AI Transparency

Last updated: 7 November 2026. Providers change from time to time. We update this page when they do.

1. Which features use AI, and who provides it

When you use one of these tools, the text, files or images you give it are sent to the AI provider listed, so it can produce your result. Where a tool lists more than one provider, it may switch between them to keep working. Our Privacy Policy explains, tool by tool, what data is sent, where it is processed, and how long it is kept.

OpenRouter is a US routing service that passes a request to a model host. On every route that carries your content, we tell OpenRouter to use only hosts that do not store or train on prompts, and to exclude China-based hosts. Some models (such as Z.ai GLM or DeepSeek) are developed by Chinese companies; we use them only on hosts outside China, mainly in the US, protected by the safeguards in section 12 of our Privacy Policy. The same rules apply to AI Labs and Prompt Builder's own-key testing, which use your own OpenRouter account.

We use providers' business (API) services. The terms of Anthropic, OpenAI and Mistral say they do not train their models on the content we send. fal's terms say it does not use what we send to train or develop its own products, except for models it marks "Pending Enterprise Ready"; models fal runs for partners (such as Google and ElevenLabs) follow the partner's terms. Perplexity's terms say it does not train on content sent to its answer engine. Perplexity's Search API and Exa (used for search queries in Search & Cite and Skill Generator research) may keep queries and use them to improve their services. Before a Search & Cite query goes to Perplexity's Search API we remove email addresses, phone numbers, card or ID-style numbers and social media handles. We do not remove names, so please do not put personal details about yourself or others into a claim you check. We no longer use Hugging Face.

2. Chatbots and assistants are AI

Prompt Keeper, Persona Studio characters, Agent Rooms participants and Helper Studio helpers are AI systems. They are not people, and no person reads your messages as you type them. Saved chats can be read by the website owner (Helper Studio) or by us when needed to fix problems, as our Privacy Policy explains. Helper Studio helpers say "AI assistant operated by" followed by the website owner's name, and website owners cannot turn this off. Prompt Keeper tells you it is "an AI assistant, not a person" when you open it; Persona Studio and Agent Rooms show a note next to the message box saying that you are chatting or working with AI, not a person. Persona Studio characters are fictional AI personalities, even when they are based on a type of person or role.

3. How we label AI-generated content

• Images: images you generate or edit with Image Tools while signed in, and PromptBook and Slide Studio illustrations and covers, are saved with IPTC metadata saying they are AI-generated ("Digital Source Type: trained algorithmic media", or "composite with trained algorithmic media" when you edit your own photo), plus a credit line naming our service and, where known, the model used.
• Limits of image labels: only PNG and JPEG files are marked. Not marked: WebP results, previews and thumbnails, and images made before the "Last updated" date of this page. Exports that re-create the image (PromptBook PDF, EPUB and Word files, Cover Studio text overlays, Slide Studio PowerPoint files) lose the metadata; the original image and direct downloads keep it. We do not add C2PA "content credentials", and where our tools convert an image (PromptBook, Slide Studio), credentials added by the model provider are lost. Metadata is not proof of origin: a screenshot or a re-save removes it.
• Audio: SoundLabs audio is not marked yet. Treat it as AI-generated and label it yourself when you publish it.
• Text: there is no reliable way to mark plain text invisibly, and our text exports and share pages do not yet add an AI statement, except Sheet Scorer exports (CSV, Excel and copied tables), which carry an "AI-generated" column and, in Excel files, an AI-generated note in the file properties. Label AI-assisted text yourself where readers expect to be told.
• Metadata can be removed by other software, and not every platform shows it. If you publish AI-generated content, especially realistic images or audio of real people or events, label it yourself where the law or platform rules require it. Our Acceptable Use Policy asks you not to remove our labels.
• Some of our own blog articles are drafted with AI assistance.

4. What AI can and cannot do

• AI can be confidently wrong. It can invent facts, quotes, sources, prices or links, misread documents, and reflect biases in its training data. Our tools include checks to reduce this, such as source checking in Search & Cite, but none can remove the risk entirely.
• It is not professional advice. Don't rely on AI output for legal, medical, financial, careers or safety decisions without checking with a qualified person.
• Results vary. The same request can give different answers, and output can resemble content other people receive.
• AI detection is uncertain. Our AI Detector, like every AI detector, makes mistakes in both directions. Its result is an estimate, not proof, and must never be the only basis for accusing someone of using AI.
• About the AI Humanizer: it rewrites text so that it reads more naturally. We do not claim that it makes text "undetectable", and we have reworded older pages on our site that suggested it did. You must not use it to hide AI use from someone entitled to know, such as a school, university, employer or publisher with AI rules.

5. Automated decisions

• No solely automated decisions with legal or similarly significant effects. We do not use AI on its own to make decisions about you that have legal effects or affect you in a similarly significant way.
• Checking public submissions: text submitted to our public libraries may first be checked by an AI model (from OpenAI), which looks for spam, illegal content and content that does not fit the library. It can only approve a text submission. Anything it does not approve is reviewed by a person, and every rejection is decided by a person. Image submissions are always reviewed by a person before they are published.
• Safety filters: our image, audio and some text tools automatically block requests that look like prohibited content (see the Acceptable Use Policy). If you think a request was blocked by mistake, tell us.
• Introductory offer checks: an automatic check stops the FREE first month of Pro being claimed more than once by the same person. If you think it got it wrong, contact us and a person will review it.
• Moderation decisions about your content or account are explained to you, including whether automated tools were involved, and you can appeal to a person.

6. The rules we follow

We design our AI features with the transparency requirements of the EU AI Act (Article 50), UK data protection law, UK consumer law and our AI providers' usage policies in mind. The EU AI Act's rule that people must be told when they are interacting with an AI system has applied since 2 August 2026; its rules on machine-readable marking of AI-generated content (Article 50(2)) apply to systems placed on the market before 2 August 2026 from 2 December 2026 (Article 111(4) of the AI Act, added by Regulation (EU) 2026/1744, the Digital Omnibus on AI). If you have questions about our use of AI, or want a person to review something an AI tool did, email steve@thepromptindex.com.

Last updated 7 November 2026