Artificial intelligence is becoming infrastructure.
It is moving into search, office work, healthcare, education, research, and manufacturing. It is also entering the small gaps where everyday decisions happen: what to do first, which option uses fewer resources, whether a complicated claim deserves our trust.
That changes the question we should be asking. “What can AI do?” is no longer the most important question. The harder question is this: as AI capabilities advance faster than our institutions, energy systems, and social agreements, how do we make sure AI expands human possibility instead of concentrating efficiency, resources, and power in fewer hands?
This is not an argument that AI will save the world. It is not an argument that AI will destroy it either. It is a working hypothesis: AI can become an amplifier for sustainable development, but only after we decide what is worth amplifying.
That question is close to the way we think about AirJelly. A proactive AI companion should not simply create more output or interrupt people with more suggestions. It should help people understand context, reduce low-value friction, and keep human judgment in the loop. The same standard should apply when we think about AI at the scale of society.
01. Progress needs a wider definition
For much of the last few decades, technological progress has been measured through speed, scale, and output. Models become faster. Factories become more automated. Content becomes more abundant. Decision chains become shorter.
Those measures matter, but they are not the same as human development.
The United Nations Sustainable Development Goals describe development as a connected set of responsibilities: ending poverty and hunger, improving health and education, creating decent work, reducing inequality, taking climate action, and building peaceful and accountable institutions. Growth is not a single leaderboard. It is a set of relationships that must remain viable over time.
AI creates a particular challenge because it is very good at optimizing local objectives.
If the goal is to minimize average customer-service response time, a system may reduce human involvement. If the goal is to maximize engagement, it may produce more stimulation. If the goal is to improve predictive accuracy, it may make minority experiences less visible. Every local metric can improve while the overall quality of life gets worse.
Sustainable AI needs a broader definition of success: reliable, fair, recoverable value for more people over a longer period of time.
That means product metrics should include less glamorous but more consequential questions. How much energy and material did the system save? How much repetitive work did it remove? Who gained capability, and who lost choice? When the system fails, who can notice, correct, and challenge it?
The best AI products are not the ones that make people feel surrounded by intelligence. They are the ones that help people spend their attention where it matters.
02. AI can help us see systems. It cannot make the value judgments for us.
Climate and ecological problems are difficult not only because we lack information. They are difficult because the systems are complex, feedback is delayed, responsibility is distributed, and many costs do not appear on a single organization’s quarterly report.
AI is well suited to handling parts of that complexity. It can help researchers combine satellite imagery, sensor data, weather records, and supply-chain information. It can help power grids forecast demand and integrate renewable energy. It can help architects compare materials, energy use, and lifecycle impacts. It can help farmers identify patterns in irrigation, pests, and soil conditions.
The International Energy Agency’s Energy and AI report points to meaningful opportunities for AI in electricity systems, buildings, industry, transport, and energy innovation. It also makes the other side of the equation impossible to ignore: data centres and AI themselves require more electricity.
AI is not inherently green or ungreen. Its impact depends on the system in which it is deployed, the energy that powers it, and the objective it is asked to optimize.
The same model could help a grid reduce peak demand, or it could generate an endless stream of low-value computation and consumption. We should not ask only how much energy a model uses. We should ask whether it helps the wider system save more energy, material, time, or risk than the model consumes.
This is why AI should be treated as an observer, coordinator, and scenario-builder—not as the final judge. It can produce forecasts, reveal trade-offs, and make complex patterns legible. It cannot decide on its own what a river is worth, whether a wetland should be developed, or whose risk is acceptable.
Those decisions still belong to public debate, law, and human judgment.
03. Computation is part of the material world
When we talk about large models, it is easy to see only intangible outputs: text, images, answers, and predictions. Every one of those outputs still depends on physical chips, servers, cooling systems, electricity, land, and water.
Data-centre demand is becoming an infrastructure question. The IEA’s Electricity 2024 analysis projected that electricity consumption from data centres, AI, and cryptocurrency could more than double by 2026. The exact increase will vary with efficiency, regional energy mixes, and demand, but the direction is clear: AI cannot scale independently of the power system.
This does not mean research or useful applications should stop. It means we need a better definition of efficiency.
A larger model does not automatically create public value proportional to its cost. A longer answer is not automatically more useful. Sending every task to the most expensive model is not automatically better product design.
A mature system chooses a model according to the difficulty of the task. It reuses results where possible. It avoids unnecessary calls. It makes important resource boundaries visible. It treats computation as something to spend deliberately.
This is a simple product principle: do not let computation substitute for thought.
When a difficult problem needs careful reasoning, AI is a tool. When a system keeps generating, reminding, and re-evaluating simply to look intelligent, it is closer to waste.
04. Sustainability also means not taking human capability away
If AI is used only to reduce labour costs, sustainable development becomes a euphemism for making fewer people do more work. That is an easy road to take, and a dangerous one.
The International Labour Organization’s 2023 research on generative AI and jobs suggests that generative AI is more likely to transform and augment tasks within many occupations than to immediately eliminate entire occupations. But the effects are uneven across countries, industries, genders, and types of work. Average gains do not tell us who pays the transition cost.
A human-centred AI system should replace the question “How many jobs did we automate?” with more useful questions:
- Did it reduce dangerous, repetitive, or low-value work?
- Did it give people more time to learn, care, create, and exercise judgment?
