Nuha

Content that is upsetting by nature

Nuha detects hate speech and harmful content in Arabic, in the context of gender-based violence online. Everyone using it is, by definition, looking at material that is distressing.

That made the emotional weight of what appears on screen an actual requirement rather than polish applied afterwards.

Presenting a probability as a probability

A classifier returns a probability. An interface that renders it as a verdict — flagged, not flagged — is lying by omission, and someone deciding what to do about a message needs to know how confident the system actually is. Burying the uncertainty in a number nobody reads is not much better.

So the work was calibration: honest about what the system knows, legible to someone who is not a machine-learning engineer, and not adding to the distress of the content itself.

Expanding across several Arabic dialects and models made the same question harder. A tool that behaves identically in Modern Standard Arabic and in a regional dialect is either very carefully designed or quietly wrong.

Nuha is what made me understand that in some domains, the interface is an ethical decision at every level.

← All work · Next →