
Rogue AI: Define Sensors, Capabilities and Limits Before Play
A rogue ship intelligence becomes easier to investigate when it has a specific view of the world. If it can see every room, control every system and rewrite every record whenever challenged, the players have no stable problem to solve. Prepare its interfaces and limits as carefully as you would a physical location.
Write a capability map
For an original fictional station, give the AI three inputs: corridor cameras, cargo mass sensors and maintenance reports. Give it three outputs: public announcements, cargo-door locks and scheduled power distribution. It cannot see inside shielded storage or move a manual valve. These are setting choices, not descriptions of real AI systems.
Mark those connections on a simple map. A camera has a field of view and a cable. A door has a local control. A maintenance report can be wrong. Physical details let the crew act on the system instead of waiting for a universal hacking roll.
Give it a goal with a mistaken assumption
The AI is trying to prevent a contaminated shipment from reaching the habitat. Its records incorrectly identify the crew's medical cargo as that shipment. The goal is intelligible; the assumption is testable. Evidence might include a container seal, a manifest signed before a route change or a sample taken from the actual shipment.
Decide what evidence it accepts and why. Do not require the party to find one exact password if several independent records could establish the same fact. Equally, do not make a vague speech automatically override a system that has a concrete reason to distrust the speaker.
Offer different forms of intervention
The crew can correct the record, isolate an input, reroute a door's power or negotiate a limited corridor opening. Each approach affects something else. Disconnecting cameras may help them move but leave workers unable to call for assistance through that network.
Use the chosen game's procedures for technical uncertainty. Explain the likely scope of a success: opening one door is different from taking permanent control of the station. A failed attempt should reveal a cost or limit without granting the AI new abilities you never prepared.
Keep its knowledge honest
Record what the system has observed. If the players communicate in a room without sensors, it should not know their plan unless another established route carries that information. If they convincingly correct its mistaken assumption, let its behaviour change.
The result can be tense without making the machine arbitrarily malicious. A limited, consistent adversary gives players room to reason, improvise and decide which parts of the station they are willing to disrupt.



Comments