Where the numbers come from
Most public figures on AI's footprint are re-cited until nobody remembers what they measured. My research traces them back and puts them on equal terms. None of these papers is peer-reviewed yet.
The Hidden Ledger
An accounting framework for the manufacturing carbon of AI chips. Carbon per job swings 2x across the 3- to 6-year depreciation lives companies disclose, a choice no reporting rule sets; the paper proposes Depreciation-Aligned Attribution.
Boundary Drift
How AI training-emission figures lose their system boundaries as policy reports re-cite them. A pilot study recomputing the Stanford AI Index figures on one standard shows the effect: on equal terms, the widely cited 20x carbon gap between BLOOM and GPT-3 falls to 13.5x, and to 2.5x on energy. More than 15 professors have given feedback.
“Three water numbers, none of them comparable”
A 1,200-word industry column tracing three conflicting AI water-use figures to their original sources, with 21 citations. Accepted; scheduled to run Nov 23, 2026. The link goes here once it runs.
Flowlink Technical White Paper
A 30-page methods paper mapping per-job energy, water and embodied-carbon math to 10 climate-reporting frameworks, including the EU's CSRD and California's SB 253. An earlier draft went to 91 faculty for comment in September 2026, and Brian Prest, an economist at Resources for the Future, reviewed the full paper with written comments in October 2026. Prof. Angel Hsu's Data-Driven EnviroLab at UNC-Chapel Hill is now reviewing its methods. We are also pursuing a GPU validation pilot on Vanderbilt University's ACCRE computing cluster with Vanderbilt Law researchers.
Execution and Resource Accounting for Evidence-Gated Business Reply Workflows
A 7-stage AI reply workflow tested against single-pass replies on 20 fictional cases (120 replies). Submitted to an IEEE BigData 2026 workshop (SE4AgenticAI).