Chapter 4 · Accountability and Society
Environment, Labour and Inclusion: The Wider Costs and Benefits of AI
- Page 19 of 22
- 19 min read
So far the people in our examples were close by: a customer, a pupil, a colleague. AI also has costs and benefits that are far from the screen. A chat answer is produced in a data centre that uses electricity. Some of the data a model learned from was sorted, labelled or checked by people whose working conditions rarely appear in the product description. And the quality of the answer you receive can depend on which language you write in. These questions are real, and they are also hard to measure. This page teaches you to ask them calmly and to be honest about what is uncertain.
A word about numbers. Headline figures about energy, water and pay change fast, depend on how they are measured, and are often repeated without their conditions. This page therefore teaches direction and boundaries: which way a cost points, what a figure does and does not include, and how to ask for evidence. It does not print headline numbers. Where a point rests only on news reports, we say so.
What you will learn
- What kinds of energy and resource cost an AI service can have, and how training, inference and data-centre demand differ.
- When a small model can be enough, and how to ask for an environmental claim to be backed by comparable evidence.
- Why energy figures from different systems are often not comparable, and how to say so politely.
- Who does the human work behind data collection and labelling, and which questions point to fair working conditions.
- How language and regional gaps appear in AI products, and how to compare open and closed model arguments without choosing a side.
Energy and resource costs of AI use
Every AI answer uses physical resources somewhere: electricity to run computers, and the buildings, networks and equipment around them. The International Energy Agency (IEA) published a report in April 2025, Energy and AI. Its overview states that AI needs electricity for data centres, and the report models how AI electricity use may grow over the next ten years, where that electricity could come from, and what that means for energy security, emissions, innovation and affordability. The direction it describes is that data-centre electricity demand is growing and AI is a major driver. The projections come with large uncertainty. We do not repeat figures here, because they depend on the model's assumptions.
Water is part of the picture too. A 2023 research paper by Li, Yang, Islam and Ren argues that the water footprint of AI has received less attention than its carbon footprint. It distinguishes water withdrawal (water taken from a source) from water consumption (water that does not return to the source, for example because it evaporates). It also notes that water efficiency varies by place and by time. The paper gives estimates, but they rest on modelling assumptions, so treat them as a sign of the question rather than a settled number.
Source note (as of 2026-10-12): IEA, Energy and AI (published 2025-04-10), overview page at iea.org; Li, Yang, Islam, Ren (2023), arXiv:2304.03271. Both are research or analysis documents, not laws. Check the current versions for any figures you plan to use.
What to take from this: an AI service has a physical footprint, it is not weightless, and the size of that footprint is uncertain and depends on how the service is built, where it runs and how much it is used. Saying "AI is clean" or "AI is wasteful" in one sentence goes beyond what the evidence supports.
Training, inference and data-centre demand
When people talk about the "energy cost of AI" they may mean three different things. Mixing them is the most common source of confusion.
| Term | What it means | When it happens | Question to ask |
|---|---|---|---|
| Training | Building the model by running it over a very large amount of data. | Before release, and again for each new version. | Is the figure for one training run or for all the experiments before it? |
| Inference | Using the finished model to answer one request, such as your question. | Every time anyone uses it. | Is it per request, per 1,000 requests, or per day? For what kind of task? |
| Data-centre demand | The electricity (and, as shown above, water) needed to run the whole facility that hosts training and inference. | All the time, including when machines wait for work. | Does the figure cover only the chips doing the task, or the wider facility? |
Research gives a sense of why the distinction matters. Strubell, Ganesh and McCallum (2019) estimated that training large language models for text can carry a substantial energy and financial cost, and recommended ways to reduce it. Luccioni, Jernite and Strubell (2023) looked at the other side: ongoing use. They measured energy and carbon per 1,000 uses on benchmark tasks and reported that multi-purpose generative models used far more than task-specific systems for a range of tasks, in their words "orders of magnitude" more. Their study concerns specific models and hardware, so it shows a direction rather than a rule for every tool. The practical lesson is that a one-time training cost and a cost that repeats with every use are different things, and the second grows with the number of users.
