How this site works
This site asks companies to explain their AI in plain words. So here is how this site works, in plain words, including what isn't finished.
Who runs it
Robert Sweetman, on his own. No money, sponsors, or investors are behind it. No ads, no tracking, no analytics.
Open by design. The code is public on GitHub under the MIT license. The writing is free to reuse with credit, under CC BY 4.0.
How it was built
Robert built it with AI help. Claude wrote most of the code with him, and several AI models helped research and draft the guide; the footer on each page names them. A person is still reviewing every page.
Services we use, and what each one sees
- Vercel hosts the site and its firewall. It sees each visit, including IP addresses, in short-term logs.
- Neon is the database, in Frankfurt, Germany. It holds votes, stored with a scrambled browser code, and suggested statements.
- OpenRouter passes each suggested statement to an AI model. We set it to use only services that promise not to keep your text or train on it. We can ask for that promise but can't inspect it. What OpenRouter itself keeps is covered by its own privacy policy.
- Mistral Small, with Google's Gemini as a backup, reads each suggestion on EU servers to flag names and contact details. It only flags; a person decides.
- GitHub stores the public code. Proton handles email.
- Cloudflare runs the site's DNS, the address book that points your browser to this site. It answers that lookup and doesn't see the pages you open.
- The fonts come from Google Fonts. They are copied when the site is built and served from this site, so your browser never contacts Google for them.
- The Internet Archive keeps saved copies of our sources. We ask it to save public pages. It learns nothing about you unless you open one of its links.
What we keep, and for how long
- Votes: until you press “Clear my votes.” The browser code lasts 180 days. We haven't set an automatic deletion date yet.
- Suggestions: rejected ones are deleted. Approved ones stay while they're shown.
- Letters: what you type into a letter stays on your device. It isn't sent to this site or saved.
- Never kept with your vote: your name, email, or IP address.
What isn't finished or perfect
- Human review of the guide is still in progress.
- The help pages and letters haven't been reviewed by a lawyer, legal clinic, or consumer organisation yet. How a review works is set out under “Reviewing a page” on this page. Reviewers are named only with their permission.
- The translations were made with AI help. Native speakers haven't checked all of them yet.
- Suggestions only appear in the language they were written in.
- Results appear only after 20 or more votes per language, so early on there are none.
- The AI check can be wrong. A person makes every decision.
- To keep costs low, the AI checks at most 100 suggestions a day. After that, a person still reads each one, without the AI check. After 200 suggestions in a day, the suggestion box pauses until the next day.
- Voting without accounts can't stop someone from voting more than once, for example with a second browser or a script. A speed limit on votes from one connection slows scripts down, but it can't stop them. So the results are a rough signal, not a count of people.
Reviewing a page
Can you review the help pages or letters? Lawyers, legal clinics, consumer organisations, and native speakers are all welcome. Write to reviewers@aiconsumerrights.org. We name reviewers only with their permission, and a no needs no explanation.
- What we ask. About 15 minutes on one page, for one country, in your own field. Tell us what is wrong, out of date, or misleading.
- What happens next. We fix what you find and list it under Corrections, on this page.
- How you are credited. You choose: by name, as an organisation, anonymously, or not at all. Nothing is published without your yes to the exact words.
- What you can reuse. The guides and letters are CC BY 4.0. You can use them with your own clients, with credit.
- What a review means. It covers only the part the reviewer looked at. It is not approval of the rest of the site.
- We ask each organisation only once. If the answer is no, or there is no answer, we don't write again.
Who has reviewed what
No one has reviewed a page yet. When someone does, it will be listed here.
Corrections
When something on this site is wrong, we fix it and list it here, newest first. Each entry says what was wrong and what changed. We name the person who flagged it only if they ask.
- What was wrong
- The source list said every link had a copy saved at the Internet Archive. 38 of the 99 linked sources had none. It also gave a single check date, 23 September 2026, though sources were added and checked until 29 September. And it labelled the GDPR and the EU AI Act "Link not yet opened" next to "Read by Robert Sweetman".
- What we changed
- 32 more sources now link to a saved copy; 6 pages could not be saved, and their entries show no copy. The text now says each entry links to its archived copy when there is one. The list shows the first and the latest check dates. The two EU laws now say a person opened them, because their website blocks automated readers. A test keeps these statements in step with the data.
- What was wrong
- The EU part said anyone who uses AI to make a deepfake must disclose it. That was broader than the law. The US part said platforms must remove "known copies".
- What we changed
- The EU part now says who the rule covers: companies, public bodies, and people using AI in their work. Purely personal use and police work allowed by law are left out, and art and satire have a lighter rule. The US part now says "known identical copies", as the law does.
- What was wrong
- Two lines said that when a machine alone makes a serious decision, you can ask for a person to review it, with no conditions. The law has exceptions. The India column also described a draft rule as still open for comment, though comments closed in May 2026.
- What we changed
- Both lines now say you can "usually" ask, as the help pages already did. The India column now says comments closed in May 2026 and that no final version was found as of 27 September 2026.
- What was wrong
- This page said voting without accounts can't fully stop someone from voting twice, "for example by using a second browser". That understated it: a script could cast many votes. Separately, the footer and the About page said a person reviews each page "before launch", though the site was already live.
- What we changed
- The page now says a script could vote many times, and that results are a rough signal, not a count of people. The footer and the About page now say the review is still under way.
- What was wrong
- The Brazilian Portuguese guide called Brazil's data protection body by its old name, "Autoridade Nacional de Proteção de Dados".
- What we changed
- It now uses the official name, "Agência Nacional de Proteção de Dados" (still ANPD), as the agency's own website does. This was found while checking the agency's complaint page.
- What was wrong
- The note under the suggestion box stopped naming the AI tools that read each suggestion. We removed the names by mistake, thinking our rule against singling out companies applied.
- What we changed
- The note names them again: Mistral Small, with Google's Gemini as a backup, on EU servers, through OpenRouter. That rule is about what the site says about others. It never hides the tools the site itself uses.
- What was wrong
- Forum results could get stuck on the first answer the server ever gave, so new votes might not show up in the results.
- What we changed
- The site now asks the database for fresh numbers every time. A test checks this, so the problem can't quietly come back.
- What was wrong
- The Italian site name said "nell'IA", which reads as "inside AI". The forum was called "Foro", which in Italian can mean a hole or a law court. Two PAUSA steps had grammar mistakes.
- What we changed
- The site name is now "Diritti dei consumatori di fronte all'IA", the forum is the "Forum Voce e Visione", and the two PAUSA steps were rewritten in proper Italian.
Tell us what's wrong
Criticism is welcome when it comes with a reason. Tell us what's wrong, why it matters, and, if you can, how to fix it.
Email feedback@aiconsumerrights.org, or open an issue on GitHub.