How Citizen Analyst Uses AI
Citizen Analyst Documentation

How Citizen Analyst Uses AI

Skip to main content
Table of Contents
< All Topics
Print

How Citizen Analyst Uses AI

Last updated: November 14, 2026

Citizen Analyst uses AI to help find records, draft and edit text, make illustrations, search our own articles and answer readers’ questions in a chat window. A human editor decides what is included, how claims are worded, and whether a correction is made. AI output is never our evidence: every factual claim we publish must trace to a court record, an official document or credible reporting.

This page says where AI is used, what people always decide, how you can tell when something was made by AI, what we never do, and what information leaves our servers when you use our chat tools.


Where We Use AI

Research assistance. AI tools help us locate, organize and cross-reference public records, court filings and news reporting. Our editorial tool sends research questions to Claude, the AI model made by Anthropic, and to Tavily, a web-search service. The results are leads for an editor to check, not sources.

Drafting and editing support. Editors can ask Claude to draft a section, suggest revisions or check a document against one of our methods. The editor reads, changes and checks that text before it goes anywhere.

Illustrations. Our editorial tool can generate illustrations with two image models: Flux, from Black Forest Labs, and Gemini, from Google. These appear in articles, infographics, slideshows and newsletter covers. How we label them is described below.

Chat on our site. The AI Rights Advisor answers questions about civil rights, voting and civic action. Its answers are written by Claude and grounded in excerpts from our knowledge base. It gives general information, not legal advice.

Search over our articles. To let the chat find relevant articles, the text of our published articles is converted into a search index by an AI model.

Connections between people and organizations. Some connections shown in our relationship graphs, and the summaries of each cluster of connections, are produced by AI from our published documents. They are labelled “AI-Generated” where they appear.


What a Human Always Decides

AI proposes; an editor decides. Our Editorial Review Process sets out the full division of work. In short:

  • Who is included. Whether a person or organization receives an accountability profile is an editor’s decision. A private citizen must first pass our inclusion gate, which requires documented conduct and not speech alone. See Accountability Profiles — Standards FAQ.
  • How criminal and legal claims are worded. AI may propose language; an editor approves it.
  • Corrections. Our publishing pipeline flags a changed factual claim on a published document automatically, but no correction is recorded on the strength of that flag. A human editor reviews and confirms every correction. See Corrections & Transparency Policy.
  • Publication. Accountability profiles and our editorial-standards pages are reviewed by a human editor before they are published. Some reference material, such as voting-guide updates and investigative-tool documentation, is published by our pipeline and reviewed by an editor afterward. The Editorial Review Process lists which is which.
  • Images of real people. An editor checks every AI illustration before it is placed in a document (see below).

How You Can Tell

AI illustrations carry a visible label. Every illustration our editorial tool generates has the words “AI-generated illustration” in a dark band along its bottom edge. The label is part of the image itself, not hidden data, so it survives a screenshot, a re-post or a resize.

Content Credentials. Since November 14, 2026, illustrations made in our editorial tool on our main editing computer are also signed with a C2PA Content Credential. That is a signed record, attached to the image file, which states that a trained AI model made the image and names the model and our tool. You can inspect it with a Content Credentials verifier. Some limits apply:

  • Illustrations made before November 14, 2026 are not signed, and we have not gone back to sign them. They carry the visible label only. Illustrations made on our second editing computer are not signed yet either.
  • We sign with our own certificate, so a verifier will show the signer as “unrecognized”. The record is still there; the verifier simply does not know our certificate.
  • Any crop, resize or re-save removes the credential. The smaller copies WordPress makes for page layouts, video made from slideshow stills, and re-posts on social media will not carry it. The visible label stays in all of them.

AI-derived connections are labelled. Connections and cluster summaries in our relationship graphs that were produced by AI are marked “AI-Generated”. Treat them as starting points, and follow the citations on each profile for the documented evidence.

AI inference is never presented as evidence. Our Rules & Editorial Standards classify AI inference as an excluded source. When data has been inferred, estimated or generated by AI rather than taken from an authoritative record, it carries a visible indicator.


What We Never Do

  • We never treat AI output as evidence. A claim that only an AI model supports does not appear in our content.
  • We never publish an AI image of a real, identifiable person in a scene that did not happen. An invented scene of a real person is a false claim about them, and a label does not make it true. No software checks for this; an editor does, before an AI image goes into a document.
  • We never cite, publish or keep people-search results. Results from people-search sites and data brokers are leads at most. We never cite them, never publish anything taken only from them, and never store them in our knowledge base. We never run such a lookup on a private citizen before our inclusion gate has passed for that person. A lead is confirmed from a primary record, such as a court filing or a government register, and that record is what we cite.
  • We never get past a login or a bot check. Our data-collection tools stop at a login page, a CAPTCHA or a bot challenge. They do not solve, bypass or disguise themselves to get past one. They follow a site’s robots.txt and terms of service unless an editor records a public-interest exception before the collection runs. They throttle their requests and keep only the fields a story needs.
  • We never ingest police-scanner audio automatically. A person listens. Scanner traffic is a lead, never a record, and we never publish names, plates or addresses heard on air or the location or timing of a live enforcement operation.

The same standards apply to text an AI helped draft as to any other text. Every percentage names its base, every rate gives its denominator and year, and a web page cited for a contested claim is captured in an archive when we cite it.


What Leaves Our Servers

Citizen Analyst requires no account. The site does not ask for your name, email or phone number. The User Manual describes the site’s privacy measures in full. Here is what happens to text when AI is involved.

When you use the AI Rights Advisor. Your messages go from our server to Anthropic, which runs Claude. With them we send our own instructions and short excerpts from our knowledge base, so the answer is grounded in our work. The request is made by our server, not your browser, and our key for Anthropic’s service stays on our server. The advisor’s code does not write your messages to a database, and its request log records only the address requested, the result and the time taken, never what you typed.

When you use the site’s other chat assistant. Some pages carry a chat assistant run by a WordPress plugin. It works the same way: your message goes from our server to the AI model that writes the reply, to a web-search service when it looks something up (limited to a list of vetted news and reference sites), and to our own research service, which finds related entries in our relationship graph.

When we build the search index. The text of our published articles is sent to OpenAI to create the search index, which is stored with Pinecone, a search-database service. Password-protected articles are left out.

When editors work. Drafts, research questions and illustration prompts go to Anthropic, Tavily, Black Forest Labs or Google, as described above. These contain our own working material. They do not contain anything about readers.

We have not described here how those companies store or use what we send them. Their own policies govern that.


How to Report a Problem

If you think something we published is wrong, including anything AI helped produce, tell us. Email factcheck@citizenanalyst.ai with the page, the claim and a source that supports the correction. Your name is optional. Our Corrections & Transparency Policy explains how we review a submission and how a correction is shown, and the Corrections Log lists every substantive correction we have made.


Related Documentation

Was this article helpful?
0 out of 5 stars
5 Stars 0%
4 Stars 0%
3 Stars 0%
2 Stars 0%
1 Stars 0%
5
Please Share Your Feedback
How Can We Improve This Article?