A comparative study of AI and human programming on environmental sustainability

Area of Science:

• Computer Science
• Environmental Science
• Artificial Intelligence

Background:

• Growing concerns exist regarding the environmental impact of artificial intelligence (AI).
• Previous studies have claimed AI has a lower carbon footprint than human writers.
• These comparisons often neglect the crucial factor of output quality and correctness.

Purpose of the Study:

• To objectively compare the environmental impacts of human and AI programmers generating functionally equivalent code.
• To quantitatively assess AI's environmental impact in code generation while controlling for output correctness.
• To evaluate the trade-offs between AI model size, efficiency, and environmental cost.

Main Methods:

• Utilized the USA Computing Olympiad database to evaluate multiple GPT-based models.
• Developed infrastructure for correctness-controlled assessment of AI code generation.
• Implemented a multi-round correction process to address AI response inaccuracies.
• Calculated AI emissions (usage and embodied) and human emissions (estimated computing power consumption).

Main Results:

• Smaller AI models can achieve environmental parity with human programmers when successful, but frequently fail.
• Widely-used, standard AI models demonstrate significantly higher environmental strain.
• GPT-4 emitted 5 to 19 times more greenhouse gases (CO2eq) than human programmers for equivalent code generation tasks.

Conclusions:

• The environmental cost of AI code generation is highly dependent on model size and performance.
• Larger, more capable AI models like GPT-4 present a substantial environmental trade-off compared to human programmers.
• Further research is needed to optimize AI for environmental sustainability in code generation.