AI Ethics in Practice: What I Actually Do (Not What Conferences Talk About)

I used to roll my eyes at "AI ethics." Not because it doesn't matter — it does, enormously — but because most discussions about it are so abstract they're useless. "Be fair." "Be transparent." "Avoid bias." Great. How?

Then I built a resume screening system that scored female candidates 12% lower than male candidates with identical qualifications. Not because I wanted it to — because the training data reflected 10 years of biased hiring decisions, and the model learned the pattern perfectly.

That was my wake-up call. AI ethics isn't philosophy. It's engineering. It's testing. It's catching problems before they reach users. Here's what I actually do in practice.

For the business case for ethics, see my companion post: Responsible AI as Competitive Advantage.

The Four Things That Actually Matter

After building AI systems for years, I've distilled ethics into four practical concerns. Not because other concerns don't exist, but because these are the ones that actually cause harm in production systems.

1. Bias: Your Model Learned Your Organization's Worst Habits

Every dataset reflects the biases of the people and processes that created it. Historical hiring data? Biased. Medical records? Different treatment patterns by demographics. Credit decisions? Decades of discriminatory lending baked in.

What I do: Before any model touches real decisions, I run bias audits across demographic groups. Not because regulators require it (though they increasingly do), but because a biased model is a broken model.

The resume screening fix was straightforward once we identified the problem: we retrained on a balanced dataset, added demographic parity constraints, and now audit monthly. The model is more accurate and more fair — because evaluating candidates on qualifications rather than gender-correlated patterns is just better prediction.

Minimum viable bias testing: Compare your model's outcomes across demographic groups. If accuracy or acceptance rates differ by more than 5-10%, investigate. You probably have a data problem.

2. Transparency: Users Deserve to Know When AI Is Making Decisions

I have a simple rule: if an AI system makes or influences a decision about a person, that person should know. Not buried in a terms of service — explicitly, at the point of the decision.

This means:

• "This recommendation was generated by AI" labels on content
• "Your application was processed using an automated screening system" disclosures
• Source citations on AI-generated answers (see my hallucination prevention guide)
• Clear explanations of why a decision was made, not just what the decision was

The practical challenge: explainability in deep learning is hard. A transformer model can't easily tell you "I rejected this candidate because of X." But you can build proxy explanations: "The key factors in this decision were: years of experience (high impact), relevant skills match (medium impact), education (low impact)." Not perfect, but infinitely better than a black box.

3. Privacy: Collect Less, Protect More

The AI industry's appetite for data is insatiable. More data → better models → more value. The problem: that data belongs to people, and those people increasingly care about how it's used.

My practical approach:

• Collect only what you need. Don't vacuum up data "just in case." Every data point you store is a liability.
• Anonymize aggressively. Hash IPs, strip PII, aggregate where possible. My view tracking system uses session-based anonymous tracking — no user accounts, no personal data.
• Differential privacy for sensitive models. If your model trains on personal data, add noise to prevent individual records from being extractable. It costs 2-5% accuracy. Worth it.
• Data retention policies. Delete what you don't need. If training data is older than 2 years, question whether it's still representative.

4. Accountability: Someone Has to Own This

Every AI system that makes consequential decisions needs a human who is accountable for those decisions. Not "the algorithm decided" — a specific person who signed off on the system's design, monitored its performance, and can be asked to explain its behavior.

In practice, this means:

• Document your design decisions. Why this model? Why this training data? Why these thresholds? Write it down. Future you (or a regulator) will thank you.
• Human-in-the-loop for high-stakes decisions. Automated screening is fine. Automated rejection without human review is not. See my agent guide for how I implement human oversight.
• Regular audits. Not just at launch — quarterly. Models drift. Data changes. The system that was fair in January might not be fair in July.

The Regulatory Reality

I won't pretend to be a lawyer, but here's what's relevant for builders:

EU AI Act (2025+): Classifies AI systems by risk level. High-risk systems (hiring, credit, healthcare) need conformity assessments, documentation, and ongoing monitoring. If you're building for European markets, start classifying your systems now.

US (fragmented): No comprehensive federal AI law, but sector-specific rules are tightening. EEOC is watching AI hiring tools. FDA is scrutinizing AI medical devices. FTC is enforcing against misleading AI claims.

My advice: Build for the strictest regulation you might face. If you document your systems, test for bias, and maintain transparency now, you'll be ready regardless of what laws pass. It's cheaper to build responsibly from the start than to retrofit compliance later.

The 30-Minute Ethics Check

I don't expect every team to hire an AI ethics officer. But every team shipping AI should spend 30 minutes on these questions before launch:

1. Who does this system affect? List every group of people who will be impacted by this system's decisions.
2. What happens when it's wrong? For each group, what's the worst case if the model makes an error?
3. Could it affect different groups differently? Run your outputs through demographic slices. Look for disparities.
4. Would you be comfortable if your decision process was on the front page? If not, fix it before someone puts it there.
5. Who's accountable? Name the person. If nobody wants their name on it, that's a sign.

This won't catch everything. But it catches the obvious problems that most teams ship without thinking about. And those obvious problems are the ones that cause the biggest damage.

---

Building an AI system that makes decisions about people? I help teams build fair, transparent systems that hold up to scrutiny. Let's talk.