Artificial Intelligence (AI) is becoming an integral part of our lives and is transforming various sectors such as healthcare, finance, transportation, and entertainment. Sustainability is increasingly a critical consideration in all sectors. Given the amount of power AI models typically consume, is there an inherent contradiction between developing AI technology and acting sustainably? Those investing in AI will want to ensure that in so doing they don’t impede their path to net zero. New terms of Red AI and Green AI introduced by Schwartz et al[1]represent two distinct principles regarding the development and application of AI systems and those investing in AI would do well to understand the differences.

Red AI: Power and accuracy

Red AI is characterised by an approach that seeks to obtain state-of-the-art results in accuracy or related measures through the use of high-performance computational power.

Red AI systems are designed to maximise performance and aim to deliver optimal outcomes leading to significant advancements in areas such as scientific and medical research, complex problem-solving, and many others.

Take for example, the Human Genome Project (HGP), at a cost of approximately US $2.7 billion and 13 years to map the human genome. The HGP’s outcome was originally viewed with scepticism due to its cost and the lack of immediate scientific breakthroughs. Today, we map an individual’s genome in a few hours for less than US $1,000 using sequencing technology that relies on HGP technology so while the HGP lacked efficiency, it nonetheless helped pave the way for personalised medicine.

Clearly, Red AI will excel in high-stake scenarios, such as autonomous vehicles making split-second decisions in life-threatening situations and contributing to breakthroughs in fields like drug discovery or climate change modelling to name a few, where speed and accuracy are paramount.

Green AI: Sustainability and ethics

In contrast to Red AI, Green AI is characterised by an approach that yields novel results while taking into account computational cost and encouraging a reduction in resources spent. Where Red AI has resulted in rapidly escalating computational capability and thus carbon costs, Green AI has the opposite effect. It emphasises long-term sustainability, fairness, transparency, and accountability.

Green AI can be a contributor to environmental sustainability in many industries such as agriculture by increasing crop yields while minimising the use for both fertilizer and water; energy by using predictive insights to manage demand and supply of renewable energy; or even in transportation & logistics by reducing traffic congestion, optimising the transport of cargo, and enabling autonomous driving capability.

Proponents of Green AI, often claim that developers of AI systems typically report accuracy or other similar measures but omit any mention of cost, energy efficiency or societal impact. The focus on this single metric ignores the economic, environmental, and social cost of attaining the reported results.  Red AI is not necessarily irresponsible AI. For example, Red AI for health and for climate science is often justified, since its benefits to human and ecosystem life can outweigh the cost of its creation. Conversely, critics of Green AI often argue that an excessive emphasis on ethics and sustainability could hinder progress and limit AI's potential.

AI itself requires large amounts of energy to operate the training and machine learning processes that make the AI models fit for purpose. To be able to be a net positive contributor, AI must be optimised for maximum energy efficiency and trained using clean energy sources which no doubt will require incentive policies from governments. 

Moving toward a greener AI

While Red AI and Green AI represent two contrasting approaches, it is important to recognise that a binary choice between them may not be the best path forward. We need AI systems that are novel, powerful, and efficient while being environmentally friendly, ethical, and fair. This move to a greener AI is being achieved by a combining of a variety of approaches:

• The advent of AI chip development.
  We have seen the development of chips that excel in AI workloads. These AI chips can complete more computations per unit of energy consumed than Computer Processing Units (CPU) or generic Graphical Processing Units (GPU). This demand has led to Google’s development of tensor processing units (TPUs) and pushed the main players in the chip market such as Intel, Qualcomm and NVIDIA towards this new AI trend.
• The replication and sharing of AI research.
  It is not uncommon that AI research is published without code, and on occasions independent researchers cannot reproduce results even with the code. Researchers can also face internal hurdles in making their work open source. Inevitably this leads to duplication of efforts and prevents efficient sharing. This situation is changing slowly with the establishment of online community platforms for data scientists and machine learning practitioners such as Kaggle and Google Colab.
• The Democratisation of AI models.
  The democratisation of AI refers to the process of making AI accessible and usable by a broader range of individuals and organisations, irrespective of their technical expertise or financial resources. It involves simplifying the tools and technologies associated with AI, providing educational resources, and creating platforms that enable people to develop, deploy, and utilise AI systems. ChatGPT is the latest example of democratisation of AI. Companies like OpenAI and Microsoft with Azure OpenAI provide platforms using the latest AI Large Language Models (LLM) that allows companies to fine tune those models to their purposes. Free Open Source LLMs are also available in the form of GPT4ALL or h2oGPT.
• The use of efficient programming languages for AI model implementation.
  Research conducted by Pereira et al[2] reported that the general assumption that energy consumption is related to execution time is not always accurate and therefore not all programming languages are created equal for energy efficiency purposes. Overall, they concluded that the C programming language is likely to be the fastest and most energy efficient. Of course, the developer may not always have a choice on the programming language depending on the hardware platform, software frameworks and libraries required. Nevertheless, selection of programming language should be a criterion for consideration in an energy efficient design.
• The implementation of energy efficient user interfaces.
  Ultimately, AI Algorithms need to interact with humans via a clear and interpretable user interface without ambiguity. Energy consumption of user interfaces is typically overlooked, yet it is an important consideration in AI portable battery-operated devices. Design techniques for energy-efficient graphical user interface exist such as minimising screen brightness while maintaining readability, using dark or neutral colours, reduced animation, and touch gesture optimisation.

Striking a balance between power and efficiency on one hand, and sustainability and ethics on the other, is crucial to harness the full potential of AI. This is just the start of building a future where AI technologies not only drive innovation and progress but also uphold the values and ethics that are essential for a thriving society.

[1] Green AI by Schwartz et al 1907.10597.pdf (arxiv.org)

[2] Energy Efficiency across Programming Languages by Pereira et al https://greenlab.di.uminho.pt/wp-content/uploads/2017/09/paperSLE.pdf