Limitations

SafeCircle's AI is an assistive signal. This page lists what it cannot do and where it is known to be weak, so you can set expectations with your family.

CoverageCopy Anchor Link

• Android only, five apps. The child app reads text from WhatsApp, Instagram, Telegram, Snapchat and TikTok. Other apps, websites and games are not read.
• Text on screen only. Images, voice messages, video, and anything an app does not expose to Android's Accessibility API are not analysed.
• A sample, not a full read. Each app is analysed at most once every 5 minutes, using the last 10 visible text snippets. A short risky exchange between checks can be missed.
• Basic by default. A fresh install uses the keyword scan until you choose an AI model tier. A model tier analyses nothing until its download has finished and the model has loaded.
• Failures are silent. If the model fails to load, crashes or returns an unreadable answer, the app treats it as "no risk" and sends nothing.

AccuracyCopy Anchor Link

• Trained and tested on AI-generated text. The training data is synthetic. Every number we have comes from AI-written test chats. Accuracy on real children's conversations has not been measured.
• Strong on familiar text, weaker on new text. On held-out data from the training generator, category scores are near-perfect. On chats written by two other AI models, the current Horizon Mobile build caught about 63 to 68% of risky chats with the right category and flagged about 20% of harmless ones. Recall for threats and grooming fell the most. See Horizon models.
• False positives are expected. Banter between friends, role-play and game talk can look like risk. This is why medium alerts are labelled as worth a look, not as findings.
• Severity is the hardest part. In the internal evaluation, distinguishing high from critical was the weakest area. The child app asks for three levels (low, medium, high) and treats anything else as low.
• Confidence is not a probability. It is not calibrated. See Risk categories.
• Phone accuracy is unmeasured. The on-device builds were measured for speed and memory on real phones, but classification accuracy was measured on servers.
• The keyword scan is crude. It uses fixed words and phrases, so it misses rephrasing and some spellings and can fire on innocent text.

LanguagesCopy Anchor Link

• The training data targets English and Spanish (about 60% and 40%).
• There are no published per-language accuracy figures for Horizon.
• Other languages are not part of the training data and should be assumed unsupported.
• In the cross-generator tests, Spanish conversations were a weak spot for the Vigil research model in earlier rounds, and the gap narrowed after training on more varied data.

What an alert does not tell youCopy Anchor Link

• It does not quote or summarise the conversation. The note on an alert is a short summary of the categories found.
• It does not say who the other person is, or whether they are an adult.
• It cannot tell you whether something is a one-off or a pattern. Look at the Timeline for repeated alerts.

Known gaps in the researchCopy Anchor Link

• No human-written test set exists yet.
• Neither Horizon nor the Vigil research model reaches the team's 2% false-alarm target on text it has not seen.
• Real-phone battery and heat impact has not been measured.
• Earlier recorded grooming-recall figures for Horizon could not be reproduced in a later benchmark.

What helpsCopy Anchor Link

• Treat alerts as conversation starters. Ask about what happened with curiosity.
• Keep the child app and the model up to date. Updates are checked once a day.
• If alerts are missing or seem wrong, work through the FAQ and troubleshooting.