergy example (about 70% less carbon, but over 30 times the water and about 100 times the land; press release) is a dramatic case.
  • Why hydro is tricky (2.2). The size of hydro’s water footprint depends on the accounting method. Gross reservoir evaporation can be much larger than the net change in the basin’s water balance. So “hydro is thirsty” is true under one method and contested under another. This does not undo her point (trade-offs exist), but it means you should not memorise one number.
  • Water is a local good. A litre is not a litre. A footprint in litres without a place and a season is of limited use. That is why she keeps saying “context specific” [14:31].
  • Why carbon gets the attention. Carbon is global (a tonne anywhere warms the same planet), standardised (GHG Protocol), and priced in some markets. Water and land are local and have no common unit. Carbon wins the policy debate because it is easy to compare, not because the others matter less.
  • Planetary boundaries (Rockström et al., 2009; Steffen et al., 2015). A wider frame: climate change is one of several limits (freshwater, land-system change, biodiversity, nitrogen and phosphorus, chemical pollution). Aczel’s argument is the AI-sized version of this.
  • Embodied footprint is missing. Chip fabrication uses large amounts of ultra-pure water and energy; mining uses land and water. These are scope-3 items and are outside her numbers. A full “AI footprint” would be larger than the report’s.

Where others push back.

  • “The numbers are too small to worry about.” Critics like Masley argue that per-prompt and even national-level data-centre water is small next to agriculture (summary). The reply: totals are growing fast and are concentrated in a few places; national averages hide basin-level stress (2.2).
  • “Carbon is the right priority.” Some climate analysts say that a focus on water and land can be used to slow clean energy, which has overwhelming carbon benefits. Aczel anticipates this: we’re not anti-AI. Renewable energy is wonderful [13:08].
  • Method disputes. Because the report’s method is not shown in the press material I read, I cannot judge its water and land factors. Independent checks of the full report would be useful (not confirmed).

What she leaves out. Direct cooling water (said openly), chips and mining, e-waste beyond a brief line in the report, biodiversity, noise, heat, and water quality (only briefly at [42:39]).

Check yourself

  • Give two examples where a low-carbon option raises another footprint. What does each cost?
  • What is the difference between withdrawal and consumption, and why does it change a “water footprint” number?
  • Why does the report’s focus on electricity make its water number lower than a full-chain number?

2. The numbers — what the report says and what a prompt costs [35:32]–[41:03]

🎙 In the talk. Aczel gives few numbers. She says ChatGPT is “estimated to have 2.5 billion new prompts per day, that really adds up” [35:32]. She says the report was a first pass screening overview [23:44]. She says “a short answer could be a fraction of a watt hour and that variation between models and settings is huge” [41:03]. She says “thinking” models “use a lot more tokens which increases the electricity usage” [40:20]. She compares image generation: start in low resolution, then regenerate in high resolution only for the final [36:53].

📚 Deeper theory — the report’s numbers (from the press release; I did not read the full report)

ItemFigure
Data-centre electricity, 2025 (est.)448 TWh
Projected 2030945 TWh (about 3% of world use)
Share from inference80–90%
Training GPT-3 / GPT-4 (est.)1.3 GWh / 50–70 GWh
ChatGPT, 2.5 bn prompts/dayabout 383 GWh per year
2030 carbon, water, land (electricity-linked)399 Mt CO₂; 9.3 trillion L; over 14,500 km²
One AI image, waterabout 29 mL
One complex AI video, waterabout 4.1 L
One chat query vs text classificationabout 200× the energy
One AI image vs text classificationabout 1,450× (about 10 W LED for 17 min)

My arithmetic, to check consistency (not in the talk or the report).

