What Makes an AI Assistant Actually Undetectable
The word invisible gets used loosely. In practice there are three very different levels of hiding, and only one survives a screen recording.
The three levels
- —A normal window: visible to everyone, hidden from nobody
- —A borderless always-on-top window: still captured by screen share
- —A compositor-level excluded overlay: absent from the capture stream itself
- —Process visibility: a separate question from display capture
- —Behaviour: what a human reviewer notices regardless of software
How Blackbird fits in
Screen sharing, recording and proctoring extensions all read the same Windows capture pipeline. A tool is only invisible if the OS leaves it out of that pipeline — everything else is cosmetic.
Blackbird is a lightweight native Windows overlay drawn at the system compositor level. It is not a browser tab, an Electron window, or a shared surface, so it does not show up in screen recordings, screen-share streams, or active-window checks. It stays on top, accepts clicks without stealing focus, and disappears instantly with a hotkey.
Practical tips
- +Test any tool by sharing your screen to a second device before trusting it.
- +Remember behaviour is half the equation: gaze, pace and phrasing.
- +Avoid pasting; that is what editor-based detection is built for.
- +Use one monitor when a platform bans a second display.
Frequently asked questions
Is anything truly undetectable?
Display capture can be handled reliably. Human observation and behavioural analysis cannot — that part is on you.
Do proctors scan running processes?
Some client-side lockdowns check names against a list of known remote-control and recording software.
Does a second monitor help?
Usually the opposite — several platforms end the session when they detect one.
How big is Blackbird?
It is a small native Windows overlay with no browser engine bundled.
You are responsible for following the rules of any exam, employer or institution you use Blackbird with. No tool can account for human observation or behavioural review.