AI Policy & Scope
Last updated: 15 May 2025 · Transparency about how we use AI in pdfonweb
In plain English: pdfonweb uses AI (Anthropic's Claude) for one specific purpose — visually redesigning PDF documents when you ask it to. The AI reads your PDF's text, applies a design template you choose, and outputs a new, beautifully formatted PDF. That's the full scope. The AI makes no decisions about your account, does not moderate your content, and is never used to analyse your behaviour.
1. What AI Does in pdfonweb
Artificial intelligence in pdfonweb is used exclusively for the AI Redesign feature. This feature:
- Extracts the text content of a PDF you choose (up to 15,000 characters)
- Sends that text, along with a design template instruction, to Anthropic's Claude API
- Receives back a complete HTML document with your content laid out in the chosen template's visual style
- Renders that HTML into a new PDF using headless Chromium (Puppeteer) on our own server
- Saves the new PDF to your dashboard as a separate document
Nothing else in pdfonweb uses AI. Your dashboard, analytics, viewer, upload processing, and account management are all rule-based software with no AI or ML component.
2. The AI Pipeline Step by Step
1
You choose a PDF and a template
From your dashboard, click Redesign on any ready PDF. A template picker shows 8 design options across Regular and Premium tiers. You select one and confirm.
2
Text extraction (on our server)
We use pdf-parse, an open-source Node.js library, to extract readable text from your PDF. This runs entirely on our server — nothing is sent externally yet. We cap extraction at 15,000 characters to stay within model context limits and reduce cost. Images, fonts, and vector graphics in your PDF are not extracted and are not sent anywhere.
3
Prompt construction
We combine the extracted text with the template's design instructions (colour palette, typography rules, layout directives) into a prompt. The prompt instructs Claude to produce a complete, self-contained HTML page that faithfully represents your content in the chosen visual style.
4
Claude API call (external)
The prompt is sent to Anthropic's Claude API over HTTPS. This is the only external AI call. The response is a block of HTML. No other data about you, your account, or your viewers is included in this call.
5
HTML → PDF rendering (on our server)
The HTML from Claude is rendered into a PDF by Puppeteer (headless Chromium) running on our server. This stays entirely within our infrastructure. The rendering engine has no internet access during this step — it cannot make outbound requests to external URLs in the HTML.
6
Standard processing pipeline
The rendered PDF goes through the same compression and page-image rendering as any manually uploaded PDF. It appears in your dashboard as a new document titled "Original Title (Redesigned)".
3. Which AI Models We Use
Both models are operated by Anthropic, PBC. We do not fine-tune or modify these models. We do not use any open-source, self-hosted, or on-device AI models in the redesign pipeline.
4. What Data Is Sent to AI
IS sent to Anthropic
- Extracted plain text from your PDF (max 15,000 chars)
- Your PDF's title (as document context)
- The template's design instructions (our own prompt text)
✗ Is NOT sent to Anthropic
- Your email address or username
- Your account ID or BU coin balance
- PDF page images or embedded graphics
- Viewer analytics or session data
- Other users' data or documents
- Payment or billing information
- Your IP address or device information
5. What AI Does NOT Do
- Does not moderate content — we do not use AI to scan, classify, or flag your uploaded PDFs. Content moderation is handled by our human review process on receipt of abuse reports.
- Does not make account decisions — AI has no role in account creation, suspension, or termination. These are rule-based and human-reviewed.
- Does not personalise your experience — there is no recommendation engine, no behavioural profiling, and no ML-based ranking of your documents.
- Does not analyse viewer behaviour — view analytics are simple aggregate counts (sessions, duration, page count). No ML is applied to viewer data.
- Does not learn from your data — we do not train or fine-tune any model on your content. Anthropic's API policy (as of May 2025) states that API-submitted content is not used to train their models by default.
- Does not generate images — all visual output from AI redesign is CSS/HTML-based. No AI image generation (DALL-E, Midjourney, Stable Diffusion, etc.) is used.
- Does not access the internet during rendering — Puppeteer runs in a sandboxed environment with no outbound network access during HTML-to-PDF rendering.
6. AI Output Quality & Limitations
AI-generated redesigns are produced by a language model that interprets your content and applies layout instructions. As with all AI systems, the output is probabilistic — not every redesign will be perfect. Known limitations:
- Text-only reconstruction: The AI works from extracted text only. Complex tables, charts, infographics, or image-heavy pages may be simplified or restructured in the output.
- Layout interpretation: The AI infers document structure (headings, body text, captions) from context. Unusual formatting may be misinterpreted.
- Language handling: The models perform well in English and most major languages. Highly specialised technical or domain-specific content may produce lower-quality output.
- Page count changes: The redesigned PDF may have a different page count than the original due to reflowing content into the new layout.
- No guarantee of accuracy: The AI may occasionally paraphrase, shorten, or reorder content. Always review the redesigned PDF before distributing it.
Important: You are responsible for reviewing AI-redesigned PDFs before sharing them. pdfonweb provides the redesign as a creative tool, not as a replacement for professional document review. We do not guarantee that AI output is free from errors, omissions, or copyright concerns.
7. AI & Copyright
The AI redesign feature transforms the visual layout and presentation of your document's text. It does not reproduce copyrighted images or fonts from your original PDF. The structural text of your document, which you own or have rights to, is preserved in the output.
You must only use the AI redesign feature on documents whose text content you have the right to process through a third-party AI API. By triggering a redesign, you confirm this.
Questions about the copyright status of AI-generated output are evolving legally. We recommend consulting legal counsel if you intend to assert copyright over AI-redesigned documents in a commercial context.
8. Anthropic's AI Safety Policies
Our use of Claude is governed by Anthropic's Usage Policy. Anthropic's models include built-in safety training that refuses requests to generate harmful, illegal, or deceptive content. If a redesign job is refused by the model due to content safety reasons, the job will fail and your BU coins will be refunded.
We do not attempt to circumvent Anthropic's safety systems. Our prompts are designed to produce visual document layouts — not to extract, reproduce, or amplify any harmful content that may exist in your PDF.
9. Human Oversight
All AI outputs are delivered directly to the user who requested them — no human at pdfonweb reviews AI-generated PDFs before they appear in your dashboard. You are the first human reviewer of every AI redesign.
We do periodically review system-level AI usage logs (token counts, error rates, template usage statistics) for operational and cost management purposes. These logs contain no user-identifiable information beyond job IDs.
10. Future AI Features
We may expand our use of AI in the future. Planned areas under consideration include:
- Brand kit injection — automatically placing your uploaded logo on redesigned pages
- Content summarisation — generating a short abstract for your flipbook's preview card
- Accessibility improvements — AI-generated alt text for PDF images
Any new AI features will be documented here before launch. We will update this policy and notify registered users when a materially new AI use is introduced.
11. Contact & Questions
For questions about our AI systems or to report concerns about AI output:
