The unplug test: an acceptance test for private AI
Brian Niceley · July 2026
“Runs locally” is too easy to say. The unplug test turns it into a behavior: disconnect Ethernet, switch off Wi-Fi, and run the client’s real workflow from beginning to end. If it fails because a model, API, license check, font, or hidden dependency lives somewhere else, the system is not ready.
- 01Start from the normal, documented launch path.
- 02Disconnect every network interface the machine could use.
- 03Run a representative private workflow with local test data.
- 04Verify the answer, document retrieval, and any approved agent action.
- 05Open the local action receipt and confirm what happened.
- 06Restart the relevant services while still offline and run it again.
What it catches
The obvious failures are cloud inference and hosted embeddings. The useful failures are quieter: a package that downloads a model on first use, a remote identity check, a web font, a telemetry call that blocks startup, or an integration the architecture diagram forgot. A fresh restart matters because caches can make a connected system look self-contained once.
What it does not prove
Passing does not make a machine invulnerable or a workflow automatically compliant. It says the tested workflow completed without a network. Physical security, encryption, access control, backups, retention, professional duties, and untested code paths still need their own controls.
Why it belongs at handoff
The client should run the test with me, not receive a screenshot of me running it elsewhere. The point is not theater. It is to make the privacy claim repeatable by the person who owns the machine, using the workflow they are about to trust.