All guides
    Mission
    5 min read

    Why Navōn filters on the server and not in the browser

    Where a filter runs decides whether it can be switched off, bypassed or forgotten. Navōn runs its checks on its own servers, and here is what that changes for a parent or a school.

    Why Navōn filters on the server and not in the browser

    When a parent asks how a kosher AI platform keeps content out, the first useful follow-up question is not what the rules are. It is where they run. A rule that lives in the browser is a suggestion made to software the user controls. A rule that lives on the server is a condition of getting an answer at all. Navōn puts its checks on the server, and this post explains why that one decision shapes most of what follows.

    What a browser-side filter can't hold

    A filter that runs in the browser, whether an extension, a managed profile or a script inside the page, depends on the device and on the person using it. It stops working when the child opens a different browser, a different phone or a friend's laptop. It can be disabled in settings, removed, or sidestepped by a page that loads the same tool another way. None of that requires cleverness. It only requires a second device. Our guide to server-side versus device filtering goes through the school version of this problem in more detail.

    There is a quieter weakness too. A browser-side check sees what is on the screen. The thing that matters for AI is what is sent to the model and what the model sends back, and a filter in the page is only ever a guess at those.

    What runs on Navōn's servers

    In Navōn the browser or app is only a window. When you send a message, it goes to Navōn's servers first, and the checks happen there before anything is handed to an answering model:

    • The request is screened against the content policy before a model sees it. If it fails, the request is not forwarded.
    • The answer is scanned again before it is shown, so a model that drifts off course doesn't reach the screen unchecked.
    • Pictures are checked as part of the same flow rather than left to the app.
    • Moments where a child writes about distress are recognised in the same shared place, so the guardian alert does not depend on each feature remembering to raise it.

    Because the app never talks to a model directly, there is no setting on the device that turns this off. A student can use a different phone and still be on the same account with the same rules. The Features page describes this in a single line: every response is checked before it reaches you, on our servers rather than in the browser.

    One place to get it right

    Server-side placement has a benefit that is easy to overlook: there is one set of checks to maintain, not a copy per app and per browser. Our own engineering notes make this point about the guardian alert. If every part of the product had to remember to do the safety-critical step itself, sooner or later one would forget. Keeping the step in a shared module means that a new feature inherits it.

    It also means the rules a school sets are applied in the same place. A school's standards, described in our post on how a school sets its own standards in Navōn, are enforced where the request is checked, not by a setting that a device might fail to load.

    What this does not mean

    Server-side does not mean perfect. The checks are made by AI, and AI classifiers make mistakes in both directions: they sometimes stop a harmless question and they can miss something. We do not claim absolute safety, and we say so in our guide to what to do when something gets through the filter. What the placement gives you is consistency. The same check applies whichever device is used, and it cannot be quietly removed from the child's side.

    It also does not replace the rest of the home. Server-side filtering covers what happens inside Navōn. It says nothing about the other apps on a phone, which is why a browser or network filter and a conversation about expectations still have their place.

    A cost we accept

    Putting checks in the path of every request adds a step, and it means that when the checking service cannot run, nothing is released. We chose that on purpose, and the post on why Navōn fails closed explains the trade. A filter that can only be trusted while everything is working is not much of a filter.

    See what runs on every request and every answer in Navōn.

    Common questions

    Is Navōn's filter a browser extension?

    No. The checks run on Navōn's servers before a request reaches a model and before an answer is shown, so there is nothing on the device to disable.

    Can a child bypass an AI filter by using another device?

    A filter that lives on one device can be bypassed that way. Navōn checks on the server, so the same account has the same rules on any phone or computer.

    What is the difference between server-side and client-side filtering?

    Client-side filtering runs on the user's device and can be switched off or avoided. Server-side filtering runs on the provider's systems before content is sent or shown.

    Does server-side filtering make AI completely safe?

    No. The checks can make mistakes in both directions. Server-side placement makes them consistent, not flawless.

    Published by Navōn. How these guides are written and checked.

    Read next