AI transparency

auria uses AI to turn a spoken or typed sentence into a suggestion: an appointment, a task, an item on the shopping list. This page explains how that works and is at the same time our transparency statement under Article 50 of the AI Act (EU) 2024/1689.

The core in one sentence: the AI may suggest an assignee, but it never assigns by itself. It executes nothing, and without your confirmation nothing happens.

What the AI does

You speak or type a sentence such as “collect Emma at 3 pm on Thursday, buy nappies, wash the sports kit”. That text goes to the model, and back comes a structured breakdown: an appointment next Thursday at 3 pm, an item on the shopping list, and a task.

That breakdown is a suggestion. It appears as a list you work through. For each entry you have three options:

• Accept: the entry is adopted as it stands
• Adjust: you change the title, date, time, list, or suggested assignee before it is adopted
• Reject: the entry disappears without anything being saved

After adoption, every action can be undone. There is no path by which an AI suggestion reaches your household without your tap.

What the AI does not do

It does not assign a task to a person by itself. If you mention a family member in the Brain Dump, the model can return that person’s short-lived reference as a suggested assignee. You see and can change the suggestion in the app; only your confirmation creates the entry with that assignment.

It has no hands. The model has no access to your database, no tools, and no way to retrieve, store, delete, or send anything. The only route to the AI runs through a single server function that puts text in and gets a structured answer out. That interface can do no more than that.

It does not see your other content. The text you enter, current date, and time zone are transmitted. For family members actually recognised in the text, only first names, short-lived references, and the self marker for the speaker are added. Your existing appointments and tasks, email address, and account identifier are not transmitted.

It makes no decisions about you. There is no automated decision-making in individual cases within the meaning of Article 22 GDPR. There is no profiling, no rating of people, and no analysis of your behaviour.

Which model is used

We use only models from Google’s Gemini family, currently gemini-3.1-flash-lite as the primary model, through the Gemini interface operated by Google Ireland Limited. If a technical attempt fails, a second Gemini model, also operated by Google, may be used as a fallback. No other AI provider is integrated.

The model does not answer in free text but in a fixed schema (structured output). A suggested person can only come back as an allowed short-lived reference and is checked against the household on the server.

Processing may take place in Google data centres outside the European Union. This is safeguarded by standard contractual clauses and Google’s certification under the EU-US Data Privacy Framework. We use the paid tier of the interface for regular user operation. A developer access without active billing may temporarily be used in the strictly internal technical test environment. That test mode is not approved for external testers or production. Under Google’s terms, paid-service data-handling rules apply to access from the EEA, Switzerland, and the United Kingdom even when unpaid quota is used; inputs and outputs are not used to improve general models under those rules. Google may currently log them for up to 55 days for abuse and security monitoring, so we do not promise processing restricted to the request duration alone.

What data the AI sees

The text of your Brain Dump, the current date and time zone, and first names with short-lived references for an assignment suggestion only if recognised in the text. If you dictate a Brain Dump, your device’s speech recognition first turns the recording into text; only that text goes to the model. An audio recording never reaches us or the Gemini API.

The text is exactly what you put into it. If you mention health information, children’s names, or other sensitive matters, that information travels along as part of the text. Such details are not required for household organisation. Please do not enter diagnoses, medication plans, or other highly sensitive details about yourself or third parties in the Brain Dump; use a concise direct form entry where necessary.

Our privacy notice explains how this fits together with the rest of the processing.

Limitations

The AI-assisted Brain Dump is available only in adult login accounts. Account holders must be at least 18; children can be managed as family members without their own login.

Language models get things wrong. auria will misread data, misplace a time, turn one thought into two tasks, or miss a nuance. Ambiguous wording makes it especially likely: “next Friday” can mean two different days.

That is exactly why every suggestion is a question and not a statement. Checking is up to you, and it is deliberately built as a separate step that cannot be skipped. Do not trust the suggestions blindly, particularly not for appointments whose omission hurts.

We check the quality of the suggestions before every change to the system prompt using an automated test suite that also covers edge cases and deliberate attempts to break it. That lowers the error rate; it does not remove it.

Feedback

If you come across a suggestion that is clearly off, or that does something not described here, write to us at hello@uxflows.de. Reports like that are worth more to us than any internal test suite.