Beyond carbon: a framework for the full cost of AI conversations — environmental, social, epistemic, and political.
20+
Cost categories across 5 dimensions
100-250 Wh
Energy per long conversation
CC0
Public domain, no restrictions
The problem
Most tools for measuring AI's impact stop at energy and CO2. But the costs that matter most — cognitive deskilling, data pollution, algorithmic monoculture, power concentration — are invisible precisely because no one is tracking them. This project names and organizes those costs so they cannot be ignored.
What makes this different
Existing tools like EcoLogits and CodeCarbon measure environmental metrics well. We don't compete with them — we complement them. This methodology adds the dimensions they don't cover:
Social — annotation labor conditions, cognitive deskilling (CHI 2025), linguistic homogenization
Epistemic — code quality degradation, data pollution (Nature, 2024), research integrity
Political — power concentration, data sovereignty, opaque content filtering
The goal is not zero AI usage but net-positive usage. The framework includes positive impact metrics (reach, counterfactual value, durability) alongside costs.
What's here
A methodology covering 20+ cost categories with estimation methods where possible and honest acknowledgment where not.
A toolkit for Claude Code that automatically tracks token usage, energy, CO2, and cost during sessions.
A related work survey mapping existing tools and research so you can see where this fits.
Help improve the estimates
Many figures have low confidence. If you have data center measurements, inference cost data, or research on the social costs of AI, your corrections are welcome.
Limitations: The quantifiable costs are almost certainly the least important ones. Effects like deskilling, data pollution, and power concentration cannot be reduced to numbers. This is a tool for honest approximation, not precise accounting.
How this was made:
This project was developed by a human
directing Claude (Anthropic's AI assistant)
across multiple conversations. The methodology was applied to itself:
across 3 tracked sessions, the project has consumed
~295 Wh of energy, ~95g of CO2, and ~$98 in
compute. Whether it produces enough value to justify those costs depends
on whether anyone finds it useful. We are
tracking that question.