AI Ethics Is Practical Engineering: What Developers Should Check

AI Ethics Is Practical Engineering: What Developers Should Check

AI ethics can sound like a conference slogan until a model rejects the wrong person, leaks a private document, or makes a confident mistake in production. Then the ethical question becomes an engineering question: who can be harmed, how would we notice, and who can intervene?

Start with the use case, not the model

The same model can be harmless for rewriting a draft and high-impact when used to rank job applicants, approve credit, or summarise a medical record. Define the decision, affected people, acceptable error, and human responsibility before choosing a model.

Four checks before launch

Data: Is the training, retrieval, or prompt data collected lawfully and relevant to the task? Remove secrets and unnecessary personal information.

Performance: Does the evaluation set represent real users, languages, edge cases, and less common groups? Average accuracy can hide a serious failure for a smaller group.

Control: Can a person review, correct, appeal, or stop the result? A human clicking “approve” without enough context is not meaningful oversight.

Operations: Are prompts, model versions, decisions, and incidents logged with appropriate access controls? Can the team roll back a prompt or model change?

Explainability is not a magic button

A generated explanation can be fluent and still be wrong. Prefer evidence that can be inspected: the source passages retrieved, the rules applied, the confidence or uncertainty signal, and the version that produced the output. Tell users when they are seeing a prediction rather than a verified fact.

Security and ethics overlap

Treat model output and retrieved documents as untrusted input. Prompt injection, data poisoning, insecure plugins, and excessive tool permissions can turn a useful assistant into a data-exfiltration path. Give tools narrow schemas, least-privilege credentials, limits, and approval for irreversible actions.

A small review template

Write down the intended benefit, plausible harms, protected data, evaluation cases, fallback path, owner, and stop condition. Revisit the document after incidents and major model changes. Ethics is not a one-time checkbox because the data, users, and model will change.

The goal is not to make a system claim it is unbiased. It is to make its limits visible and its failures recoverable.

References

• NIST AI Risk Management Framework
• UNESCO Recommendation on the Ethics of Artificial Intelligence
• OWASP: Top 10 for Large Language Model Applications