Total Cost per Successful Goal (TCSG): A dual-axis reporting framework for the environmental and human cost of AI-assisted work

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Most measures of AI cost count outputs: energy per query, carbon per image, tokens per response. Generative AI work rarely ends in one output. A person asks, rejects, re-prompts and corrects until something is usable. Each rejected attempt draws energy and water and emits carbon, and each costs the person time and attention. This paper proposes Total Cost per Successful Goal (TCSG), a dual-axis reporting framework. Its environmental axis (ECSG) sums energy, carbon and water across every attempt tied to a goal, failures included, and divides by the number of goals independently verified as successful. Its human axis (HCSG) applies the same denominator to a profile of burden measures. The axes are reported side by side and never summed. The paper builds on Panigrahy and Tyagi's Energy per Successful Goal (2026) and does not claim the success-normalised principle as new. In the sources reviewed, none combines failure-inclusive environmental accounting beyond energy with a same-goal human-burden profile. The main contribution is a denominator-governance protocol: the disclosure fields that make an outcome-based figure auditable and hard to game. TCSG is unvalidated. The paper sets out a measurement specification and reports a feasibility pilot of the measurement harness, which produced no verified successful goal and therefore no ECSG value, and which showed in practice why the zero-success and missing-data rules are needed. It then describes a redesigned single-operator instrumentation protocol, catalogues the public datasets and their licensing limits, and states what evidence would be needed before any disclosure standard could use TCSG. Status: preprint, not peer reviewed. No ECSG or HCSG value is reported; worked examples use assumed numbers, and the only empirical material is a feasibility record of the measurement harness.