- Do users understand the system’s limits, or are they simply expected to accept its conclusions?
- Can people affected by a decision appeal, request review, and opt out?
UNESCO’s Recommendation on the Ethics of Artificial Intelligence places human dignity, autonomy, transparency, explainability, accountability, and environmental protection at the centre of responsible AI. These can sound abstract until they become product decisions: preserve human review; show uncertainty; keep a traceable decision history; do not treat silence as consent; let people correct what a system believes about them.
This is also where personal context matters. An assistant that remembers the right things can reduce cognitive overhead. An assistant that silently turns memory into control can do the opposite. At AirJelly, the point of context is not to predict a person perfectly. It is to help a person use their own accumulated context with more agency.
If AI makes people less able to remember, judge, organize, or disagree, it may improve short-term efficiency while weakening society’s long-term resilience.
05 The dividing line is not whether people have AI, but who can use and govern it
The deepest sustainability risk is often not the technology itself. It is the unequal distribution of its benefits and risks.
Advanced computing, data, talent, and infrastructure are concentrated in a relatively small number of countries and companies. Wealthier regions may use AI for climate modelling, medical screening, and personalized education. Under-resourced regions may receive external systems that score, monitor, or classify people without having a meaningful role in their design, audit, or governance.
That can reproduce existing inequality. People with more data receive better services. Organizations with more resources gain more efficiency. Vulnerable groups become more likely to be predicted, categorized, and managed.
An AI sustainability agenda therefore cannot focus only on deployment. It must also build public capability: open standards, affordable infrastructure, multilingual systems, local knowledge, public-interest research, and mechanisms that allow governments, schools, communities, and civil-society organizations to inspect important systems.
Fairness is not a “public good mode” added after a product is finished. It begins with the problem definition, the data, the deployment context, the pricing model, and the way failures are handled.
The same principle applies to personal AI. A useful companion should not turn a person’s context into a proprietary cage. It should make context more usable while leaving ownership, correction, and direction with the person.
06 Making AI a long-term collaborator
If AI is going to support sustainable development, we need principles that are simple enough to use and strict enough to matter.
Start with public value, then choose the model
Do not begin with “Can we add a model here?” Begin with “Which real problem is worth solving?” A system that reduces nurses’ administrative burden should not be evaluated by the same metric as a system that merely adds more content to an already noisy feed.
Put lifecycle cost into product decisions
Cost includes more than an API bill. It includes energy, hardware, data governance, maintenance, migration, training, and the social cost of errors. In high-impact settings, lifecycle assessment should be as important as feature evaluation.
Show uncertainty to the person making the decision
Sustainable systems cannot pretend to be certain all the time. Confidence, data freshness, scope of applicability, and known failure modes should be visible when people act. Saying “I don’t know” honestly is often more useful than giving a fluent wrong answer.
Let affected people participate in design
People who are evaluated, served, or managed by a system should not exist only as rows in a dataset. Teachers, nurses, workers, farmers, disabled people, and community organizations need a real voice in design, testing, and review.
Preserve the right to exit and correct
Not every workflow needs automation. Not every automated workflow should be irreversible. Essential services need alternative channels, human review, and a traceable appeals process.
Measure maturity by what the system chooses not to do
A mature AI product does not hand more and more decisions to a model. It knows which decisions should remain human. Reducing unnecessary generation, notifications, and decision loops is also progress.
07 The final question is still: what kind of people do we want to become?
Technology tends to amplify the way organizations already work.
If we treat education as answer production, work as a set of compressible tasks, nature as an asset to optimize, and people as predictable users, AI will accelerate those assumptions.
But if we treat education as the cultivation of judgment, work as shared creation, nature as a system we must live with, and people as dignified, relational, and partly unknowable beings, AI can help us protect time, see complexity, collaborate across distance, and make room for long-term goals.
Sustainable development has never meant making the world infinitely efficient. It means preserving humanity’s ability to choose, repair relationships, and care for the future on a finite planet.
Whether AI becomes a partner in that work depends on the tasks, boundaries, and institutions we give it. Models will keep getting stronger. But no model will decide for us what is worth doing.
That part is still ours.
References
- Intergovernmental Panel on Climate Change (IPCC). AR6 Synthesis Report: Climate Change 2023. 2023. https://www.ipcc.ch/report/ar6/syr/
- United Nations. Transforming our world: the 2030 Agenda and the Sustainable Development Goals. 2015. https://sdgs.un.org/2030agenda
- International Energy Agency (IEA). Electricity 2024: Analysis and forecast to 2026. 2024. https://www.iea.org/reports/electricity-2024
- International Energy Agency (IEA). Energy and AI. 2025. https://www.iea.org/reports/energy-and-ai
- Gmyrek, P., Berg, J., & Bescond, D. Generative AI and Jobs: A global analysis of potential effects on job quantity and quality. ILO Working Paper 96, International Labour Organization, 2023.
- UNESCO. Recommendation on the Ethics of Artificial Intelligence. 2021. https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence
- United Nations Environment Programme (UNEP). Emissions Gap Report 2024: No more hot air ... please!. 2024. https://www.unep.org/resources/emissions-gap-report-2024
- Stanford Institute for Human-Centered Artificial Intelligence. AI Index Report 2025. 2025. https://hai.stanford.edu/ai-index/2025-ai-index-report
- OECD. OECD AI Principles. Updated 2024. https://oecd.ai/en/ai-principles