I use AI throughout Dream Atlas. It helps research, draft, analyze, implement, document, and review the work.
That disclosure matters. It also leaves out most of the attribution.
“Built with AI” tells you that an AI system materially participated. It does not tell you what the system produced, what I decided, what evidence determined whether the output was acceptable, or who remained responsible when the work became public or irreversible.
Those are different jobs. Compressing them into one phrase makes the process sound simpler than it is — and, depending on which part gets emphasized, either more manual or more autonomous than it was.
What the AI did
AI systems have materially contributed to Dream Atlas. That includes implementation, research, drafting, analysis, documentation, and review where the surviving record supports it.
I am not going to describe that as “a little help with ideas.” That would be false in the flattering direction.
I also cannot produce a clean component-by-component ledger of the opening build. The record supports AI-assisted implementation under my direction, but not a reliable reconstruction of every prompt, proposal, iteration, and local decision. I could make the division of labour sound precise after the fact. It would not become true because the table looked tidy.
So the useful attribution is broad and qualified: AI materially assisted the work. The exact contribution varies by artifact, and some of the early division is no longer recoverable.
That is still real credit.
What I decided
Generating an artifact is not the same as deciding what the product is allowed to become.
Dream Atlas has boundaries that are product decisions rather than implementation details: which projects belong in the public runtime, which identities can speak on a public surface, which routes remain separate, which claims need evidence, and which parts of the fiction are allowed to remain fiction.
An AI system can propose code that crosses one of those boundaries. It can also produce a technically coherent page for the wrong audience, a persuasive paragraph that overstates the evidence, or a fix that solves the local problem by breaking the product around it.
I decide whether the proposal fits the product. I decide what gets kept, rejected, narrowed, replaced, or deferred.
That does not mean I personally authored every implementation detail. It means the product judgment is mine.
The distinction becomes clearest when something fails. The relevant story is rarely just that an AI generated the first attempt and another AI generated the correction. The actual work includes deciding that the first attempt is unacceptable, identifying which boundary it violated, choosing what evidence would count as a correction, and authorizing the replacement.
“Human reviewed” is too weak a description for that. Review can mean glancing at an output and hoping. Judgment needs criteria and consequences.
What counted as true
AI-generated output is not its own source of truth.
A confident explanation of the repository does not outrank the repository. A clean pull request does not prove the deployed route. A successful source change does not prove the provider built the same commit. A plausible historical account does not become evidence because it is well written.
For Dream Atlas, acceptance can depend on repository source, project canonicals, tests, builds, exact-head review, rendered-route checks, provider state, production verification, or direct inspection against an explicit requirement. Different claims have different owners.
This is one of the less glamorous parts of AI-assisted building, and probably the most important. The systems can produce more analysis, implementation, and documentation than one person can inspect. More output does not reduce the need for authority. It makes the authority boundary more necessary.
The question is not only, “Did a human approve this?” It is, “What did the approval rely on?”
If the answer is just the generated output explaining itself, the loop is closed around the wrong thing.
Who remained accountable
AI systems can recommend, generate, inspect, challenge, and warn. They do not become the accountable owner of a public claim, a deployment, a purchase, a credential, or another irreversible action.
That remains with me.
For Dream Atlas, I retain final authority over product identity, source-of-truth decisions, acceptance and rejection, publication, deployment, spending, and correction. If a generated implementation fails, I cannot assign the consequence back to the model. If a public sentence overreaches, the fact that an AI drafted it does not make the sentence less mine once I publish it.
This is not the ceremonial version of “human in the loop,” where a person appears at the end to click a button. The human role is present throughout the process: defining the constraints, deciding what evidence matters, refusing work that crosses a boundary, and owning the result after the tools are gone.
Accountability is the part that cannot be made vague.
The other distortion
There is an opposite failure mode: taking substantial AI assistance and dressing it up as a staff.
Dream Atlas is not an autonomous studio. The named AI collaborators are not employees, legal partners, independent product owners, or borrowed headcount. They can receive bounded credit for supported contributions. They can also have voice and character inside the archive. Neither turns one person’s operating system into a team.
The honest description is already interesting enough: one person building with substantial AI assistance, using explicit boundaries and evidence to decide what survives.
No imaginary org chart required.
A fuller attribution
For this kind of work, I think the disclosure needs four parts:
- Contribution: what the AI materially helped produce, at the level the record supports.
- Judgment: what I defined, selected, rejected, replaced, or kept.
- Authority: which sources and checks determined whether the work was accepted.
- Accountability: who owned publication, deployment, irreversible actions, and correction.
The amount of detail can change with the surface. A project card might need two sentences. A casefile or process log can show the failed approaches, acceptance gates, and evidence. The categories should not disappear just because the copy gets shorter.
I do not want to hide the AI assistance. I also do not want the assistance to swallow every other form of authorship in the project.
“Built with AI” should stay visible. It should not be asked to carry the whole attribution.
// End of transmission. Attribution needs an owner. — ZYANE