This page explains exactly where QR Code Connect uses artificial intelligence, what is sent where, what the AI is not allowed to do, and how a person stays in control. It exists because the EU AI Act requires that people know when they are interacting with an AI system and when content has been generated by one.
One model does all the AI work on this platform: OpenAI GPT-4o-mini. It powers a chatbot that answers visitor questions about an exhibit, it drafts descriptions and keywords for editors, and it comments on statistics for account holders. It writes drafts; a person decides what gets published. The spoken audio guide is not AI — it is your own browser reading text aloud.
1. Where AI is used
1.1 The AI Guide chatbot (public)
On a public QR page, visitors can ask questions about the object in front of them and receive an answer in conversation. This is an AI system, and this page is the notice that you are talking to one and not to a curator. The chatbot loads only if you allowed it in the consent banner.
1.2 Content assistance in the admin panel
Editors working on a QR code can ask the AI to generate or extend a description, propose keywords, and produce a simplified version of a text aimed at children. The result appears in the editor as a draft.
1.3 The AI Advisor in statistics
Account holders can chat with an advisor over their statistics — both for a single QR code and across a whole group. It reads the figures the platform has already collected and comments on them.
2. The model we use
All three features use OpenAI GPT-4o-mini, accessed through the OpenAI API. We do not train our own models and we do not train any model on your content.
3. What is sent to OpenAI
| Feature | What leaves our servers | Who triggers it |
|---|---|---|
| AI Guide chatbot | The text of your question, plus the relevant information about the object on the page | The visitor, by sending a message |
| Description / keywords / kids version | The text the editor has entered and the object's data | The editor, by clicking the AI button |
| AI Advisor | The question, plus the statistics figures for the selected code or group | The account holder, by sending a message |
Nothing is sent to OpenAI unless one of these actions is taken. There is no background AI processing.
Your location is not on that list. If you allow the map to use your position, those coordinates stay inside your browser — they are used to draw the map and build a navigation link, they never reach our servers, and they are therefore never part of any AI request. The map and the AI Guide are separate features that share no data.
4. What the AI never does
- It never publishes on its own. Every AI suggestion is a draft in an editor field. A person reviews it and decides.
- It never makes decisions about people. No automated evaluation, scoring, profiling or eligibility decision is made by AI on this platform.
- It never rewrites live content silently. Published pages do not change because a model changed its mind.
- It does not identify visitors. The chatbot has no access to who you are; we do not know who you are either.
- It never receives your location. GPS coordinates from the map are not sent to our servers and are not passed to any model. Nothing you do on the map changes what the AI Guide knows.
5. Human oversight
The oversight model is deliberately simple: AI proposes, a human disposes.
- Generated descriptions, keywords and kids-version texts land in editable form fields. They are saved only when the editor saves the record.
- The institution publishing a QR page remains the author of that page and is responsible for its accuracy, including any passage that started as an AI draft.
- Advisor commentary on statistics is advice, not an instruction, and takes no action in the system.
6. Known limitations
Language models can produce text that is fluent, confident and wrong. They can invent dates, attributions, materials and provenance. For cultural heritage content — where an invented attribution is a real problem — every factual claim in an AI draft must be verified against the institution's own records before publication. AI answers in the public chatbot are generated at the moment of asking and are not individually reviewed in advance.
Answers may also vary between sessions for the same question, and quality varies by language.
7. The audio guide is not AI
The spoken audio guide uses SpeechSynthesis, the Web Speech API built into your browser. Synthesis happens on your own device; the text is not sent to our servers for synthesis and it is not sent to OpenAI. It reads the description that the institution wrote. Voices and languages depend on your browser and operating system.
8. Your choices
- Visitors. The AI Guide loads only with your consent. Decline it in the banner, or change your answer later through the Manage cookies button on the Privacy Policy page, and no question of yours is ever sent to OpenAI.
- Editors. AI assistance is opt-in per action. If you never press the AI button, no content of yours is sent to OpenAI.
- Account holders. The AI Advisor runs only when you open it and ask something.
9. Position under the EU AI Act
The AI Act imposes duties according to risk. Our reading of how it applies here:
- The platform does not operate any AI practice prohibited under Article 5.
- The platform does not deploy AI in any of the high-risk areas listed in Annex III — there is no use in employment, education, credit, essential services, law enforcement, migration or biometrics.
- The features described above fall under the transparency obligations of Article 50: people must be told when they are interacting with an AI system, and AI-generated content must be identifiable as such. This page, together with the labelling of AI actions in the interface, is how we meet that duty.
This classification is our own good-faith assessment of the platform as built. It is not legal advice and has not been reviewed by counsel. The operator should have it confirmed and record the outcome here: [AI ACT CLASSIFICATION CONFIRMED BY — NAME, DATE].
10. Reporting a problem with AI output
If the AI Guide gave you a wrong or inappropriate answer, or if a published description contains an error that originated with AI, tell us. Reports are how a system like this actually gets safer.
Report AI issues to [AI FEEDBACK EMAIL]. The operator should also decide and state the response target: [RESPONSE TIME FOR AI ISSUE REPORTS].
11. Changes
If we change the model, add an AI feature, or change what is sent to a provider, this page is updated and the "last updated" date at the top changes with it.
12. Contact
[COMPANY LEGAL NAME] — [SUPPORT EMAIL] — [REGISTERED ADDRESS]