# Plan: Competitive landscape analysis

**Target sub-goals**: 7 (multiply impact through reach), 12 (honest arithmetic)

## Problem

Before sharing the project, we need to know what already exists so we can
position honestly. If a better alternative exists, we should point people
to it rather than duplicating effort.

## Landscape (as of March 2026)

### Tools that measure energy/carbon

| Tool | Scope | Covers social costs? | Per-conversation? |
|------|-------|---------------------|-------------------|
| [CodeCarbon](https://codecarbon.io/) | Training energy/CO2 | No | No |
| [EcoLogits](https://ecologits.ai/) | Inference energy/CO2 via APIs | No | Yes |
| [ML CO2 Impact](https://mlco2.github.io/impact/) | Training carbon estimate | No | No |
| [Green Algorithms](https://www.green-algorithms.org/) | Any compute workload | No | No |
| [HF AI Energy Score](https://huggingface.github.io/AIEnergyScore/) | Model efficiency benchmark | No | No |

### Published research with per-query data

- **Google/Patterson et al. (Aug 2025)**: 0.24 Wh, 0.03g CO2, 0.26 mL
  water per median Gemini text prompt. Most rigorous provider-published
  data. Environmental only.
  ([arXiv:2508.15734](https://arxiv.org/abs/2508.15734))
- **"How Hungry is AI?" (Jegham et al., May 2025)**: Cross-model
  benchmarks for 30 LLMs. o3 and DeepSeek-R1 consume >33 Wh for long
  prompts. Claude 3.7 Sonnet ranked highest eco-efficiency.
  ([arXiv:2505.09598](https://arxiv.org/abs/2505.09598))

### Frameworks that go broader

- **UNICC/Frugal AI Hub (Dec 2025)**: TCO + SDG alignment. Portfolio-level,
  not per-conversation. No specific social cost categories.
- **CHI 2025 deskilling research**: Empirical evidence that AI assistance
  reduces critical thinking. Academic finding, not a measurement tool.
- **Oxford "Hidden Cost of AI" (2025)**: Descriptive survey of social costs.
  Not quantitative or actionable.

### What no one else does

No existing tool or framework combines per-conversation environmental
measurement with social/cognitive/political cost categories. The tools
that measure well (CodeCarbon, EcoLogits) only cover environmental
dimensions. The research that names social costs is descriptive, not
actionable.

## Our positioning

**Honest differentiator**: We are the only framework that enumerates 20+
cost categories — environmental, financial, social, epistemic, political —
at per-conversation granularity.

**Honest weakness**: Our environmental estimates have lower confidence than
Google's or EcoLogits' because we don't have access to infrastructure data.
Our social cost categories are named and described but mostly not
quantified.

**We are not competing with**: CodeCarbon, EcoLogits, or AI Energy Score.
These are measurement tools for specific environmental metrics. We are a
taxonomy and framework. We should reference and link to them, not
position against them.

## Tasks

- [ ] Add a "Related work" section to `impact-methodology.md` citing the
  tools and research above, with honest comparison
- [ ] Calibrate our energy estimates against Google's published data
  and the "How Hungry is AI" benchmarks
- [ ] Link to EcoLogits and CodeCarbon from the toolkit README as
  complementary tools