For starters: your setup requires a good bit of human incompetence.
If the biometric tests are not happening on the same network as the AIs live, or they operating in a read-only capacity, then this question is a non-starter. Since the AI has no physical body, the AI has no way of giving the authentication system input. Even if the station's IT team has a very specific reason they can not isolate the video camera, you still have the option of point-to-point encryption. As long as the camera encrypts the video, and the security server decrypts it, then the AI cannot use a man-in-the-middle attack to inject content. So, for this to even begin to be a problem, we have to assume that the station's IT team is pretty sub-par for the AI to even be able to find an exploit in the station's camera systems that allow it to inject images.
That said, your IT team does not need to be security experts for the security features that are already a part of many modern biometric scanners to be really good at proving images are falsified.
Anti-spoofing tech is currently advancing faster than spoofing techniques.
Many communications companies are currently prioritizing the use of anti-spoofing image detection into consumer devices like cell phones and tablets. This tells us that the likely future trend is to assume that nearly all communication devices have nearly uncrackable anti-spoofing systems built into them, the same way nearly all websites now operate under nearly uncrackable TLS connections.
High-end biometric systems include many techniques that can be used together for recognizing spoofs that most humans can not spot. Any one of them used alone, the AI could beat, but as you stack them, the AI will eventually be unable to overcome its own limitations.
The most important of these features we see emerging is the use of active flash confirmation. As some biometric cameras takes an video of a person, it flashes a complex, randomized pattern of infrared lights on them and time stamps each frame as it comes in. When it proccessess the video of the subject's face, it looks for patterns of reflected light and shadows in the image that match the patterns that were projected on the person.
If an AI were to simply pass a pre-compiled clip of a person to the authentication system, the light patterns would be missing and the system would know that the image was forged. If the AI were to intercept the light projection specifications when they are sent, it would need to spend a moment computing all the right raytracings to generate the false video. This would create a time delay which would not match up with the timestamps; so, the system would know that the image was forged. Since the light patterns are declared by the recipient of the communication, and not the sender, injecting masterfully faked imagery alone will not overcome the system.
External communication is where things get a little tricky. Let's say you are communicating with a ship that is outside of the station, and the AI wants to fool someone on the ship, light pattern verification can still be used to prevent pre-compiled video, but unpredictable latencies could make the times stamp verification unreliable. The AI could just fake a longer latency giving it time to compile the video... or could it?
As data flows from the ship, it first hits the access point, then the network, then the camera, and back again. But when the AI becomes a MITM, data will flow from the access point, to the network, to the AI, back to the network, then to the camera and back again. By using its position in the chain the AI could falsify any time stamps that flows through the it, but not the access point. If your communication system pings the access point directly, you have an honest TTL on the lag over distance; so, if the AI tries to hide proccessessing time through faked latency, the direct pings to the access point will reveal it.
In summary, active flash detection puts a very tight time constraint on the AI giving it only milliseconds to generate flawless fakes. Then you stack that on top of other technologies like deep learning, 3d camera verification, biomechanical verification, projection smear detection, rendering artifact detection, etc, it is really easy to layer up until you have enough confirmation tools in place that it takes several minutes for the AI to falsify a frame in a way that fools the interpreter.
Then there is ofcourse the really low-tech solution...
In our age of computer controlled everything, we often forget that technology is just hardware. If you give it power, it works, if you deprive them of power, it does not. Let's say for example you don't want an AI to be able to gain control of your communications system, you just need to give it a power button. Then the ultimate test of human-hood becomes a simple "Can you press the button?" A human can hold down a button that gives power to the communications controls. If a person is not holding the button down, then an AI can not control it no matter how badly compromised your computer systems are.