Small and large models
A larger model can handle a wider range of tasks and harder ones. A smaller model is usually cheaper to run, faster, and sometimes can run on your own device. Many everyday jobs are narrow: fix spelling, sort messages into five labels, pull a date out of a form. For narrow jobs, the study above suggests that a task-specific system can cost far less per use than a general one. That is a reason to ask "is a smaller tool enough?", not a rule that small is always better.
| Question | Points toward a smaller model | Points toward a larger model |
|---|---|---|
| How narrow is the task? | One clear job repeated often (labels, spelling, extraction). | Open-ended work with many unpredictable requests. |
| What does a mistake cost? | Low cost, easy to check. | Hard cases where a weak model would fail often (check this by testing). |
| Where must the data stay? | Sensitive text that should not leave your device or team. | No such limit. |
| How much will it be used? | Thousands of requests a day, where small costs add up. | Occasional use. |
| Is the quality enough in your language? | Test it first; a small model may be weaker in some languages. | Test it too; size alone does not guarantee quality in your language. |
The only reliable way to answer these is a small test with your own examples, as you learned on Policies, Vendors, Records and Incidents: Responsible AI in a Small Team.
Asking for environmental claims to be supported
You will meet claims such as "carbon-neutral AI", "30% more efficient", or "uses less energy than a web search". Some are accurate; many are incomplete. We have no verified source here for what companies actually disclose, so the useful skill is the question, not a verdict. Ask the five questions below. They are the same whether the claim comes from a vendor, a news story or a social media post.
| Question | Why it matters |
|---|---|
| What is inside the boundary? | Training only? Use only? The whole facility? Equipment manufacture? Water? |
| What is the unit? | Per request, per 1,000 requests, per task, per user per day, or in total? |
| What method and period? | Measured or estimated? Over what days or months? By whom? |
| Compared with what? | Same task, same quality of result, same kind of hardware, same place and grid? |
| Can someone else check it? | Is the method published, or is it a single number in an advertisement? |
The limits of comparing energy figures
Even honest figures often cannot be compared. The two examples below are invented and use made-up units.
Toy data:
Toy data (made-up units, illustrates the problem only):
Vendor A says: "0.3 units of energy per request".
Vendor B says: "2.0 units of energy per 1,000 requests, measured on a lab server".
At first glance A looks 150 times larger than B. But:
- A may measure the whole data centre; B only the chip while it works.
- A's request may be a long document; B's a one-line label.
- A's answer may be a full essay; B's output a single word.
- A is measured in a place with one kind of electricity supply; B in another.
So the two numbers do not describe the same thing, and neither is "wrong".Differences that make comparison hard include: what is measured (the boundary), the task and the length of input and output, the quality of the result, the hardware, the electricity supply and the time of day, and whether the figure is a lab measurement or an estimate for a typical day. The calm response is not "you are lying" but "please tell me the boundary, unit, method and period, so I can compare like with like". Studies such as the one by Luccioni and colleagues compare on the same benchmark tasks and say so, which is exactly the kind of care to look for.
Where reasonable people disagree
How much should an individual user's choices count?
| View A: systems matter more than single choices | View B: individual choices still matter |
|---|---|
| One person's use is tiny next to decisions by providers, regulators and electricity suppliers. Public attention is better spent on how data centres are powered and on honest reporting. | Many small uses add up, and demand signals what gets built. Choosing a smaller tool for a narrow task is cheap, easy and consistent with other habits of care. |
Both views can be held together: ask your provider for evidence (the systems level) and choose proportionately for yourself (the personal level). Some also argue that AI may help reduce resource use in other fields. That is a claim to be examined case by case with evidence, not accepted or dismissed in general.
The human work behind data
Many AI systems rely on people. Some write or select the examples a model learns from. Some label images and text, for example marking which comments are harmful. Some rate model answers so the system can be improved. This work is sometimes done by employees, sometimes by contractors, and sometimes through online platforms, often in a different country from the company that sells the product.
Reports by journalists have described low pay and exposure to distressing material for some data workers. Companies named in such reports have sometimes disputed details, and we have not checked an original report, so we state only the general point: reports describe low pay and difficult material for some of this work. We give no amounts, and we do not claim that all data work is poor. The responsible position is to ask, not to assume.
| Question for a vendor or a supplier | What a good answer shows |
|---|---|
| Who labelled or checked the data, and who employs them? | The supply chain is known, not hidden behind several layers. |
| How is pay set compared with local living costs? | The supplier has a reason it can explain, not just "market rate". |
| What support is there for people who see harmful content? | Breaks, rotation, the right to decline a task, access to counselling. |
| Can workers raise concerns without losing their work? | A real channel, protected from retaliation. |
| Are contracts clear, and are payments reliable? | Written terms, and a way to dispute them. |
Where reasonable people disagree
| View A: this work creates opportunity | View B: this work needs stronger protection |
|---|---|
| Data work can bring income and skills to people with few local options, including in regions where jobs are scarce. Withdrawing the work could leave them worse off. | Distance and many layers of subcontracting can hide poor conditions. Wages that look acceptable on paper may be too low where people live, and some tasks carry real psychological risk. |
The two views meet at one practical point: ask about conditions and prefer suppliers that can answer. Whether a particular arrangement is fair depends on facts about that workplace, not on a general rule.