  • 9.3 trillion L ÷ 945 TWh ≈ 9.8 L per kWh. ChatGPT: 383 GWh at that rate ≈ 3.8 billion L, which fits “about 500,000 people’s minimum domestic water”. So the report seems to use about 10 L of water per kWh of electricity across the board.
  • 383 GWh ÷ (2.5 billion × 365) ≈ 0.42 Wh per prompt on average. At about 10 L/kWh, that is about 4 mL of water per prompt (indirect, electricity only). That is above Google’s 0.26 mL (on-site only, one product, median prompt) and below the “bottle of water” claim sometimes repeated. The two numbers measure different things.
  • An image at 29 mL ≈ about 3 Wh, which matches the “10 W bulb for 17 minutes” (about 2.8 Wh). The report is internally consistent.
  • So the water number is basically energy × a fixed factor. This means its accuracy rests on the energy estimate and on that one factor. The factor (about 10 L/kWh) is a global average over very different grids.

Reading the energy numbers.

  • Per-prompt values vary hugely: Google’s median text prompt at 0.24 Wh [source above] is about half the report’s 0.42 Wh average for ChatGPT. Reasoning (“thinking”) modes, long contexts and agent loops can use far more. The median hides a heavy tail.
  • Image and video cost far more than text, as Luccioni et al.’s measurements and the report both say.
  • Inference dominates the total because it is repeated billions of times. This is why her advice targets use, not training.
  • Forecast risk. The 945 TWh figure is a forecast. The IEA itself says uncertainty widens after 2030. Forecasts of AI demand have been wrong in both directions before (Masanet et al. showed earlier doom forecasts for data-centre energy missed because efficiency kept pace).

Where others push back.

  • Per-query numbers cut both ways. Companies report small per-prompt numbers (Google); critics say these exclude embodied costs, training and the system-wide build-out. Independent measurement is thin.
  • Counting method. “Footprint of electricity” depends on the grid average used. Marginal emissions (what the next kWh comes from) can be higher than average ones when new demand is met by gas.

What she leaves out. Water from cooling (said), uncertainty ranges (not given in the talk), and the rebound dynamic in the volume forecast.

Check yourself

  • Why is the report’s water number roughly “energy times 10 litres per kWh”? What would change if the grid is wind-heavy?
  • Why is Google’s 0.26 mL per prompt not comparable to the report’s implied 4 mL?
  • Why does she say inference, not training, is the main target for individual action?

3. Local trade-offs and who decides [14:31]–[16:20], [42:39]

🎙 In the talk. The host asks who gets to decide which footprint to reduce when “you can’t have all three” [13:48]. Aczel: that’s the $64,000 question … very context specific and localized … there’s no one-size-fits-all [14:31]. Her worked example: “if a certain community is water stressed but has really abundant solar energy, maybe it’s better to … have a larger land footprint if there’s abundant land available but water is a point of stress. Generally lower carbon is better” [14:31]. She admits the report “were looking at countries overall. We didn’t get the granularity of community by community” [15:33]. The fix: “engaging early and often with local communities. Often these are things that aren’t captured in national electricity generation metrics. Traditional knowledge, community experience …” [16:20]. At the end she adds that priorities vary: “if you’re in a severely water stressed area, water impacts are likely going to be the most impactful there” and that people also worry about “noise … pollution” [42:39]. “The greatest wealth of knowledge is going to come from the people who live there” [42:39].

📚 Deeper theory

  • Multi-criteria decision-making. When goals conflict (carbon, water, land, cost), there is no single best answer; there are weights. Methods (multi-criteria analysis, deliberative valuation) exist to make the weights explicit. Her proposal is that local people supply the weights. This is a political claim about legitimacy, not a technical one.
  • Scale mismatch. Carbon is a global commodity, so national numbers work. Water is basin-scale; land is parcel-scale; noise is neighbourhood-scale. A national-average study (like hers) is structurally blind to the scales where most conflict occurs. She says so [15:33]; the broader field says the same (Aqueduct-type basin tools exist for this reason).
  • Cumulative impact. One data centre may be fine; ten in one county may not. Her recommendation (in the press release) of “cumulative impact assessments” for operators is a standard idea in environmental law: judge the stack, not the single project.
  • Ireland as a case. Data centres were 21% of Ireland’s metered electricity in 2023, and the grid operator paused new approvals around Dublin until 2028 (press release). This is a grid constraint turned into a siting rule.
  • Participation is not “a tick in the box.” She had said this about her Berkeley work [05:29]. In theory: Sherry Arnstein’s “ladder of citizen participation” (1969) separates tokenism (informing, consulting) from real power-sharing (partnership, control). “Early and often” is a claim about where on the ladder communities should be.
  • Local knowledge as data. Water stress at particular times of year [16:20] is something farmers and residents know and national data lack. Citizen science and participatory monitoring are the methods.

