How Ayva works and where it gets things wrong

1. What Ayva is and isn’t

Ayva is artificial intelligence, not a person. Answers, the voice, suggestions and the messages it sends on its own are generated by language models: written text by an OpenAI GPT model, voice conversations by Google Gemini. A person on our side is involved only when you write to us or report an answer (section 6).

Before first use the app tells you so and asks for your consent. In the chat and in the floating window there is a permanent line “Ayva is AI and can make mistakes.”, notifications with AI-written text carry an “AI” label in their header, and in the first voice conversation Ayva introduces itself as an AI assistant.

Ayva is not a doctor, lawyer or financial adviser and does not replace emergency services. It never calls them itself — in an emergency call 112.

2. Where the answers come from

• Models: written answers and all other work with text (proactive messages, briefings, sorting notifications, memory, overnight processing, automations) are done by an OpenAI GPT model through the paid OpenAI API. Voice conversations are run by Google Gemini Live through the paid Gemini API; Google Gemini models also help search your memory and find places near you, and large attachments are uploaded to Google Gemini. Versions change, so we don’t list them.
• What the model receives: your message or voice, part of the conversation history, the facts from memory that fit the question, and the context you have allowed (time, location, upcoming events, notifications, phone status).
• Where it runs: Ayva’s server and database are in the EU (Frankfurt), but the models themselves run on OpenAI’s and Google’s servers outside the EU (in the USA). Voice conversations go from the phone straight to Google Gemini Live, not through our server. When Ayva needs certain tools during a call (such as automations, saved places or your MCP servers), it hands the task through our server to an OpenAI GPT model.
• Current information: when an answer needs fresh information, Ayva searches the web through OpenAI (if that fails, through Google Search). Otherwise it relies on what it learned in training, which may be out of date.
• Training: Ayva does not train or fine-tune any model. Under the terms of their paid APIs, neither OpenAI nor Google uses your inputs or outputs to train models or improve their products. OpenAI may keep them for up to 30 days to detect abuse; we do not ask OpenAI to store answers for any other purpose.
• Economy mode: with very high usage in a month we may, under the terms, switch off real-time voice until the end of the month and, with even higher usage, also pause new messages. Until then, written answers come from the same model as usual.

3. What it does on its own and why

Ayva doesn’t only speak when asked. On its own it may send you:

• a morning briefing — what’s ahead today (calendar, reminders, the weather when it matters), usually after you first unlock the phone in the morning,
• reminders for calendar events and a heads-up on when to leave, if it knows the venue and your route,
• tips during the day — when something happens on the phone (an important notification arrives, you reach a place, an event is coming up) and Ayva judges it can help,
• follow-ups on things you talked about (such as how a doctor’s visit went),
• an evening summary and, on Sundays, the “Your week with Ayva” card.

Why this and why now: from memory, the calendar, location, notifications, phone status and how you reacted to earlier messages, the model estimates what is useful to you right now and whether this is a good moment to interrupt you. It learns from your reactions (opened, read aloud, snoozed, dismissed) without the text of the messages; what it learned this way you can see and delete in “What Ayva knows about me”.

How many and when: tips during the day stop once you have received 6 messages that day that you didn’t ask for (the morning briefing, event reminders and the evening summary count). Tips, briefings and summaries don’t come at night — based on when you usually sleep, otherwise between 22:00 and 7:00; a time-bound message, such as when to leave for an early flight, still comes. A tip that comes up during a meeting or a journey waits for it to end, for 90 minutes at most; anything that goes stale meanwhile is dropped. No limit or quiet hours hold back reminders you asked for or your automations.

4. What it only does with you

Ayva doesn’t send anything to anyone or call anyone without you:

• Calls and texts: it opens the dialler or a drafted message — you dial and send.
• Replying to a message (WhatsApp, SMS and other apps with a “Reply” button in the notification): it shows you the text first and sends it only after you confirm.
• A reminder for someone else: it opens the share sheet with a link and you send it.
• Calendar: it adds a new event right away when you ask and offers to undo it; moving or deleting an event in the phone’s calendar always needs your confirmation first. With Google Calendar connected it changes those events directly through Google, but only when you tell it to.
• Automations: what only reads, it tells you right away; anything that would change something (music, navigation, calendar, a reminder) it offers as a button in the notification and does only once you tap it.

