An honest case
Making art in the
age of AI.
AI is not going away. I chose to learn the craft beside it — eyes open about how it was built, and aligned with the company doing the most to build it right. Here is the case, the real proportions, and the missteps I won’t hide.
Why use it at all
An oracle on the desk.
For most of history, a great teacher was a matter of luck and money. You had one if you could afford one, or if you happened to find one. That is changing. A person with a question can now hold a two-minute conversation in place of an afternoon of searching, and come out the other side knowing more.
That is the part the worry tends to bury. Used well, this is the widest door to learning ever built — a patient tutor for anyone with a question and the will to ask it. I am not interested in pretending that good away. I am interested in using it without harm.
What it learned from
The internet, mostly in the open.
No lab publishes its exact recipe, so these are the public reference datasets the field is built on — GPT-3, LLaMA, The Pile. The shape holds across all of them.
Proportions are drawn from public datasets, not any one company’s undisclosed mix, and rounded for reading. Two honest notes: books are high-value data and get weighted up, so a small share carries more influence than its size suggests; and a small percentage does not make piracy acceptable. The point is the shape, not a decimal.
The line that matters: free to read is not the same as free of copyright. The overwhelming majority of training was learning from work any person could already access — ruled transformative, not theft. The narrow wrong was different in kind: pirated books that were never free to take.
The line my tool holds
Not an AI that writes for you.
The grievance writers have is real: machines that replace the writer, or launder other people’s prose into new prose to sell. My tool does neither. It stands on either side of the page and never steps onto it.
Before the page
Build the story
Structure, character, the architecture of the thing — the thinking that happens before word one.
— tool helps here
The page
You write
every word
The prose that gets read, sold, and credited is yours, start to finish.
× tool stays off the page
After the draft
Read it back
Analysis and feedback on the story you finished — what works, what doesn’t, what’s missing.
— tool helps here
It develops the story before you write and reads the draft after you finish. The creative work that carries your name is written by you, and only you. There is no other writer’s prose going in, and none coming out.
The company behind it
The same questions, asked of each company.
Five things I weighed. The green column is where I landed. One row goes to a rival on purpose — a comparison that never loses isn’t worth reading.
| What I asked | Anthropic | OpenAI | Meta | |
|---|---|---|---|---|
| Mission locked into the structure | LeadsPublic benefit corporation since 2021, plus an independent Long-Term Benefit Trust that can seat and remove directors.4 | Became a public benefit corporation in 2025, under its nonprofit Foundation; critics flag weakened profit caps.5 | A division of Alphabet, a standard public company. | A division of Meta, a standard public company. |
| Safety written into the model | LeadsConstitutional AI plus a public Responsible Scaling Policy — principles built into training.6 | Publishes a Preparedness Framework. | Publishes a Frontier Safety Framework. | Publishes a frontier framework; ships open-weight models. |
| Owning the copyright problem | LeadsUsed pirated books, then became the first major lab to pay creators — $1.5B, the largest copyright settlement on record. Denied wrongdoing.3 | Suits consolidated and ongoing; signed forward licensing deals; no past-piracy settlement paid.7 | Relies on a fair-use defense; faces active suits. | Training ruled fair use; piracy claims still live; nothing paid.8 |
| Telling the public its energy and water cost | No per-query figures published. | A rough per-query estimate from its CEO. | LeadsPublished the most detailed production data of any lab.9 | Little per-query data. |
| Founders putting wealth toward the harm | LeadsSeven cofounders pledged 80% of their wealth — more than $21B — toward AI-driven inequality.10 | Nonprofit Foundation holds about 26% equity for mission spending.5 | Corporate giving; no comparable founder pledge. | Corporate giving; no comparable founder pledge. |
Snapshots as of mid-2026, each tied to a source below. The law and these structures keep moving; the links let you check them against the day you read this.
The line they wouldn’t cross
A $200M contract, refused over principle.
In 2025 Anthropic became the first AI approved for the Pentagon’s classified networks, under a $200 million contract — on the condition that Claude not be used for mass surveillance of Americans or for fully autonomous weapons.11 When the Pentagon demanded those limits be dropped, the company refused. In February 2026 the administration blacklisted Anthropic as a “supply chain risk” — a label usually kept for foreign adversaries — and Anthropic sued rather than give the limits up.12
In fairness: some argue working with defense at all sits uneasily with “force for good,” and the government disputes Anthropic’s account. Weigh it for yourself. What is not in dispute is that the company took a real loss to hold a line it had drawn.
Every company that built a large AI model learned from the internet. Most of it — 80 to 90% — is the open web, free for any person to read, and a court has ruled that learning from it is transformative, not theft. You can disagree with the ruling. It is still the law.
A smaller slice crossed a line: pirated books that were never free to take. Every major lab has some version of this in its past. One has paid for it — Anthropic settled with authors for $1.5 billion while the others litigate or license forward.
None of that makes them spotless. It makes them the one, of the options I have, that keeps choosing the more accountable path. AI is not going away. I would rather learn to be an artist beside the company reaching for the right side of history than pretend I can sit the era out.
Not spotless
So the record is straight: the piracy was real — roughly seven million books pulled from pirate libraries, an act a judge found was not protected.3 Anthropic paid, but it denied wrongdoing, and it settled after losing the piracy ruling and facing a trial that could have ended it. The check covers a fraction of the works involved. The benefit trust’s power is untested, and the 80% pledge is a promise, not a completed gift.
I am betting on direction, not on a finished record. Of the companies that could power this work, this is the one trying hardest to set its own house in order — and the tool I built keeps the writer, and the writer’s words, at the center.
Sources
- Meta AI — LLaMA paper: Common Crawl, C4, Wikipedia, and code dominate the training mix; all books combined are ~4.5%.
- EleutherAI — The Pile: composition across 22 sources, including books and web text.
- Norton Rose Fulbright — Bartz v. Anthropic: training ruled fair use; pirated books not protected; $1.5B settlement. · Authors Alliance — the settlement denies wrongdoing.
- Harvard Law School Forum on Corporate Governance — Anthropic’s PBC form and Long-Term Benefit Trust.
- OpenAI — 2025 restructuring into OpenAI Group PBC under the OpenAI Foundation. · The OpenAI Files — criticism of profit-cap and oversight changes.
- Anthropic — Constitutional AI and the Responsible Scaling Policy.
- Copyright Alliance — 2025 review: OpenAI cases ongoing; licensing deals across the field.
- Norton Rose Fulbright — 2026 update: Kadrey v. Meta, training fair use, piracy claims live.
- Google Cloud — published per-prompt energy and water data.
- Transformer News — the cofounders’ January 2026 pledge to give away most of their wealth.
- Mayer Brown — the $200M Pentagon contract and Anthropic’s acceptable-use conditions.
- CNBC — the “supply chain risk” designation and Anthropic’s legal challenge.