Artificial intelligence (AI) is no longer a futuristic concept. The technology is being woven into the fabric of our daily lives. From the facial recognition features on our phones to Netflix’s algorithmically tailored show recommendations, AI powers countless conveniences. It’s even revolutionizing fields like healthcare, helping doctors analyze CT and MRI scans more efficiently. 

Tools like ChatGPT can craft emails you’re overthinking or summarize lecture topics just in time for the midterm you’re procrastinating studying for. Yet, behind its seamless functionality lies a cost that most users are unaware of. To truly understand the price of AI, we must first explore how it is developed and operated.

The cost of AI development 

AI refers to computer software designed to mimic human intelligence and learning. It is developed using machine learning, a research branch of computer science that employs data analysis algorithms to simulate how humans learn. Instead of programming every possible input and its corresponding response, machine learning enables software to identify patterns and connections in new information based on the data it was trained on — like images, texts, and videos. Over time, it can adapt and improve its responses by learning from its mistakes, much like how humans learn. 

Generative AI systems like ChatGPT are capable of creating new content, whether it’s text, images, or video. These systems use an even more sophisticated subset of machine learning called deep learning in order to make artificial neural networks, which are AI models inspired by the makeup of the human brain. 

The human brain is made of specialized cells called neurons, which exist as an interconnected network. These neurons communicate with each other through links called synapses, which are weakened or strengthened based on how actively the synaptic connection is used. The brain builds stronger connections between different pieces of information as it becomes more familiar. For example, certain synapses will become stronger when you learn that ‘hola’ means ‘hello’ in Spanish. In contrast, human babies initially have the neuron links to differentiate between the faces of monkeys, but because these synapses are not commonly utilized as children grow up, these links are quickly broken.

Deep learning aims to recreate similar neural networks within AI. However, attempting to reach even a fraction of the brain’s complexity is an expensive task. 

Training these models is resource-intensive. For example, training OpenAI’s GPT-3 reportedly cost an estimated 4.6 million USD in computational resources required to train the model alone. More advanced models, like GPT-4, cost around 63 million USD to train. By 2027, the largest AI training runs are projected to exceed $1 billion in costs. Beyond hardware and computational costs, companies also invest heavily in teams of researchers, engineers, and coders to design and refine these systems.

The environmental costs of AI

AI is also incredibly energy-expensive. Based on its current rapid rate of development, major AI companies like NVIDIA are expected to ship around 1.5 million AI server units annually by 2027. Running at full capacity, these servers would use around 85.4 terawatts of electricity in a year — comparable to the energy used by a country like Sweden, Argentina, or the Netherlands in a year. 

If Google ran a neural network as robust as ChatGPT, its search engine alone would require as much electricity as all of Ireland. If every data center adopted AI at this scale, the global energy consumption of data centers would increase tenfold, which is ridiculously unsustainable. 

Efficiency improvements are one way to mitigate this demand. For example, newer AI models use techniques like knowledge distillation to create smaller, more efficient systems without sacrificing performance. Knowledge distillation transfers what a ‘teacher model’ has learned to a ‘student model,’ saving time and resources. However, as AI becomes more efficient, demand tends to increase, which offsets any lowered energy costs. 

How can the growing energy demands of AI be met with renewable resources? Transparency about energy usage and sustainability practices from AI developers is scarce. As climate instability worsens, addressing the environmental footprint of AI is critical. Do we truly need AI systems of this scale and complexity?

These are just some questions to ponder the next time you use ChatGPT. After all, every convenience has its cost.