A strong sustainability approach to AI development
Artificial intelligence is often framed as a force for boundless progress, a technology that can supercharge economic growth, accelerate scientific discovery, and transform everyday life. But a growing body of research, including new work by Prof. Daniel O’Neill and Prof. Felix Creutzig, asks a more fundamental question: What kind of future is AI actually creating, and is it sustainable?
Beyond the “cost–benefit” mindset
Most discussions about AI’s impact focus on trade-offs. We weigh economic gains against environmental costs, or innovation against social disruption. This approach, often grounded in traditional cost–benefit analysis, assumes that different types of value; financial, environmental, social, can be reduced to a single metric and balanced against each other.
But authors of the recent Nature Machine Intelligence article argue that this way of thinking is no longer sufficient.
Some things, they suggest, should not be traded off at all. A stable climate. Social cohesion. Human wellbeing. These are not interchangeable with financial gains or technological progress. When we reduce everything to a single value, we risk obscuring the real stakes, and ignoring irreversible damage.
A stronger definition of sustainability
Instead, the paper calls for a “strong sustainability” approach to AI.
At its core is a simple but powerful idea:
Human prosperity must be achieved within the limits of the Earth’s systems – not at their expense.
This perspective recognises that natural systems, from climate stability to biodiversity, have critical thresholds. Once crossed, the consequences may be severe or irreversible. Similarly, societies depend on foundational conditions – such as equality, employment, and social cohesion – that cannot simply be replaced or compensated for.
The Doughnut: a compass for AI futures
To make this vision practical, the authors turn to a framework known as the Doughnut.
The outer boundary represents planetary limits including climate change, water use, pollution, and biodiversity loss. The inner boundary represents the minimum conditions for a decent life; access to health, education, energy, and participation in society. Between these two boundaries lies a “safe and just space” where humanity can thrive.
Applying this to AI reveals a more complex picture than the usual narratives of progress.
AI might:
- Improve healthcare, education, and connectivity
- Enable more efficient use of land and resources
But at the same time, it could:
- Increase energy demand through large data centres
- Exacerbate inequality and job displacement
- Undermine social cohesion if benefits are unevenly distributed
Improvements in one area do not justify damage in another.
Seeing the system, not just the technology
Another crucial shift is scale. Many current frameworks for “responsible AI” focus on individual systems, organisations, or use cases. While important, they miss something larger: AI operates within complex economic and social systems.
Its effects ripple outward, shaping patterns of consumption, labour markets, and global resource use. For example:
- Efficiency gains can lead to rebound effects, increasing overall consumption
- Automation can reshape employment and income distribution
- Infrastructure expansion can drive long-term environmental pressures
To understand these dynamics, the authors point to emerging modelling approaches that integrate economic, environmental, and social factors – moving beyond narrow metrics like GDP or carbon emissions.
Choosing the future we want
Ultimately, the article turns a common assumption on its head. Rather than asking how AI will shape our future, it suggests we should start by defining the future we want:
- A society where everyone can meet their basic needs
- An economy that operates within planetary boundaries
- A world where technological progress strengthens, rather than undermines, social foundations
From this perspective, AI becomes a tool, not a driver of change. The goal is not to adapt to AI’s trajectory, but to steer it.
This reframing has profound implications.
- Policymaking must go beyond efficiency and growth, focusing on limits and thresholds
- AI governance should explicitly address inequality, social cohesion, and environmental ceilings
- Innovation should be evaluated not just by what it enables, but by whether it keeps us within a safe and just operating space
The message is clear: Technological advancement alone is not enough. Direction matters.
