Your Own Ethical Checklist as an AI Builder

From Theory to Shipping

You have spent lessons learning about bias, copyright, safety, alignment, and regulation. If you actually build and ship AI — even a personal project with a few users — all of it turns into concrete decisions you make at your keyboard. This lesson is the working checklist, not a lecture.

Before you build

1. Problem definition: what specific decision will your AI make? Who is affected?
2. Stakes: what happens to a person if it is wrong? Can they appeal?
3. Regulated domain check: is this hiring, credit, healthcare, law enforcement, education, migration? If yes, EU AI Act high-risk obligations plus local equivalents apply.
4. Data legality: do you have rights to the training/fine-tuning data? Did users consent to their data being used?
5. Alternatives: is a simpler non-AI solution actually better here?

While you build

1. Baselines: measure performance not just overall but across relevant demographic slices.
2. Ground-truth leak check: verify your test set is not in your training set.
3. Human-in-the-loop design: default to advising decisions, not making them.
4. Explainability: can you tell an affected user what the model considered?
5. Red-team pass: try to make it produce harmful outputs yourself before shipping.
6. Model card: write a one-page description of intended use, training data, limitations, evaluation results.

At launch

1. Transparency: users know they are interacting with AI (EU AI Act Art. 50).
2. Opt-outs: users can refuse AI involvement where it matters.
3. Incident channel: a clear way to report harms, with a real human reading it.
4. Kill switch: you can turn it off quickly if things go sideways.
5. Logging: enough to audit decisions after the fact, within privacy law.

After launch

1. Drift monitoring: performance on representative slices over time.
2. Feedback integration: real user complaints are the most valuable signal you have.
3. Retraining cadence: plan it, do not wait for a crisis.
4. Disclosure: publish periodic stats on decisions, overrides, and corrections.
5. Sunset plan: what happens when this product ends? Do users get their data back?

The ten questions to answer before you ship

Red flags that mean rethink, not ship

• You cannot explain why the model did what it did on a failure case
• Evaluation was only done on the group you expected to use it
• Opt-out exists only in a cookie banner
• Monitoring plan is 'we'll check back in a quarter'
• Stakeholder consultation was limited to your team
• The business model depends on users not fully understanding what is happening

What this does NOT mean

• Refuse to build. Most AI is net positive when built well.
• Add friction for its own sake. Bad UX is not ethics.
• Endless committee approval. Speed matters; so does rigor.
• Treat users as hostile. Assume good faith, design for the edges.
• Chase every new framework. Pick one solid one and apply it consistently.

Ethics is not a checklist you pass. It is the thing you do between the checklist items, when nobody is watching.

— Margaret Mitchell, co-author of Model Cards paper

The big idea: AI ethics as a builder is not philosophy. It is a set of engineering decisions, made early, revisited often. The builders who think about these questions at the start ship better products. The ones who do not end up in a headline they did not want. You choose which story you want to be part of.