I Don't Need an AI Ethics Lecture. I Need a Checklist.
2026-02-01 · 12 min read · Janaina Maia
Most AI ethics frameworks I've read are useless for practicing designers.
There. I said it.
They're long documents written by policy teams, filled with abstract principles like "be fair" and "avoid harm" that give you zero guidance when you're in Figma at 3pm on a Wednesday trying to decide how to display a risk score.
I don't need a philosophy lecture. I need a checklist I can use while I'm working.
So I built one.
When to Use It
Not at the end. Not as a final review before launch when it's too late to change anything meaningful. At three specific points:
- Feature definition — Before you design anything
- Design review — During your crit
- Pre-launch — Final check with real-world testing results
Section 1: Data and Bias
- ☐ Can you describe the training data in plain language? If not, you can't assess bias.
- ☐ Who's missing from the data? What happens when an underrepresented case shows up?
- ☐ Can the AI's output perpetuate historical bias?
- ☐ Is there a bias monitoring plan — not just at launch, but ongoing?
Section 2: Transparency
- ☐ Can users see WHY the AI made its recommendation?
- ☐ Is AI-generated content clearly labeled as AI?
- ☐ Are confidence levels visible?
- ☐ Can a non-technical stakeholder understand the explanation?
Section 3: Autonomy and Control
- ☐ Can users override the AI? Every recommendation. No exceptions.
- ☐ Is override easy or buried? (If it's 5 clicks deep, nobody will do it.)
- ☐ Can users opt out of AI entirely?
- ☐ Does the system respect professional judgment in domains with liability?
Section 4: Impact and Consequences
- ☐ What's the worst case if the AI is wrong? Map out the failure cascade.
- ☐ Who is affected beyond just the user? Clients, communities, the public?
- ☐ Are high-stakes decisions flagged for human review?
- ☐ Have you defined unacceptable error rates — explicit thresholds where the feature should be disabled?
Section 5: Accountability
- ☐ Is there an audit trail? Every AI decision logged.
- ☐ Who is responsible when the AI is wrong? (Not "the AI." A human.)
- ☐ Is there an escalation path when users discover errors?
- ☐ Is there a kill switch? Not a code deployment — an operational toggle.
Making It Stick
The checklist is useless if it lives in a doc nobody opens. Here's how to actually integrate it:
In design critiques: Add a standing 5-minute agenda item. "Ethics check — anything we should pressure-test?"
In your design system: Create components for ethical patterns — AI attribution labels, confidence indicators, override buttons. When they're in the system, they get used.
In sprint rituals: Add acceptance criteria like "User must be able to override AI in ≤2 clicks" and "Audit trail must capture model version."
In your team: Designate a rotating "ethics champion" whose job is to bring up the uncomfortable questions everyone's avoiding.
Fair warning: this checklist will surface difficult conversations. You'll discover nobody knows what the training data looks like. You'll find there's no plan for bias monitoring. You'll learn the "human review" process is a rubber stamp.
Good. Those conversations need to happen now, not after launch.
Your AI product will reflect your team's values whether you're deliberate about it or not. The checklist just makes sure you're deliberate.