Where others push back.

  • Efficiency of siting. Developers argue that speed matters (AI race, grid interconnect queues) and that long consultations slow needed capacity. Her counter: fast-track only where the public benefit is clear [18:09].
  • Who is “the community”? A host town may want the tax revenue and jobs; neighbours may not. Local democracy has internal conflict. She does not discuss how to resolve it.
  • Global vs local. Local optimisation can export harm (build where people have least power). Environmental justice scholarship calls this a “path of least resistance” in siting.

What she leaves out. Concrete mechanisms (who convenes consultation, who pays for it, what happens if the community says no). Also benefits that communities may want: jobs, tax base, grid upgrades, waste-heat use.

Check yourself

  • Why is a national-level footprint study blind to the problems that matter most to residents?
  • In her water-stress/solar example, which footprint does she trade against which, and why does she say “generally lower carbon is better”?
  • What would “meaningful engagement” look like on Arnstein’s ladder?

4. Extraction, justice and who pays [18:09]–[23:02]

🎙 In the talk. Asked about patterns from fracking, Aczel says there are “certain applications of AI where we could say these are for clear public benefit … that maybe should be fast-tracked and other technologies or applications that … warrant a little bit more assessment” [18:09]. For “an extractive technology that has clear environmental implications that are physical and really related to how close you are to the area … it’s harder to justify that kind of rapid fast-tracking” [18:09]. Climate justice, in her words, is “who benefits, who’s paying the cost, where” and the “geographical distinction” between them [19:23].

At [20:16] the captions blur the speakers. The host raises Alexandria Ocasio-Cortez visiting communities near data centres with “dark brown” tap water, and a small number of companies in the Bay Area holding control. The next part is most likely Aczel: she sees “the same type of extractive patterns”; she says AI can “entrench a slow decarbonization because of this massive demand”; she contrasts “extractive energy” with AI, where “there’s been incredible developments whether it’s genome sequencing or tornado forecasting”; and she warns that “certain companies are able to buy up electric grids, pay off people, sign a lot of NDAs, especially in the US like with fracking” [20:16]. Which speaker said which sentence is not certain.

📚 Deeper theory

  • The “resource curse” and extractive institutions. Economists (Sachs and Warner, 1995; Acemoglu and Robinson, Why Nations Fail, 2012) show that extraction tends to enrich a few and weaken local institutions unless rules force benefit-sharing. “AI as extraction” is an analogy to this: concentrated gains, dispersed costs.
  • Concentration of control. Compute is held by a few firms and two countries (press release: over 90% of AI-specialised capacity in the US and China). Karen Hao and Crawford both argue that AI’s material footprint is inseparable from this political concentration.
  • Her claim of “entrenching slow decarbonisation.” This is a prediction. The mechanism: demand growth outpaces clean supply, so utilities keep gas and coal plants open longer. There are 2025–26 reports of utilities delaying coal retirements for load growth (from memory; I did not verify specific cases). The counter-mechanism: tech firms buy clean power and finance new nuclear and geothermal, which can speed clean supply. The IEA treats both as live.
  • NDAs. Data-centre developers often use code names and non-disclosure agreements with local officials to avoid land-price spikes and competition between towns. Critics say this blocks public debate before deals are signed. In fracking, private settlements with confidentiality clauses limited the public evidence base. Her point that NDAs affect what the public can learn ties back to the transparency theme.
  • Public benefit and fast-tracking. Her middle position (fast-track the clear-benefit uses, assess the rest) needs a test for “clear public benefit.” This is a hard design problem: who judges? Tests used in other areas (public-interest tests, proportionality) exist but are not specified in the talk.
  • Slow violence (Nixon): noise, vibration, water quality shifts and heat arrive gradually. The host’s cascade [44:25] (noise → birds → insects → pollination) is the ecological form.

