What Face ID for My Mac Taught Me About Trusting Machines
I built my own face unlock for macOS. The recognition was the easy part. Measuring it honestly is what changed how I read every claim in AI.
Liam Coverdale · 7 April 2026 · 2 min read
My Mac unlocks when it sees my face. Not with Apple hardware, but with Doorman, a thing I built: a camera-based recogniser that cooperates with macOS rather than fighting it. It works and I use it every day. Building it permanently changed how I evaluate any "it works" claim in AI, including my own.
The afternoon version
The first version worked in an afternoon. Face recognition is a solved problem in 2026, and the careful part was integrating it respectfully with the operating system's security model. I sat down, it recognised me, the desktop appeared. I nearly filed it under done. Then I asked the question that ruined my week: how do I know this is actually any good?
Working and measured are different claims
Answering honestly meant finding, one at a time, the traps in my own testing. I'd been measuring myself under a single set of conditions. Same chair, same lighting, same time of day. I'd been counting successes and treating failures as anecdotes. Every trap had the same shape. The system wasn't wrong, my measurement was, and always in the direction that flattered the system.
"It worked when I tried it" is the weakest sentence in engineering and the most common one in AI.
So I rebuilt the unglamorous part. Log every attempt, pass and fail, across conditions I didn't get to handpick. The honest numbers came in worse than the demo had felt, and the honest numbers were the only ones worth having.
The part I still haven't measured
Full disclosure, because it's the point of this piece. I haven't rigorously measured the spoofing question, meaning the adversarial cases every biometric system eventually has to answer. I have informal observations. I don't have data. Until I do, the honest description of Doorman is "a working convenience with an unmeasured attack surface", not "security".
Saying that plainly costs me nothing and buys me the right to trust my own claims elsewhere. The alternative is quietly upgrading "unmeasured" to "probably fine", which is how security theatre gets built by people who meant well.
The AI parallel
Map all of that onto AI, because it's the same trap with bigger stakes. A demo that works when the builder runs it. Evaluation conditions chosen by the party being evaluated. Successes logged, failures anecdotal. And an attack surface, whether that's prompt injection or adversarial inputs or the confidently wrong answer, that everyone has opinions about and almost nobody has measured on their own system.
When I handed an AI the keys to my machine, the same discipline applied. Approval logs rather than impressions. Track record rather than vibes. Trust tiers that widen on evidence. I trust my tools exactly as far as I've measured them, plus a small overdraft I'm honest with myself about.
Trust isn't a feeling. It's a measurement with a maintenance schedule. Doorman still opens my Mac every morning, and every morning it quietly tells me what the number is.