toward sustainable AI?

Artificial intelligence (AI) promises revolutionary advancements across various sectors, such as healthcare, finance, and environmental science. AI's capacity to analyze complex datasets and perform early cancer detection in healthcare, predict stock market trends in high-frequency trading, or optimize energy consumption in large data centers showcases this transformative potential.

However, a critical yet overlooked issue lies beneath this mask: the significant environmental impact of training large AI models. The training process uses expansive datasets and increasingly complex algorithms, which require immense computational power.

---

A data center in Odense, Denmark.

---

This issue highlights the multifaceted problem of unsustainable energy consumption and a layer of inequality when only a few companies can afford to train such large models. This dichotomy between the promise of AI and its environmental impact calls for a critical reassessment of our current approaches to developing AI models. It highlights a need to set up regulations to ensure long-term sustainability. This essay explores the nuances around this issue and motivates the needed actions.

Karen Hao, AI reporter.
Hao calls for better energy sustainability in AI.

The sheer scale of the amount of energy needed to train large AI models such as Generative Pretrained Transformers (GPT) is astounding. It is estimated that training a single model can emit as much carbon as five cars in their entire lifetimes¹. GPT-3, the predecessor to GPT-4, emitted an estimated 126 cars’ worth of carbon during its training phase, yet it pales compared to GPT-4, a model with nearly a thousand times more parameters. The progression from GPT-3 to GPT-4 exemplifies how rapid advancements in AI capabilities correspond with escalations in environmental costs. Unsurprisingly, Sam Altman, OpenAI’s CEO, recently expressed his intent to raise trillions of dollars to procure more computational power for training AI models².

The University of Massachusetts at Amherst conducted a study which found that that training a single AI model could consume more than 626,000 kWh of electricity, equivalent to nearly five times the annual electricity consumption of an average American household³. We need to emphasize these financial and environmental factors and actively push for more energy-efficient AI research and development approaches.

The environmental implications of AI progress echo historical challenges in other industries, such as the automotive industry, which faced similar concerns when the number of personal cars used worldwide increased in the 1950s and early 1960s. In 1965, the state of California imposed emission standards on all automotive vehicles in an attempt to control the situation⁴.

---

OpenAI CEO, Sam Altman, wants to raise trillions of dollars for training AI models.

---

In the years since, advances in engineering led to the development of more fuel-efficient and, eventually, electric vehicles. At the same time, international agreements and national regulations, such as the Kyoto Protocol and the European Union’s emission standards⁵, further incentivized reductions in carbon emissions. This transformation was driven by governmental policies, shifting consumer preferences toward more environmentally friendly vehicles, and intentioned innovations in green transportation methods such as Tesla in recent years.

Similarly, a combination of technological advancements, such as the development of more efficient computing architectures and algorithms, paired with more robust regulatory frameworks and improved awareness among consumers and developers about the environmental impacts of AI, could catalyze a shift toward more sustainable AI. Companies including Google, Apple, and OpenAI have made commitments to power all their operations with renewable energy and reduce the carbon footprints of their AI models.

However, who is holding them accountable? Much of the regulatory framework is nonexistent, which needs to change.

Worse yet, the AI industry prioritizes short-term gains in model performance (accuracy) over environmental sustainability, adopting a "not my problem" stance that ignores broader implications. This attitude ignores the interconnectedness of environmental challenges, downplaying the carbon emissions accompanying the development and training of large models. Schwartz et al. point out that even the minification of existing models—such as Google’s recent reengineering of its Gemini model to Gemma, a model with fewer parameters, hence less computational costs⁶—tends to be motivated by the need to have models be more accessible on smaller computational devices such as personal computers and phones, not the need to have them be more energy efficient⁷. Hovy and Spruit identify this lack of ethical awareness and discourse as concerning⁸.

Prompt action is needed. First, governmental regulation must catch up with the status quo in the AI industry. However, this poses a unique challenge for state and federal governments: many big AI companies have international operations. Global bodies such as the UN could be better suited to set up regulatory frameworks. The European Union’s General Data Protection Regulation (GDPR) could be a good precedent for setting and enforcing such regulations. Nevertheless, the EU has faced difficulties in managing multinational tech giants like Meta, illustrating the complexity of enforcing standards across borders ⁹.

Furthermore, the AI sector must embrace "Green AI," advocating for research that achieves novel results without escalating computational costs, thereby reducing the carbon footprint. This shift towards including energy efficiency as a primary evaluation criterion, alongside accuracy, would ensure that the energy conversation does not get overshadowed by vanity metrics such as a few extra percentage points on an accuracy benchmark, especially when such an improvement comes with a disproportionate reduction in energy efficiency⁷.

---

The European Union has faced unique challenges in trying to enforce regulations across borders.
Any regulatory body will likely face similar issues.

---

The transition to efficient AI also hinges on technological innovation: companies must intentionally seek to create more efficient models, not just more accurate ones. This very closely links to consumer influence. If the user only cares about a model’s accuracy, companies will do their best to improve it. However, users caring about other metrics, such as energy efficiency would force companies to attend to their performance on those other metrics.

The AI industry must urgently balance its pursuit of technological advancement with environmental responsibility. The drive for more powerful AI models results in significant energy consumption and carbon emissions, raising critical ethical concerns. The sector must adopt sustainable practices to ensure that the development of AI technologies does not come at the expense of our planet's health. Action and perspective shifts are needed from all involved: companies, consumers, and regulators.