Where others push back.

  • “AI is not extraction like fracking.” Defenders say that a data centre does not remove a resource from the ground; it uses electricity and some water, adds tax revenue and jobs, and benefits can be local. Aczel herself says “I don’t think that with extractive energy technologies we can say that there is an overall social or public good to this use whereas with AI I think there’s been incredible developments” [20:16]. So her view is mixed, not a plain analogy.
  • On the brown-water stories. The Georgia case is disputed: Meta said a study found an effect on groundwater “unlikely” and families sued in 2026 (sources in Part 1). The host’s version is a story told in conversation, not a measured finding. Aczel does not endorse it as causal.
  • On “slow decarbonisation.” A forecast, with evidence on both sides.

What she leaves out. Evidence for her extraction analogy beyond her fracking experience; what strong protections would be; the role of the global South (the report notes over 150 countries have little sovereign compute), including India, where data-centre build-out and water stress overlap (my addition; not in the talk).

Check yourself

  • What is the difference between distribution, recognition and participation in justice theory? Find each in her answers.
  • Why does she treat fracking and AI as similar in pattern but different in public benefit?
  • What does “fast-track only clear public benefit” require that the talk does not supply?

5. The nutrition label and disclosure [19:23], [23:44], [34:13], [38:40], [42:39]

🎙 In the talk. “I think we need something like a nutrition label for AI, but that highlights … there’s no judgments or decisions that are made on nutrition labels. You pick them up and you get the information” [19:23]. It should show “where all the different parts are coming from, how they’re sourced, what the electricity and water and land footprints are” [19:23]. She adds that it is “really about information and empowering people” [19:23].

On feasibility [23:44]: she suggests real-time combination of grid data (Electricity Maps), query routing, and company claims of renewable use: “when companies claim to be running on renewable electricity, it would be great for people to see where that renewable electricity is coming from.” She welcomes the UN AI Environmental Transparency Initiative [34:13], and says disclosure should include “where that energy is coming from, what the entire supply chain looks like” [35:00]. Her one requirement for every AI company: “sharing information where the electricity is coming from, where the data centers are located”, plus real community engagement [38:40]. The label would show “this query was processed at a data center in this location. The electricity use was provided by this type of source” [38:40]. The host adds “a fair trade symbol” [41:44]. At [42:39] she says “disclosure is probably one of the biggest” priorities.

📚 Deeper theory (see 2.5)

  • Disclosure as a policy tool. It works when information is simple, comparable and used. A nutrition label works because food is bought many times a day by people who can compare shelf items. AI is different: users often cannot pick the data centre, and many get AI inside other products.
  • Who is the label’s audience? (a) Consumers, (b) enterprise buyers and procurement teams (who can set rules and may be the largest lever), (c) regulators and grid operators, (d) communities, (e) investors. The report lists all of these. Aczel in this talk talks mostly to (a) and (d).
  • Fair trade vs nutrition label. The host’s analogy is different. Nutrition labels describe; fair-trade labels certify against a standard. Aczel prefers description with no verdict [19:23]. A certification would need a standard, auditors, and a view on what is “acceptable” — the very judgment she wants to leave to people.
  • Hard parts. (1) Allocation: a data centre serves many models; whose footprint is which? (2) Marginal vs average carbon. (3) Temporal granularity: hourly or annual. (4) Embodied impacts. (5) Commercial secrecy and security. (6) Gaming: renewable certificates may not match real-time power. Her own difficulty finding data [38:40] is the evidence for the problem.
  • The UN initiative asks for voluntary measurement and disclosure by major firms and a renewable-power commitment for data centres by 2030 (UNU-INWEH). Voluntary schemes have a mixed record (see the history of corporate carbon reporting); mandatory ones (EU) have more teeth (from memory).
  • Real-time labels need standard data feeds. Carbon-intensity APIs exist. What is missing is the firm’s side: which region ran this request.

