Sustainable AI or Just Efficient? The Real Impacts of AI

Photo by Daniil Komov on Unsplash

It’s not a lie that humanity loves a quick fix; efficiency is a driving force in most modes of life, in business, in leadership and our day to day routines. Do more with less, limit waste and maximise output. Artificial Intelligence (AI) is shaping our view of efficiency, especially from an environmental standpoint, but do faster results combat the extremities of energy drained by using AI? Recent research questions this and the answer is simple – efficiency is not enough to make machine learning environmentally sustainable. It is, in fact, the framing of the issue that is causing the largest problem, feeding our expectations of efficiency, ignoring the reality.

Efficiency is but the tip of the iceberg. When machine learning is related to sustainability, it is laced with technological jargon, quantified in bits per joule and flops per watt, or how much training time has been decreased through its application. What is missing from the narrative is that while models may use fewer cycles per training task, it does not use less energy overall. Even more significantly, efficiency improvements can paradoxically increase total emissions, cheaper computation requires more experimentation, larger datasets and more frequent model retraining, overall expending more energy, not less. Lastly, consider the larger broader environmental impacts, such as rare earths mined for hardware, supply chain carbon from global server manufacturing, the invisible effects in the energy equation.

Being more efficient does not equate to being more sustainable, it would be similar to state that climate change has been solved by reducing car engine emissions by 10%, but the quantity of cars on the road has doubled.

In Wright and colleagues’ paper, the researchers argue that it would be beneficial to adopt ‘systems thinking’; recognition of the entire lifecycle of AI systems, from data collection to training, to deployment to the production of hardware and, finally, disposal. Recognising the larger system leads to recognition of the wider context; what is sustainability? What information is being fed to us and what is being omitted? What happens if the feedback loop continues, when cheaper computing results in larger models?

The underlying mechanism is that sustainability is not a checkbox, but a relationship between technology and society, between intention and impact and between ambition and consequence. It is difficult to separate the two, the claim that more efficient models creates more favourable consequences, and it feels actionable. But what if sustainability actually requires a trade off? Perhaps it is a case of slowing down instead of speeding things up, rethinking incentives and questioning why we would even need larger models.

In leadership coaching it’s called shadow work, the parts of the self that are unwanted and ignored. In sustainability, efficiency is the shadow we hide behind; something that looks good, feels good, and can be done without confronting why we are choosing the efficient route. It is a question of which problems are being solved; are they the right ones or just the easy ones? Optimising for metrics that don’t capture what really matters or KPIs that track activity but not influence may seem efficient in theory, but is wasted energy. The sustainability of AI works the same way, like a trap, optimising energy per operation but not in the grand scale.

The implication of this research is not just that AI is harming the planet, it is that we repeat old patterns with new technology, using intelligence as a substitution for wisdom. Before sustainability, it must be acknowledged what kind of future is being built. It is not a technical issue – it’s a human one.

From a depth psychology perspective, obsession with efficiency reveals the collective shadow, not what is evil, but what is not acknowledged. We are drawn towards growth, control and acceleration with compulsion. Efficiency is a socially acceptable mask that feeds the compulsion to continue. The comfort in the story that cleaner, faster and cheaper AI justifies its expansion, but in actuality it is a projection, a displacement of our own responsibility, our human values, into optimisation, replacing ethics with engineering. The shadow migrates into larger models, denser infrastructures and more extractive supply chains.

Technological rationality becomes a defence mechanism, a tool for avoiding deeper existential questions. Why must everything scale? Does intelligence need to be externalised? Is ‘more’ truly progressive?

When leaders develop, the shadow work requires slowing down long enough to tolerate discomfort, so the same must be applied to the technological aspects we are growing to rely on. While it feels counterintuitive, attitude change requires a psychological shift, accepting the paradox that something can be both innovative and destructive. Sustainability then becomes less about carbon accounting and more about maturity and understanding limits; recognising that not every capability must be maximised and not every efficiency gain is worth exploiting.

For the full paper published in arXiv, 2025, find the link here: [2309.02065] Efficiency is Not Enough: A Critical Perspective of Environmentally Sustainable AI

Follow The Heretic for more psychological viewpoints on the effects of AI in both personal and professional life