This is an English translation provided for convenience. The Romanian version is the legally binding text.
City Twin includes an assistant that answers questions about the city in natural language. This page explains, without jargon, what it does, what it relies on, where it makes mistakes and what it is not allowed to do. We publish this information because a tool used in public decision-making must be open to verification, not simply taken on trust.
1. What it does, in practice
It receives a question put by the user and the city model, a data structure containing the areas, the known problems, the city hall's projects, the public services, the indicators and the relationships between them, together with the source document for each piece of information. From these it generates an answer which may include: the text itself, the explanation of the conclusions, the assumptions made, the sources used, the data limits and gaps identified, and suggestions for next steps.
In parallel, it highlights on the map the entities it refers to, so that the answer can be checked visually, not only in the text.
2. What data it works on and what data it does not
| Data used | Data not used |
|---|---|
| The current city model: official strategic documents (PUG (General Urban Plan), PMUD (Sustainable Urban Mobility Plan), SIDU (Integrated Urban Development Strategy), PAED (Sustainable Energy Action Plan), budgets, council decisions), open data and real-time data (traffic, air quality, weather) | The internet. The assistant is not connected to web search and does not bring in information from outside the model |
| Aggregated urban data: areas, networks, indicators, projects | Personal data. The platform is not intended for it, and the terms of use prohibit entering it |
| Only the city in the uploaded model | Other cities. When asked about a city that is not in the model, the assistant is configured to flag that it does not have the data, not to improvise |
3. The limitations we know of
We list them because they are real, not because they are theoretical:
- The answer is limited by the quality of the data. If a city document is old or incomplete, the answer inherits that limitation. That is why the platform permanently displays the coverage of the model and lists the gaps.
- It may phrase things imprecisely, even when the data is correct. A language model may mislabel a relationship or emphasise one aspect at the expense of another. The sources displayed next to each answer are there precisely to allow verification.
- It does not carry out specialist technical analysis. It does not replace a traffic study, an energy audit, a technical expert report or a legal opinion.
- Estimates remain estimates. Simulations, isochrones and impact assessments are preliminary assessments based on the model's data, marked as such, not validated forecasts.
- Real-time data depends on its providers. An unavailable source means missing information, flagged as such, not filled in from memory.
4. How you check an answer
Every answer is built so that it can be checked, in four steps:
- Read the sources. The answer lists the documents on which it is based.
- Look at the confidence level displayed next to the answer. It is an indicative measure (high, medium, low), not a statistically calibrated probability: a low value is a signal for further checking, not a hidden error.
- Read the “incomplete or missing data” section (“date incomplete sau lipsă” in the interface). It states what the assistant could not support with data.
- Check on the map. The entities referred to are highlighted; you can open the file for each of them.
5. Position under the AI Act
With regard to Regulation (EU) 2024/1689 (AI Act):
- City Twin is not intended for automated decision-making about individuals and does not produce individual legal effects;
- the assistant provides support for substantiating decisions, with human validation and oversight at every step;
- the user is explicitly informed, above the question box, that they are interacting with an AI system, and the results are accompanied by sources and the confidence level, in line with the transparency obligation in Art. 50, applicable from 2 August 2026;
- the analysis sections generated by the assistant are marked as such in exported reports, so that they can be distinguished from text written by a person. Technical marking, through machine-readable metadata, is in progress.
6. Model, provider and location of processing
The reasoning runs at a European AI model provider, in the European Union, with no transfer to a third country. What is sent to it is the question and the relevant information from the city model; no data identifying the user is sent. The categories of providers are described on Providers and data transfers; the named list is made available to client institutions within the contractual framework.
Conversation history. On the public demo, without an account, the conversation thread is kept only in the server's memory, for at most two hours, so that the assistant can understand the next question; it is not written anywhere. If you have an account, your questions and answers are kept in your account, separately for each city, so that you can resume your work the next day: for 90 days, after which they are deleted automatically. Only you can see them. Before saving, personal numeric codes (CNP), email addresses, telephone numbers and IBANs are removed from the questions. We also delete them earlier, on request, if you write to [email protected].
Feedback. If you voluntarily send feedback on an answer (👍 / 👎), we keep the question, the rating and limited technical information needed to improve the service. This happens only when the button is pressed, not automatically, and without user identifiers: the feedback record does not contain the account identifier and does not contain the IP address as an attribute, and we do not use cookies for feedback. We apply filters to reduce the risk of accidentally storing personal data. We use the feedback exclusively to find and fix weak answers. The technical logs of the infrastructure are separate, may process IP addresses for security purposes and are described on the Providers and data transfers page. Details of what is processed and for how long are in the Privacy policy.
7. What we do when it makes a mistake
If you find an incorrect or incomplete answer, or a statement not supported by the data, write to us at [email protected]. We handle such reports as a priority: we check the case, correct the model or the instructions and, for client institutions, communicate what has changed. Reports of this kind are the most useful form of feedback we can receive.