Where others push back.

  • “Labels don’t change behaviour.” Evidence is mixed: calorie labels have small effects; energy labels on appliances shifted markets. A label that is read once and ignored has little effect. Defaults and rules may matter more (2.5).
  • “Information shifts blame to users.” Maniates’s critique. Aczel says she does not want guilt [36:53].
  • “Industry should be regulated, not labelled.” Mandatory caps (tokens, resolution) are in the report; the talk stays on information.

What she leaves out. Enforcement, auditing, penalties, and the case for binding limits. (The report’s demand-side caps are not discussed in the talk.)

Check yourself

  • Why does she insist the label contains “no judgments”? What is the advantage and the risk?
  • Who, besides ordinary users, could use a label to change outcomes?
  • What data would a firm have to publish for a “this query ran in X” label to be accurate?

6. What individuals can do — prompts, models, timing [23:44]–[26:14], [35:32]–[41:03]

🎙 In the talk.

  • Timing. “If it’s something that you want to get an answer for tomorrow, you could queue it and then get your answer tomorrow morning and it could run off renewable electricity at 3:00 a.m. when there’s much less demand on the grid” [24:58]. (The host extends it to trip planning [24:58].)
  • Prompt discipline. “Knowing what you want to get as an answer first and creating bounds around that answer,” such as “five bullet points” or “three sentences” [35:32], [36:53]. This “reduces a lot of the unnecessary back and forth”, and at 2.5 billion prompts a day, “that really adds up” [35:32].
  • Resolution. Start images in low definition; regenerate in high only for the final [36:53].
  • Right tool. “Intentionality and choosing the right tool and the right model for the task” [36:53]. Her cookbook example: use a cookbook for recipes, a chatbot photo when something goes wrong [37:44].
  • Lighter models. Use “an instant answer” instead of a “thinking model”; lighter defaults are generally better [40:20], [41:03].
  • Framing. “It’s not a question of guilting people” [35:32]. It is an “amazing new technology” [36:53].

📚 Deeper theory

  • What actually moves the total. Per-user choices affect the marginal load. Impact is bigger when choices are made at scale by defaults (which model answers a search query, whether AI summaries run automatically) than by individuals. Her lighter-defaults point [41:03] is therefore stronger than the prompt tips.
  • Is shorter prompting really cheaper? Output length drives cost; input length matters less per token but a long context has costs. Rounds of back-and-forth do add inference. But the biggest swing is model class: reasoning models can spend many times more tokens than non-reasoning ones for the same question (general knowledge; not verified here).
  • Carbon-aware timing (2.7). Works best when the grid has strong daily swings (solar-heavy midday; wind-heavy night). Where the grid is dirty at all hours, it helps little. Her 3 a.m. claim assumes a grid that is cleaner at night. That is not true everywhere (in solar-heavy grids, midday is cleaner).
  • Substitution check. Her cookbook example [37:44] is about not using AI where a lower-cost option works. A limit: books, search and AI have their own footprints; the comparison is intuitive, but I have no figures for it.
  • Rebound. If these tips lower the cost of each use, total use might rise (2.6). Individual efficiency is not the same as total reduction.
  • Behavioural context. Pro-environment tips have small effects when infrastructure does not support them. The host’s tap-running analogy [38:40] fits the moral register but not the scale: running the tap wastes a measurable amount; one prompt does not, which is why disclosure and defaults matter.

Where others push back.

  • Individual action is a distraction (Maniates; and Masley in a different way: he says personal prompt-shaming is not worth the effort, as other choices matter far more).
  • User time and quality. Constraining outputs can reduce quality, forcing retries. Net savings depend on the task.
  • No visible feedback. Without a live meter, people cannot learn which actions save energy. This is her label argument again.

What she leaves out. Numbers for each tip; what share of use comes from consumers vs enterprise and automated systems (agents, search, background tasks); how defaults are set.

