Most of what gets filed under "AI ethics" is useless to the people who actually build the systems. It's either abstract debate about superintelligence, which is a luxury problem, or a corporate "AI principles" page that no engineer has read and no engineer ever will. Neither one stops a bad system from shipping.
What stops bad systems shipping is habit — the small checks a careful practitioner runs almost without thinking, the way a surgeon scrubs in. You don't need a philosophy degree. You need to want the work done properly. Here is the list I keep returning to, after thirty years of watching where things go wrong.
Write down who pays when it fails. Every system has failure modes; the only question is who bears the cost. A recommender getting it wrong annoys someone. A diagnostic tool getting it wrong harms someone. A parole-prediction tool getting it wrong takes someone's freedom. Before you build, name the person who pays for a mistake. If it's someone with no power in the situation, stop and ask whether you're the right team to be doing this at all.
Look at your data as the person it excludes. Every dataset has a silent demographic — the people who aren't in it. Train face recognition on Western social media photos and it will work worse for older people, darker-skinned people, and anyone who doesn't post pictures of themselves. This isn't hypothetical; it has happened repeatedly and expensively. The habit is cheap: before training, picture the person your data doesn't know about and ask whether your system fails them without anyone noticing.
Show what the system learned, not just what it decided. A model that outputs "denied" with no reasoning cannot be challenged, audited or improved. "Denied because feature X scored 0.34 against a 0.5 threshold" is a different posture entirely, even when the feature is debatable. Explainability isn't solved, but "we tried and here's what we can tell you" beats "trust the black box" every single time.
Treat private data the way you'd want yours treated. Sounds obvious; isn't, in practice. The test I use is blunt — would I be comfortable with my own mother's medical records, loan application or emails being handled the way we're handling this? Sometimes the answer is fine: anonymised, limited-purpose, careful. Running the check keeps you straight.
Plan the exit before you collect a single record. Far too many systems have no end-of-life plan, so the data outlives everyone's memory of why it was gathered. Decide how it gets deleted, write it down, set a reminder you'll actually honour. Data you forgot you had is data that will eventually embarrass you.
Say no sometimes. The hardest one. A client will eventually ask for something that sits wrong — a predictive policing tool, a sentiment-scoring HR system, a dark pattern dressed as personalisation. Sometimes the discomfort is worth working through. Sometimes it's the signal to walk. We've turned down work. It cost money. It has never once cost me a night's sleep, and I can't say the same about the jobs I should have refused and didn't.
That's the whole list. Notice none of it requires reading a paper on alignment.
The existential debates are interesting and I don't dismiss them, but they're not where a working team should spend its ethics budget. When a system does something genuinely harmful, it won't be because nobody read the right philosophy. It'll be because someone, on an ordinary afternoon, skipped the question a more experienced colleague would have asked. The habits above are how you become that colleague — by running the checks until they're reflex. Do that and you'll catch the problem at the design stage, which is the only stage where catching it is cheap.
