Rethinking AI Innovation Through Computational Accountability: From Accuracy-Centered Evaluation to Cost-Aware Intelligence
Main article
Abstract
The prevailing evaluation culture in artificial-intelligence (AI) research treats predictive accuracy as the dominant, often sole, indicator of progress. This accuracy-centered view has produced remarkable methodological advances, yet it is increasingly at odds with the operational realities of contemporary AI systems, whose computational demands have risen by several orders of magnitude over the past decade. This paper argues that the next phase of AI innovation must be organised around the concept of computational accountability: a systematic, reproducible, and hardware-independent accounting of the resources an AI model requires to deliver its predictions. We decompose computational accountability into three mutually reinforcing pillars—algorithmic accountability captured by floating-point-operation (FLOP) counts, numerical-precision accountability captured by bit-operation (BOP) counts, and hardware-execution accountability captured by energy and carbon measurements—and argue that each pillar, while useful in isolation, becomes meaningful only when reported jointly. Drawing on a structured review of seventy prior studies spanning Green AI, quantization, hardware-aware design, and sustainable machine learning, the paper develops a conceptual framework that positions computational accountability as a methodological discipline rather than a tool choice, and proposes a unified accountability ledger that records per-run, per-model, and per-precision workload indicators. An illustrative analysis across eight representative architectures and three precision regimes shows that accuracy-normalised cost indicators reorder model rankings relative to raw accuracy, frequently by more than one quartile, and that BOP-based analyses reveal quantization benefits that FLOP-only analyses systematically under-count. The paper concludes with concrete recommendations for researchers, reviewers, editors, and funding agencies, and sketches a policy interface through which computational accountability can be integrated into publication norms, procurement decisions, and sustainability audits. The goal is not to diminish the role of accuracy in AI evaluation but to situate it inside a richer, cost-aware narrative that treats computational demand as a scientific variable rather than a hidden externality.