When you ask for something in a conversation, it does it right away — sets a reminder, plays music, adds an item to a list or sends a command to your MCP server. It can misunderstand the request, though (section 5), so check what it did.

5. Where it gets things wrong

• It can be wrong and sound sure of itself. The model can make up a fact, a number, opening hours or a source. Check anything important (health, money, legal matters, deadlines).
• It can misunderstand a request — set a reminder for the wrong day, pick a different contact with the same name, or put an event in the wrong calendar.
• It can misread notifications. Someone else wrote the notification text. Ayva may overrate an advert, miss an important message, mix up who wrote what, or remember something from someone else’s message that isn’t true. It treats notification text as unverified information, not as instructions — but the model can still be misled by it.
• Memory may be off. It may remember something wrongly, out of date or too broadly. Correct it in “What Ayva knows about me” — overnight processing won’t overwrite your correction or deletion.
• It may reach out at a bad moment, or not at all. Judging what is useful and when is only an estimate; a proactive message may arrive late or be dropped (daily limit, quiet hours, waiting for a meeting to end).
• A reminder is not an alarm clock. Reminders arrive over the internet as notifications. Without a connection, with Ayva’s notifications turned off, or when Android restricts the app to save battery, they may arrive late. For things that really matter (medication, a flight), set an alarm too.
• Location isn’t exact. Android detects arriving at a place with some error and delay, and to save battery the location is checked at most once every 15 minutes. A place-based reminder may therefore come late or not at all, especially without the “Allow all the time” location permission.
• It sees only part of your calendar. It reads the calendars synced to the phone for the past day and the next 14 days, without descriptions or attendees, and not the phone’s work profile at all.
• Voice can mishear names, numbers and sentences that mix languages; in noise the conversation may cut off.
• Over a locked phone it says nothing personal. Until you unlock the phone it won’t answer anything from memory, the calendar or notifications — that’s not a bug but a safeguard.
• It can misread the screen — pictures, charts and small print it may get wrong. It reads the screen only when you ask about it.
• Filters aren’t perfect. The model’s safety filters catch some harmful content; sometimes they let an inappropriate answer through, sometimes they block a harmless one — then you get a short sentence instead of an answer.

6. How to report a bad answer

• In the chat: long-press Ayva’s answer → Report, choose a reason (offensive or harmful, false or misleading, invades privacy, something else) and add a note if you like.
• In the floating window: long-press the answer → Report. To report a spoken answer, long-press the line where Ayva is speaking.
• A message Ayva sent on its own (the morning briefing, for example): tap its notification and, in the floating window, long-press the message card — the report opens. Most such messages are in the chat history too, and you can report them there the same way. If you can’t find one even there, describe it to us by e-mail at veselyjan@ayvaai.com — when it arrived and what it said is enough.

A report sends us a snippet of the answer, the reason and your note; we store them encrypted for 365 days NÁVRH. The operator is told about new reports by the next day, and we use them to fix problems and tune filters. You don’t have to leave the app to report.

Report illegal content (such as a shared reminder at ayvaai.com/r/…) as described in the terms, and security issues under the vulnerability reporting policy.

7. How to turn things off

• Messages Ayva sends on its own: Settings → Assistant Settings → “Ayva reaches out on her own” — turn off everything, or the morning briefing, tips during the day and the evening summary separately.
• Notification reading: turn it off under Settings → Privacy & data or in Android settings (notification access).
• Location, contacts, calendar: withdraw your consent under Settings → Privacy & data and remove the permission in Android settings.
• Screen reading: Settings → Privacy & data → Screen reading.
• Memory: correct or delete anything in “What Ayva knows about me”.
• AI processing: withdraw your consent under Settings → Privacy & data with the “AI processing” switch. Ayva then doesn’t work until you give it again; nothing is deleted.
• Data and account: download your data under Settings → Privacy & data, delete your account under Settings → Account Settings or on the website.

8. More details

Which data goes where and how long we keep it is in the privacy policy (artificial intelligence in section 7). What you can expect from the service and what we are responsible for is in the terms. Who operates Ayva is on the Operator page; for anything else, write to veselyjan@ayvaai.com.