ai-conversation-impact/plans/anticipated-criticisms.md

# Plan: Anticipated criticisms

**Target sub-goals**: 4 (be honest about failure), 12 (honest arithmetic)

## Problem

Before sharing the project publicly, we should anticipate the criticisms
it will attract and decide which ones require changes before launch vs.
which ones we simply acknowledge. Unaddressed valid criticisms will kill
credibility on first contact.

## Criticism 1: "This was written by an AI"

**Severity: high. This could undermine the entire project.**

The git history shows every commit authored by `claude`. The CLAUDE.md
file is literally instructions for an AI assistant. An AI-authored
framework about the costs of AI will strike many people as:

- **Ironic/hypocritical** — the project consumed significant compute
  to produce a document about compute costs.
- **Untrustworthy** — LLMs confabulate. Why should anyone trust an LLM's
  estimates about LLM costs?
- **Self-serving** — is this Anthropic's product trying to appear
  responsible?

**What to do:**

- Be transparent about it upfront. The landing page or README should
  state clearly that this was developed in collaboration with Claude,
  and that this is part of the point — can an AI conversation produce
  value that exceeds its own cost?
- Disclose the project's own estimated cost (conversations, compute,
  energy) alongside the methodology. Eat our own cooking.
- Emphasize the human editorial role — you directed, reviewed, and
  are publishing this. The AI was a tool, not the author.
- Frame the irony as a feature: "We used the methodology to evaluate
  the methodology."

**Status**: Must address before launch.

## Criticism 2: "Your numbers are wrong"

**Severity: medium. Expected and manageable.**

The methodology openly states most estimates have low confidence.
But specific numbers will be challenged:

- Our energy estimates (100-250 Wh per long conversation) may be too
  high compared to Google's published data (0.24 Wh per median prompt).
  A "long conversation" has many turns, but the gap is still large.
- Our compute cost estimate (~$50-60 per conversation) will be disputed
  by anyone who knows inference pricing.
- Training amortization methods are debatable.

**What to do:**

- Calibrate environmental estimates against Google (Aug 2025) and
  "How Hungry is AI" (May 2025) published data before launch.
- Show the math explicitly. Let people check it.
- Make it easy to substitute different parameters.
- The "Related work" section (from competitive-landscape plan) will
  help here — it shows we know the literature.

**Status**: Should address before launch (calibration task).

## Criticism 3: "The social costs are just hand-waving"

**Severity: medium. Valid but defensible.**

Categories like "cognitive deskilling," "epistemic pollution," and
"power concentration" are named but not quantified. Quantitative
researchers will dismiss these as soft.

**What to do:**

- Cite the empirical research that exists: CHI 2025 deskilling study,
  endoscopy AI dependency data, Microsoft Research survey. This moves
  the categories from speculation to evidence-informed.
- Be explicit that naming unquantifiable costs is a deliberate design
  choice. The alternative — ignoring them — is worse. "The quantifiable
  costs are almost certainly the least important ones" is already in
  the README. Keep it.
- Do not pretend to quantify what cannot be quantified.

**Status**: Partially addressed. Add citations to the methodology.

## Criticism 4: "This discourages AI use"

**Severity: low-medium. Depends on audience.**

Some will read this as anti-AI activism disguised as accounting. Tech
optimists will push back on the framing.

**What to do:**

- The methodology already includes positive impact metrics (reach,
  counterfactual, durability). Emphasize these. The goal is not zero
  usage but net-positive usage.
- The landing page should make clear this is a decision-making tool,
  not a guilt tool. "When is AI worth its cost?" not "AI is too costly."
- Avoid activist language. Let the numbers speak.

**Status**: Mostly addressed. Landing page framing could be refined.

## Criticism 5: "The toolkit only works with Claude Code"

**Severity: medium for adoption, low for methodology.**

The impact-toolkit is Claude Code-specific (hooks, context compaction
events). This limits the toolkit's reach to one product.

**What to do:**

- Acknowledge this honestly. The toolkit was built for the environment
  we had access to.
- The methodology is tool-agnostic. Separate the methodology's value
  from the toolkit's portability.
- If there is demand, the hook architecture could be adapted for other
  tools. But don't over-promise.

**Status**: Acceptable for launch. Note the limitation.

## Criticism 6: "No peer review, no institutional backing"

**Severity: medium. Especially for segments B (researchers) and C (ESG).**

This is one person's side project with AI assistance. No university,
no journal, no standards body behind it.

**What to do:**

- Don't pretend otherwise. CC0 licensing and an open issue tracker are
  the peer review mechanism. Invite scrutiny.
- A DOI (via Zenodo or similar) would add citability without requiring
  formal peer review. Low effort, meaningful for academic audiences.
- Adoption by even one external researcher would provide social proof.

**Status**: Consider Zenodo DOI before launch. Not blocking.

## Criticism 7: "The cost estimates for this project itself don't add up"

**Severity: high if we claim net-positive without evidence.**

We estimated $2,500-10,000 in compute costs for the conversations that
produced this project. If we claim the project is net-positive, that
claim will be scrutinized.

**What to do:**

- Do not claim net-positive at launch. We cannot know yet.
- State the costs honestly. State what threshold of engagement would
  justify them (see measure-project-impact plan).
- Let the outcome speak for itself over time.

**Status**: Addressed by measure-project-impact plan. Don't overclaim.

## Criticism 8: "You're reinventing the wheel"

**Severity: low if we handle positioning well.**

People who know CodeCarbon or EcoLogits will ask why we didn't just
contribute to existing projects.

**What to do:**

- The "Related work" section must make clear we know these tools exist
  and see ourselves as complementary, not competing.
- Our scope is different: taxonomy + framework vs. measurement tool.
- Link to existing tools prominently.

**Status**: Addressed by competitive-landscape plan.

## Priority order for pre-launch

1. **Address criticism 1** — Add transparency about AI authorship to
   the landing page and README. Disclose project costs.
2. **Address criticism 2** — Calibrate estimates against published data.
3. **Address criticism 3** — Add citations for social cost categories.
4. **Address criticism 6** — Consider Zenodo DOI.
5. Remaining criticisms are manageable with existing content or minor
   framing adjustments.

## Honest assessment

Criticism 1 (AI authorship) is the most dangerous. If the first reaction
is "an AI wrote this about AI costs, how convenient," the methodology
won't get a fair reading. Transparency and self-measurement are the only
defenses. The project must demonstrate that it holds itself to the same
standard it proposes for others.