# How the child app works (/en/docs/platform/android/how-it-works)



The child app watches for risk signals in a few messaging apps, analyses them **on the child's phone**, and sends you a short summary only when something looks risky.

## From screen to alert [#from-screen-to-alert]

<Steps>
  <Step>
    ### Read [#read]

    The app uses Android's **Accessibility Service** to read the text visible on screen in the [supported apps](/en/docs/platform/android/supported-apps), and only in those apps. For each app it keeps the most recent lines of visible text (up to 10) in memory.
  </Step>

  <Step>
    ### Analyse on the device [#analyse-on-the-device]

    Every so often it analyses that text locally. Nothing is sent to a server for analysis. It looks for these categories: grooming, bullying, sexual content, isolation, personal information, moving the conversation to another platform, and threats.
  </Step>

  <Step>
    ### Decide [#decide]

    The result is a risk level — **low**, **medium** or **high**. Low results are discarded on the phone. Only **medium** and **high** create an alert.
  </Step>

  <Step>
    ### Send a summary [#send-a-summary]

    The app sends the alert to the SafeCircle backend, and it appears in your [dashboard](/en/docs/platform/dashboard/alerts).
  </Step>
</Steps>

## What is sent [#what-is-sent]

An alert contains:

* The risk level (medium or high)
* The detected categories
* A confidence score
* A short reason, such as which categories were detected
* The name of the app where it was detected (for example `com.whatsapp`)

It does **not** contain the message text, contact names or screenshots, and the conversation is not stored by the app.

## How often [#how-often]

To avoid flooding you, the app analyses each monitored app at most **once every 5 minutes**. Alerts are created from what is on screen at that moment.

<Callout type="info">
  The app doesn't queue alerts. If the phone has no connection when an alert would be sent, that alert isn't resent later.
</Callout>

## Basic mode and the AI model [#basic-mode-and-the-ai-model]

The app has two ways to detect risks.

| Mode                     | How it works                                       | Download                             | Accuracy                                     |
| ------------------------ | -------------------------------------------------- | ------------------------------------ | -------------------------------------------- |
| **Basic** (keyword scan) | Looks for risk-related phrases in the visible text | None                                 | Lower — expect missed risks and false alarms |
| **Horizon Mobile INT4**  | On-device AI model, 4-bit                          | About 600 MB (3 GB+ RAM recommended) | Higher                                       |
| **Horizon Mobile INT8**  | On-device AI model, 8-bit                          | About 1.2 GB (6 GB+ RAM recommended) | Highest                                      |

A new installation starts in **Basic** mode, and the status screen shows **Basic protection mode — AI model not installed — using keyword scanning**. The AI models are optional and run entirely on the device.

### Switch to an AI model [#switch-to-an-ai-model]

<Steps>
  <Step>
    ### Open the tier setting [#open-the-tier-setting]

    In the child app, open **Settings** and tap **AI Model tier** under **Advanced**.
  </Step>

  <Step>
    ### Pick a model [#pick-a-model]

    In **Change AI Model**, choose a tier. The one that suits the phone's memory and storage is marked **Recommended**. Tap **Apply**.
  </Step>

  <Step>
    ### Accept the model licence [#accept-the-model-licence]

    Before the download, a **Model License Agreement** sheet shows the SafeCircle Research License (SRL-1.0). Tick the box and tap **Accept & Continue**. If you tap **Decline**, the app stays in Basic mode.
  </Step>

  <Step>
    ### Wait for the download [#wait-for-the-download]

    Keep the phone on a good connection. The download continues in the background and you can leave the screen. When it finishes, **Settings › AI Model** shows **Active**.
  </Step>
</Steps>

Choosing a tier deletes the model file currently on the phone and downloads the new one. To go back, pick **Basic (keyword scan)** and confirm **Switch to Basic**. The app warns that detection accuracy is significantly reduced.

See [Updates](/en/docs/platform/android/updates) for how model updates arrive, and [How the AI works](/en/docs/ai/overview) for more on the technology.

## Limits [#limits]

* Only the [supported apps](/en/docs/platform/android/supported-apps) are monitored.
* Detection isn't perfect. Use alerts as a prompt to talk to your child, not as proof.
* Protection stops while the Accessibility Service is off. The status shows **PERMISSION MISSING** until it is enabled again.
* Android 11 or later is required.

More in [Known limitations](/en/docs/support/known-limitations).