Language inequality and underrepresented communities
In How AI Works: LLMs, Tokens and Limitations you learned that text is split into tokens. A research team (Ahia and colleagues, 2023) studied how token counts and costs differ across 22 languages. They reported that speakers of many languages can be "overcharged while obtaining poorer results", often in regions where the service is less affordable. The mechanism in plain words: some languages need more tokens for the same meaning, so they cost more and fit less into the model's memory, and the model may also have seen less of them. We cite one study; broader benchmark evidence has not been checked here.
Invented scenario:
Nadia coordinates a district help desk in Dhaka. She tests an AI assistant with 20 questions
about forms, first in English, then in Bangla (toy results, invented):
English: 18 answers good, 2 need fixing.
Bangla : 12 answers good, 8 need fixing; two answers are polite but give a wrong office.
Most visitors to the desk write in Bangla. Nadia's staff speak both languages.This is one small test, not proof, but it shows the habit. Nadia's options include: use the tool only for staff drafting (in English, with review); use it in Bangla with a staff member checking each answer; ask the vendor for evidence in Bangla and wait; or not use AI for this service. The people most affected are visitors who cannot easily check an answer in English, so a good decision gives them a way to reach a person. This links to Fairness and Its Trade-Offs: Inclusion, Accessibility and Language: a tool can be fair on average and still serve one language community less well.
The same questions apply to any community that appears less in the data or the design team: speakers of regional languages and dialects, people who use assistive technology, older people, people with slow or costly internet access, and people whose names, places or customs differ from the examples the tool saw. Ask three things: Who tested it, and with whom? Where does it do worse, and is that written down? Can people in the affected community help shape or review it? These are questions, not accusations. Do not assume a gap exists for a given group without evidence, and do not assume it does not.
Open and closed models: arguments without a winner
"Open" is not one thing. It can mean that the model's parameters (its learned numbers) can be downloaded, that the code is shared, that the training data is described, or that the licence allows reuse. Licences differ, so check the actual terms. A "closed" model is typically available through a service run by its provider. Neither word tells you whether a model is safe, accurate or fair. Here are the strongest arguments, from different settings.
| Issue | Arguments for more openness | Arguments for a closed approach |
|---|---|---|
| Access and cost | A university lab with a small budget, or a start-up in São Paulo, can use and adapt a model without paying per request. | A hospital IT team in Pune may prefer a managed service with a support contract over running its own system. |
| Scrutiny | Outside researchers can inspect, test and report problems, including in languages the provider does not prioritise. | Providers can run central safety checks, update quickly, and withdraw a version that proves harmful. |
| Misuse | Many eyes can find and fix weaknesses; restricting access can concentrate power in a few firms. | Once a model is released, a provider cannot easily stop harmful use, so access limits can reduce risk. |
| Language and local needs | Communities can adapt a model to a local language or task without waiting for a vendor. | A well-resourced provider may fund broad testing and support for many languages. |
| Accountability | Public access to details makes claims easier to check. | With a single provider, it is clear whom to contact and who answers for a failure. |
| Dependence | Less lock-in; you can move or run it yourself. | Self-hosting needs skills and money; a team without them may be less safe running its own model. |
Note how the same fact can support both sides: access helps a small lab, and access helps a bad actor. How the balance falls depends on the model, the task, the licence, the support available and the setting. This page takes no side. Your skill is to name the issue that matters most in your case and ask for evidence on it.
The decision pattern applied
Invented scenario:
Ana runs a language school in Lima. She wants AI feedback on 200 student essays a month.
Option A: a large online assistant. Option B: a small tool running on the school computer.Worksheet, filled in (short):
1. SITUATION Feedback on spelling, grammar and structure for 200 essays a month.
2. AFFECTED Students (essays are personal), teachers, the school.
3. BENEFITS/HARMS Time saved; risk of wrong feedback; student text leaving the school;
resource use grows with volume.
4. OPTIONS A large online assistant. B small local tool. C teacher marks only.
D small tool for spelling, teacher for structure.
