Latenrgy: Model Agnostic Latency and Energy Consumption Prediction for Binary Classifiers

Responsible AI safeguards have measurable computational costs: The paper quantifies how fairness, transparency, and privacy guardrails increase the latency and energy consumption of machine learning models during inference—a previously unquantified impact that organizations need to account for when implementing ethical AI practices.

A universal prediction framework can help optimize AI deployment: The researchers developed model-agnostic equations that can predict how long inference will take and how much energy it will consume across different types of binary classifiers, enabling practitioners to make informed decisions about which models to deploy based on computational constraints.

Efficiency and ethics don't have to be opposing forces: The paper establishes a theoretical foundation showing that computational efficiency and responsible AI principles can be balanced together, rather than treated as competing priorities, which is important for organizations trying to scale AI responsibly.

One-size-fits-all approaches to AI won't work: The framework shows that latency and energy consumption depend on the interplay between classifier type, dataset characteristics, and which specific RAI guardrails are applied—meaning organizations need tailored solutions rather than generic recommendations for sustainable AI systems.

Better prediction tools can guide smarter AI choices: By enabling researchers and practitioners to benchmark and forecast inference performance before deployment, this framework helps organizations select and optimize models that are both ethically sound and computationally sustainable.