AI Disclosure

Every book made in this studio is AI-assisted. The text is drafted by a large language model. The illustrations are rendered by a separate image model. The typesetting is performed by deterministic software, not an AI. You review and approve every page before we send the manuscript to press. This page tells you how that works, where the AI ends, and where your responsibility begins.

· I ·The plain statement

Books printed through PagePrinted Studio are generated using artificial-intelligence systems. They are not written by a human ghostwriter. They are not illustrated by a human illustrator. The words and the pictures originate from a model. We disclose this clearly here, on the checkout page, and (where the customer chooses to include it) on the copyright page of the printed book itself.

· II ·Which models, and what they do

• Manuscript text — Anthropic Claude Sonnet (the Claude Sonnet 4.x family, accessed via api.anthropic.com, version-pinned at release time). The model receives your prompts, the synopsis you wrote, and the chapter structure you approved, and returns a draft. It does not have continuous memory between sessions; each chapter is generated in a bounded context. The pinned model version is recorded with the book and shown in your order detail. The published model card lives at anthropic.com/claude.
• Illustrations & theme covers — OpenAI gpt-image-2 (the brand logo on our marketing site was generated with the earlier gpt-image-1; theme covers shown on /discover were generated with gpt-image-2and are then cached as fixed PNGs in our object store — the API we expose at /api/cover/[slug] is a resize-and-format proxy in front of those cached files, not a live model call). For new illustrations inside a book, each prompt is constructed from your scene description, the manuscript context, and the visual style you chose. The model does not see images we have not given it. OpenAI's image-model documentation lives at platform.openai.com/docs/guides/images.
• Safety classifier — Anthropic Claude (same provider, smaller / faster model in the Claude family, called before every text-generation request). See Section VII below.
• Typesetting & binding — performed by deterministic in-house software using LaTeX-derived pipelines. Not AI.

We do not run any model of our own. We do not fine-tune the upstream models. We have no proprietary weights. If Anthropic or OpenAI changes a model, retires a checkpoint, or publishes a new version, that change propagates to PagePrinted at the cadence described in each provider's deprecation policy — we pin a version per release and re-test before promoting.

· III ·The human in the loop

At every step, you are the human in the loop. You write the prompt. You read the generated chapter. You approve it, edit it, or throw it away. You preview the rendered book — every page, in order, at full resolution — before you authorise print. Nothing goes to press without your explicit click. This is not a marketing claim; it's the literal control flow of the studio.

The studio is not a fully-autonomous publishing pipeline; it is a human-directed one with AI doing the heavy creative lifting. That distinction matters legally — under EU AI Act Article 50, your role as the directing human is what places the responsibility for the published content on you rather than on the model provider.

· IV ·How your words are handled

Each theme on PagePrinted carries an explicit declaration of how it handles the text you bring to it. There are two modes:

• Verbatim — themes labelled "Your text is used verbatim" pass your uploaded prose, captions, or recipes through the pipeline unaltered. The AI handles layout, typography, and illustration placement only. It does not paraphrase, summarise, or rewrite a single word. This mode is designed for oral history archives, adoptive-parent memory books, hospice life-review projects, and any other context where the exact wording is irreplaceable. Cookbook Classic and Photo Memories are currently verbatim themes.
• AI-authored — themes labelled "AI writes the text from your prompt" generate the full manuscript from the variables you set (name, age, setting, genre, etc.). No user-supplied prose is expected or stored as content. Little Explorer and Modern Novel are currently AI-authored themes.

The mode for each theme is shown on its theme-detail page before you enter the studio. If a theme's mode is wrong for your use case, do not proceed — choose a different theme or contact us.

· V ·Authorship & copyright

Norwegian and EU copyright law currently grants copyright only to human authors. Output from a generative model on its own is not copyrightable. Your contribution to the book — the synopsis you wrote, the prompts you crafted, the edits you made, the selection and arrangement of chapters, the cover decisions — is human creative input and may attract copyright.

By accepting these terms at checkout, we assign to you any rights in the assembled work that we hold and that are capable of assignment. You are the author of record. You take on the rights and the responsibilities. If a chapter contains a passage that looks like it might be too close to a copyrighted source, the responsibility to review it sits with you; we provide a similarity check at preview time, but it is advisory, not exhaustive.

· VI ·Known limits

AI models are imperfect. They occasionally invent facts ("hallucinate"). They occasionally produce text in awkward style. They occasionally render hands with the wrong number of fingers. We do not promise perfect output; we promise an honest preview, a revision workflow, and a refund policy that recognises when something has gone wrong.

· VII ·The safety classifier

We run an automated classifier in front of model calls to block prompts that violate the Terms of Service — depictions of minors in sexual contexts, instructions for weapons of mass harm, defamatory targeting of identifiable real people. Classifier blocks are reviewable: if you think the classifier got it wrong, click "Ask a human" on the block screen and a real person reads the prompt within 24 hours.

· VIII ·Training data & opt-out

Anthropic and OpenAI both offer zero-retention enterprise endpoints, and we use them. Your prompts and the model's completions are not used to train the underlying models. The models you interact with were trained on data the upstream providers selected and disclosed in their own published material; we are a consumer of those models, not a trainer of them.

Honest unknowns about upstream training data. PagePrinted does not control, audit, or have line-of-sight into the corpora Anthropic and OpenAI used to train Claude and the GPT image models. We can only point you to what each provider has publicly disclosed:

• Anthropic — see Anthropic's usage and privacy policies at anthropic.com/legal/privacy and the Claude model documentation at docs.anthropic.com/en/docs/about-claude/models. Anthropic states that Claude was trained on a mixture of publicly available internet data, third-party licensed data, and data generated by Anthropic's own researchers and users who opted in.
• OpenAI — see OpenAI's training-data policy at openai.com/policies/how-we-use-your-data and the image-model documentation linked in Section II. OpenAI states that its image models were trained on a combination of publicly available, licensed, and human-feedback data.

If you object to AI image models on principle — for example because you believe their training corpora used illustrators' work without consent — please be aware that the theme covers shown on /discover were themselves generated by gpt-image-2, and any illustrations inside an AI-authored book will be generated by the same family of models. We disclose this here rather than burying it, because the persona that cares most about this question is the one most likely to feel deceived if they find out later. If this is a deal-breaker for you, do not place an order; we will not be offended.

· IX ·EU AI Act Article 50

Regulation (EU) 2024/1689 (the EU AI Act) Article 50 requires deployers of generative-AI systems to inform users that they are interacting with AI, and to mark AI-generated content. We comply as follows:

• The studio's editor displays a persistent ribbon: "This chapter is being written by AI — your role is to direct and approve." The ribbon is dismissable but reappears on every new session.
• The checkout page shows the same disclosure above the pay button, with this page linked.
• The printed book may include an optional "AI-assisted" notice on the copyright page. The default is to include it; you can uncheck it before print, but the disclosure to you remains.
• Generated images carry a C2PA-style provenance manifest in the high-resolution download, identifying the model and the time of generation, where the upstream provider supports it.

If you publish or sell books made here, the obligation to disclose AI-assistance to your readers passes through to you under Article 50; we make the disclosure available, you carry it forward.