AI Ethics & Algorithmic Bias

Where bias actually enters an AI system

Most public conversation about AI bias starts at the output — a chatbot says something offensive, an image model renders a stereotype. That is the last and least interesting place to look. Bias is decided far earlier, in four places that rarely get audited.

• Provenance: whose documents, photographs, and speech made it into the corpus at all, and whose were never digitized.
• Labeling: who defined the categories, and what assumptions the label taxonomy encodes.
• Objective: what the model is optimized to maximize, and who absorbs the cost when it is wrong.
• Deployment: which population the system is pointed at, and whether they consented to being scored.

The under-recorded and over-surveilled problem

Working on the Living Archive Series — restoring American newspapers that documented Black civic life — makes one pattern impossible to miss. Enormous stretches of Black history exist only on decaying microfilm, in church bulletins, and in family collections that were never scanned. That absence becomes training data absence, which becomes model ignorance.

The same communities are simultaneously the most densely captured by modern surveillance systems: facial recognition, predictive policing, credit and tenant screening. So the record is thin where it would help and thick where it can be used against people. Any serious ethics practice has to hold both facts at once.

An audit checklist you can actually run

Ethics documents that no one can operationalize are decoration. These are the questions worth putting in front of a team before a model touches a real decision.

• Can you name the sources of your training data and the years they cover?
• What is the model's error rate broken out by group, not averaged across the population?
• What is the cost of a false positive to the person affected, and who pays it?
• Is there a documented appeal path with a named human decision-maker?
• If the system were pointed at your own family, would you accept the error rate?

Why this is a cultural question, not only a technical one

A model is a compression of a record, and a record is a set of decisions about what deserved to be written down. That makes AI ethics continuous with archival ethics and, in older language, with the ethics of testimony — who gets believed, and on whose authority. Traditions that have spent centuries transmitting knowledge without institutional permission have practical things to say here, and they are usually excluded from the conversation.