5. EVIDENCE Unknown: quality on our students' Spanish and English. Test 15 essays
with each; ask the vendor for the boundary and unit of any energy claim.
6. DECISION Option D if the small tool matches the teacher on spelling; stop if
it misses more than 3 of 15 errors teachers consider important.
7. ACCOUNTABLE Ana.
8. REVIEW DATE 2027-01-15.The environmental part of this decision is small and honest: it is one consideration among several, handled by asking whether a smaller tool is enough and by asking for comparable evidence. It does not claim that either option is "green".
Where this shows up at work
A procurement officer adds three questions to a tender: what the energy figure includes, how data workers are treated, and how well the product performs in the languages users speak. A school chooses a smaller tool for spelling checks. A product manager asks the team to test with real examples in the main local language before launch. A journalist, a student or a manager reads an energy claim and asks for the boundary and the unit before repeating it. None of this requires being an expert in energy or labour; it requires the habit of asking.
Common mistakes
- Treating one energy figure as the whole story. It may leave out training, the wider facility or water. Fix: ask for the boundary, unit, method and period.
- Comparing two figures that measure different things. Fix: compare only like with like, or say that you cannot.
- Assuming "open" or "closed" settles safety or fairness. Fix: look at the actual licence, evidence and support.
- Forgetting the people behind the data. Fix: ask who did the labelling and under what conditions, and prefer suppliers who can answer.
- Testing only in English. Fix: test in the languages your users actually use, and publish the result to your team.
Try it yourself
- Easy. A news post says "Model X uses 5 times less energy than Model Y". Write the five questions you would ask before repeating it.
- Medium. Using the toy data in the energy comparison above, write three reasons why Vendor A and Vendor B cannot be compared, and one message to each vendor asking for what you need.
- Hard. Invented scenario: a regional newspaper in Accra plans to use AI to draft summaries and to label reader comments, with the labelling done by an outside company. Use the eight-step pattern. Include at least one question about data-worker conditions, one about language quality, and one about the environmental claim in the vendor's brochure. Name an owner and a review date.
Answers
Easy. A strong answer: (1) What is inside the boundary: training, use, whole facility? (2) What is the unit: per request, per task, in total? (3) Method and period: measured or estimated, when, by whom? (4) Compared with what: same task, same quality, same hardware and place? (5) Can others check the method? Pass: all five, in your own words, with a reason for at least one. Weak: asking only "is it true?", or accepting "5 times" as given.
Medium. Strong reasons: different units (per request vs per 1,000 requests); different boundaries (whole facility vs chip only); different tasks and lengths (a long document vs a one-line label); possibly different hardware and electricity supply; one is a lab measurement. Strong messages ask for boundary, unit, method, period, task and output length, and ask both vendors to use the same test. Pass: at least three reasons and two precise requests. Weak: concluding that one vendor is better, or that the figures are fake.
Hard. A strong worksheet lists readers, comment authors and the outside workers as affected; options such as no AI, AI for summaries only, or AI labelling with human review of a sample; evidence needed (a test in the main local languages used by readers, vendor answers on worker pay, support and a way to raise concerns, the boundary and unit of any environmental claim); a decision with conditions; a named owner; and a date. It states what is unknown and does not assume the vendor is acting badly. Pass: three or more options including not using AI, each of the three required questions, and honest uncertainty. Weak: a verdict that outsourcing is always wrong or always fine, or no questions about the people doing the labelling.
Return to Policies, Vendors, Records and Incidents: Responsible AI in a Small Team and add three questions from this page (boundary of the energy claim, worker conditions, language evidence) to the vendor questions you wrote there.
Summary
- AI has energy and water costs that are real but uncertain. The direction is growing demand; the size depends on the boundary, method and assumptions.
- Training, inference and data-centre demand are different things. Ask which one a figure covers.
- A smaller model may be enough for a narrow task; test before you decide, and do not assume small or large is always better.
- Ask for boundary, unit, method, period and comparability. Figures from different systems often cannot be compared.
- People label and check data. Reports describe poor conditions for some of this work; ask vendors about pay, support and a route to raise concerns.
- Language and regional gaps are real in some products. Test in your users' languages. Open and closed models both have strong arguments; look at the evidence for your case.
Next: Rules, Uncertainty and Responsible Decisions: Finding What Applies to You shows how to weigh innovation against social harm, find the current official rules for your country and sector, and decide in proportion when you are unsure.