AI Use

Living document. Updated whenever usage expands, not retroactively reconstructed.

---

Position

Most objections to AI are objections to how capitalism intends to use it rather than to the tool itself. In the world that I hope to create, technology will be used to support everyone. People are not born to labor, and when we have tools that can replace or reduce labor, that should be a win for everyone. My hope is that one day, tools like AI will be used to improve the quality of life for everyone throughout the world, and not just cause environmental devastation so that a few people can hoard even more wealth.

Some people won’t want to support the work because of that. I understand, and I’d rather be clear about it now. I’m sure my position on this will evolve over time as I do more research and understand the topic more deeply. See also 2026-07-31.

---

Current uses

Area	What AI does	What it does not do
Vault management	Structure, frontmatter, templates, Bases, scripts, tooling	—
This website	Writes and maintains it — the generator config, the build pipeline, the hosting, the checks that have to pass before a page goes live	Decide what publishes
Research assistance	Finding sources, extracting metadata, summarizing what a source argues	Decide what’s true, or what I believe
Restructuring	Reorganizing, tightening, and formatting content I already wrote	Introduce claims I didn’t make
Dialogue	Thinking out loud with a model to work an idea out	Supply the idea and have me ratify it
Review	Finding contradictions, gaps, weak citations, unstated assumptions. Fixing the ones I tell it to	Decide which of them matter, or merge its own work

The site gets its own row because “tooling” hides it. Every line of the thing you are reading this on was written by a model. The layout, the deploy, the scripts that check for leaks. What it publishes is a list I keep by hand.

Boundaries

1. Manuscript prose is mine. AI does not write narrative prose. It may restructure or tighten prose I have already written, when asked. Enforced in CLAUDE.md → Rules for Claude #1.
2. No unsourced claims. AI may not introduce a factual claim without a source note. Enforced by CLAUDE.md #2 and vault_check.py.
3. AI never decides what publishes. publish: true is set by me only. Enforced by CLAUDE.md #5.
4. Dialogue is captured, not paraphrased. Conversations that produced ideas are saved in Dialogues/ with the model recorded, and cited like any other source.

How provenance is tracked

Two mechanisms, both queryable:

• ai-assisted: true and ai-role: on any note where AI did more than mechanical formatting. Roles: research, drafting, restructuring, summarizing, polish.
• Dialogue notes in Dialogues/ record model and link produced to whatever came out of them.

Models used

Model	Used for	Since
Claude Opus 5 (via Claude Code)	Vault setup, tooling, research assistance	2026-07-31

Known tensions

• Training data provenance. The models used here were trained on work whose authors did not consent. I don’t have a resolution for that. I recognize that using the tool anyway is a choice with a cost, and naming it is the minimum.
• Research assistance is not neutral. A model surfacing sources shapes which sources I see. But this usage is a starting point in a longer research process. I will not rely on AI to tell me the truth about any position I take.
• Dialogue can produce agreement rather than thought. Talking an idea through with something inclined to be agreeable risks feeling like rigor without being it.

Changelog

Date	Change
2026-07-31	Document created. Initial uses: vault management, research, restructuring, dialogue, review
2026-08-01	Added the website. It was already true on the day this document was written and went unsaid, which is the failure mode a living document is supposed to prevent
2026-08-01	Audited against the vault. The review row said AI does not resolve what it finds, and by then it had spent a day fixing what it found — under instruction, and merged by me, which is where the boundary actually sits