Check yourself

  • Which of her tips works by intensity, which by timing, and which by tool choice?
  • Why might “queue it for 3 a.m.” not reduce carbon in a solar-heavy grid?
  • Why are defaults potentially more powerful than tips, and who sets them?

7. AI as a climate tool — weighing benefit against cost [26:14]–[32:15]

🎙 In the talk. The host asks how to weigh AI’s climate benefits against its costs [26:14]. Aczel points to a colleague’s report (published “a couple of months ago”) on “the power of AI for climate change impact assessment” [28:21]. Her points: climate data collection and processing are slow at human scale; AI is “flexible”, “efficient computationally”, and “your large language models don’t need to sleep” [28:21]. It “hasn’t really been fully utilized” in this field [29:30]. Examples: wildfire modelling from historical conditions in “new conditions” [29:30]; sea-level rise [29:30]; “This Climate Does Not Exist”, where you “input … any address in the world and it will project what that location would look like under certain climate change context” [29:30], to help people “visualize … a future that hasn’t yet happened but very easily could” [30:59]. She says AI could help “encourag[e] climate friendly behavior at larger scale and … targeted messages” [31:32]. She notes the project came from “machine learning engineers who are trying to promote … the environmental benefit of AI” [32:15]. On insurers she says plainly: “I don’t have very much information” [32:15].

On benefit vs cost: “I’ve seen some studies that compare the efficiency of an AI model versus the efficiency of a human … AI is a lot lower carbon, water, land footprint than a person” [26:14]–[27:19]. She says the ability to process data “does have massive … environmental benefits” [27:19].

📚 Deeper theory (see 2.8)

  • Three-level accounting (Kaack et al.). To weigh benefit and cost, count (1) compute-related impacts, (2) the direct effect of the application, and (3) system effects (rebound, lock-in, who gains). Most “AI for good” arguments stop at (2). Most “AI is bad” arguments stop at (1).
  • Not all “AI for climate” is a large-model use. Weather and flood models are specialised and may cost far less energy per forecast than the older physics models after training; they are not chatbots. The report is about the large-model, high-use side. So “AI is good for climate” (specialised science models) and “AI’s footprint is growing” (general-purpose generative AI) can both be true and are largely different parts of the AI sector. This distinction is my addition; Aczel does not draw it explicitly.
  • Who benefits from the benefit? Forecasts and tools help most where people can act on them. Data and compute are scarce in many vulnerable regions (the report’s point on sovereign compute).
  • Behaviour-change uses. A tool that shows your own street under flood [30:59] is communication, not mitigation. Evidence on whether imagery changes behaviour is mixed (visualisation can raise concern without changing action; from memory).
  • The “AI vs human” comparison. See 2.8: it compares per-task footprints and only counts the substituted effort. If AI use adds to output without replacing work, total footprint rises. And one human’s footprint exists regardless of whether they write that page.
  • Beyond climate science. She also mentions genome sequencing and tornado forecasting [20:16]. These are claimed benefits. The strength of each depends on the application.

Where others push back.

  • “Benefits are speculative; costs are measured.” Critics (including parts of Kaack et al.’s framing) say the measured costs are mostly of consumer chatbots and image/video, while the climate benefits come from a smaller set of specialised systems that are not the main drivers of growth.
  • “Compare the right things.” Ligozat et al. (2022): many climate-AI studies omit the AI’s own LCA.
  • Optimists (including parts of the IEA’s analysis) argue that wide use of existing AI applications could cut emissions, but not automatically (from memory; not checked).

What she leaves out. Any quantified net balance; which AI uses drive growth (chat, image, video, agents) vs which deliver climate gains; and the weak evidence she herself admits about insurers [32:15].

Check yourself

  • Why can “AI helps climate” and “AI’s footprint is growing” both be true?
  • Where do the three levels of Kaack et al.’s accounting appear in her answer, and which is missing?
  • Why is a per-task comparison of AI and a human not the same as a net comparison?

