Externality Accounting: Carbon, Water, and Social

This article teaches the three externality categories — carbon, water, and social — and the measurement methods that produce defensible estimates. The estimates enter the Sustainability-Adjusted Value (SAV) framework from Article 34 and the board-grade reporting of Article 35. Externality accounting is technically demanding; the article is honest about where the methods are mature and where they remain contested.

Carbon accounting for AI

AI workloads consume electricity, which in most grids produces greenhouse gas emissions. Two primary components require accounting: training and inference.

Training carbon

Training a large model consumes significant compute over weeks or months. The carbon footprint depends on the compute hours, the hardware’s energy efficiency, the data center’s Power Usage Effectiveness (PUE), and the grid’s carbon intensity at the training location.

The canonical peer-reviewed reference is Strubell, Ganesh, and McCallum’s 2019 ACL paper “Energy and Policy Considerations for Deep Learning in NLP,” which estimated carbon emissions for several NLP training runs and demonstrated that neural architecture search in particular could emit several hundred tonnes of CO2.¹ Patterson et al.’s 2022 IEEE Computer paper provided a more balanced follow-on, arguing that training emissions will plateau and then decline as efficiency gains outpace scale growth.²

For an AI value practitioner, training-carbon accounting typically uses published estimates rather than first-principles computation:

Training CO2e ≈ GPU-hours × average GPU power draw × PUE × grid carbon intensity

Mature cloud providers now report region-specific carbon intensity for their data centers; open-source libraries like CodeCarbon, MLCO2 Impact Calculator, and the Green Software Foundation’s measurement tools wrap the computation. Stanford HAI’s AI Index Report includes training-carbon estimates for frontier models in its annual compute-cost trendline.³

Inference carbon

For most enterprise AI features, inference carbon exceeds training carbon over the feature’s lifetime — training happens once, inference happens millions of times. Inference carbon per request depends on the model size, the hardware tier, the context window, and the grid carbon intensity at the inference location.

For managed provider inference (OpenAI, Anthropic, Google, etc.), per-request emissions must typically be estimated from public figures — providers publish partial data. For self-hosted inference, the computation is direct: measured GPU-hours per request × power × PUE × grid intensity.

An accounting discipline: report inference carbon at the workload level (total CO2e per month for the feature) and at the unit level (grams CO2e per request). Both figures matter — workload figures drive organizational totals; unit figures drive comparisons across architectural choices.

Water accounting

Data centers use water for cooling; AI workloads intensify water use. The OECD AI Policy Observatory publishes measurement guidance for AI water consumption.⁴ The accounting is typically per data center and scaled to the workload’s share of data-center compute.

Water consumption is measured two ways. Water Usage Effectiveness (WUE) measures water used per kWh of IT energy. Water Consumption Effectiveness (WCE) measures total water consumed (withdrawn plus evaporated) per kWh. Data-center operators publish WUE; WCE is less commonly reported but more policy-relevant.

For AI value reporting, water accounting typically uses cloud-provider disclosures where available (Google publishes per-region water figures; AWS and Azure provide partial disclosure) and defaults to industry averages where specific figures are unavailable. Water accounting is less mature than carbon accounting; uncertainty bands are wider.

Social externalities

Social externalities are the human impacts of AI deployment that fall outside the direct buyer-seller relationship. Three categories most commonly require accounting.

Category 1 — Employment displacement

An AI feature that reduces headcount creates employment effects. Net displacement is often less than gross (workers redeployed rather than laid off), but the net figure is not always positive either. Honest accounting tracks both gross and net, along with the quality of redeployment (comparable role? reduced compensation? involuntary separation?).

Stanford HAI and MIT Sloan research have documented patterns of AI-driven workforce change over the past three years; the patterns vary substantially by industry and role.⁵ Accounting for displacement in a single organization requires distinguishing AI-driven changes from changes driven by other factors (market conditions, reorganization, automation of non-AI processes).

Category 2 — Equity and fairness impact

When AI affects individuals differently across protected classes, equity externalities accrue. A credit-scoring system that approves at different rates across demographic groups imposes a distributional cost on the disadvantaged group; a hiring system that screens out certain resume patterns imposes a cost on candidates whose resumes fit those patterns.

The Dutch Toeslagenaffaire remains the most-cited case of systemic equity failure from algorithmic decision-making. Parliamentary inquiry documentation quantified the distributional impact — disproportionate wrongful benefit claims against specific ethnic groups — and the corrective actions required.⁶ For AI value practitioners, the case illustrates that equity externalities are not abstract: they reach monetary quantification through regulatory penalty, compensation liabilities, and reputational damage.

Category 3 — Privacy and surveillance externalities

Data-intensive AI creates privacy externalities. Even when individual consent is secured, the aggregate effect of collecting and processing data at scale changes the privacy landscape in ways individuals cannot negotiate individually. Organizations under GDPR, CCPA, and analogous regimes now face explicit cost pathways (fines, remediation) for privacy externalities; broader ethical accounting assigns costs even where regulatory exposure is absent.