Programming Languages for Artificial Intelligence and Machine Learning: An Updated Analysis with Original Benchmarks on Emerging Contenders, Systems Performance Evaluation, AI Tooling Sociology Perspective, and Quantified LLM Bias

This paper reframes the study of programming languages for Artificial Intelligence (AI) and Machine Learning (ML) as both a systems problem and a socio-technical phenomenon shaped by tools, developer incentives, and Large Language Models (LLMs). This comprehensive analysis synthesizes insights from a broad range of historical and contemporary studies on programming languages for AI and ML. We trace the evolution of AI languages, evaluate key general-purpose, domain-specific, and emerging languages across four critical metrics (execution speed, memory efficiency, ease of use, and community/library support), and provide detailed methodology for scoring. To add novelty, we include original small-scale benchmarks on emerging languages (Mojo, Julia, Rust, Zig) compared to Python and C++ baselines for standard ML tasks like matrix multiplication (1000x1000 matrices) and ResNet-50 inference, using tools such as PyTorch for Python, Flux for Julia, Burn for Rust, Modular for Mojo, and standard library approximations for Zig. These are combined with aggregated benchmarks from prior sources. Additionally, we integrate systems benchmarking with an AI tooling sociology lens, treating language adoption as an emergent property of technical capability, tooling accessibility, and machine-generated bias. We introduce a quantified empirical study measuring language bias in LLM-generated code. Using a controlled prompt corpus across multiple LLMs, we demonstrate a statistically significant preference for Python in language-agnostic AI/ML tasks. We argue that language choice in AI/ML is no longer driven solely by performance or expressiveness, but by feedback loops between tooling, education, community scale, LLM training data, and biases that create self-reinforcing cycles influencing language adoption independently of technical merit. Our findings position Python as a socio-technical default, C++ as a performance substrate, and emerging languages such as Mojo, Julia, Rust, and Zig as systems-level challengers whose adoption is constrained as much by tooling bias as by technical maturity. This work introduces novel contributions through original small-scale benchmarks on emerging languages and a quantified empirical study of LLM biases in code generation, offering fresh insights into the socio-technical forces shaping AI/ML language adoption and carrying significant implications for the future evolution of AI systems and programming languages.