8. The coda — empathy, and a zero [45:00]–[52:14]

🎙 In the talk. This is the host’s standing closing question: a time Aczel received empathy [45:00]. She describes the support from mentors and others after her father died suddenly during her master’s, just before her PhD [45:35]. What mattered was that kind of recognition of … we understand what you’re going through and you’re a human … outside of what you’re doing [45:35]. She says it showed her that “you don’t need to disconnect your research from you as a person” [45:35]. When asked about him [48:43], she says he was a mathematician whose last project was the origin of zero, which took him “on this … Indiana Jones type exploration around the global south” and led him to a 1929 paper about a zero in Cambodia that was “presumed destroyed by the Khmer Rouge”, then found “with the help of a whole team of archaeologists” and now in the National Museum in Phnom Penh [48:43]–[50:02]. “It’s a zero as a placeholder and it looks like a dot” [50:02]. She says the dissertation was dedicated to him [51:29].

📚 Context (not her claim)

  • The stone is K-127, from Sambor; its date is 683 CE. The recovery is told in Amir Aczel’s Finding Zero (2015); the match is by story and surname, and I did not find a source stating the family link (not confirmed). The paper is by Georges Cœdès, 1931 (see warnings). Mathematicians note that the stone shows a zero as a placeholder in a place-value number, an early written instance, not the origin of the idea of zero (MAA / Borwein note).
  • How this links to the talk: it is an echo of her method. A thing “lost” in an unrecorded place, found by going to the place and asking local people. The episode’s theme (empathy) is the frame; the environmental content ends at [45:00].

Check yourself

  • Does the empathy theme change anything in her argument? Where does it show in her answers on community voice?

Putting it together

The mental model in one paragraph. AI is not a cloud; it is a factory chain: mine → chip → building → grid → query → e-waste. Every link has its own costs in carbon, water, land and people. Counting only one cost (carbon) makes the chain look cleaner than it is, because fixes in one place often push harm to another (burden shifting). Which trade-off is best is local (it depends on the basin, the grid, and what people there value), so the people who live there must have a say (justice). To let anyone decide, the facts must be public: a simple, judgment-free “nutrition label” with where the power came from and where the query ran. With facts, individuals can pick lighter tools and timing, companies can be pressed to disclose, and communities can bargain. Efficiency alone will not cap growth (rebound), so rules on use may be needed. And AI is also a useful tool for climate science: the aim is to count both sides fairly, not to stop it.

Four tensions to keep open

  1. Information vs binding limits. Her answer is disclosure. The report adds caps on tokens, resolution and length. Which should come first, and does a label change anything without rules?
  2. Individual tips vs system change. She gives prompt tips and also says it is not about guilt. Defaults and procurement may matter more than users’ habits.
  3. Local control vs global fairness. Communities should decide, but local choices can export harm to places with less power, and compute is held by two countries.
  4. Benefit and cost are different parts of the sector. Climate-science AI is a small, specialised use; the growth in demand comes mainly from general-purpose generative AI (my reading, not hers). Counting them together may mislead.

Things to verify before you quote this talk

  • The report’s full method, water factor and land boundary (I read only the press release).
  • Whether the “AI is lower footprint than a human” studies she cites apply to her intended comparison.
  • Any claim from the host (AOC, brown water): treat as unverified in this transcript.

Reading path

  • Start: The UNU-INWEH press release and then the report. Google’s 2025 Gemini footprint paper (see Part 1). Ren et al., “Making AI Less Thirsty” (arXiv).
  • Then: Crawford, Atlas of AI (2021); IEA, Energy and AI (2025); Luccioni, Strubell and Crawford, “From efficiency gains to rebound effects” (arXiv); Kaack et al. (2022) on AI and climate mitigation.
  • Deeper: Nixon, Slow Violence (2011); Schlosberg, Defining Environmental Justice (2007); Fung, Graham and Weil, Full Disclosure (2007); Sorrell’s rebound review (2007); Maniates (2001); Hao, Empire of AI (2025).

Q&A

(Add questions and answers here only if worth keeping.)

Graph View

Backlinks