How these notes are written

Structure

Most notes follow the same shape: what the thing is, why it is harder than it looks, the method, where the method misleads, and how to tell whether it worked.

The sections on what does not work are not optional. A description of detection accuracy that omits the base rate describes a number that will not survive deployment.

The least capable system

A principle running through the collection: for every proposed use, ask what the decision needs, then ask the least capable tier that could inform it.

An alarm when someone enters a hazardous zone needs detection. Understanding queue length needs counting. Neither needs identity. Deployments that start from what the technology can do rather than from the decision end up with data they did not need and obligations they did not plan for.

No benchmarks

Deliberately. Performance depends on cameras, lighting, scene, population and threshold. A figure from another deployment predicts nothing about yours, and published figures are self-selected.

Where a relationship is stated — that precision collapses for rare events, that errors fall unevenly across groups, that alert volume determines whether anyone reads alerts — that is the part worth relying on.

The uncomfortable material

Error rates are not evenly distributed across people, and an average conceals it entirely.

Existing camera estates are usually unsuitable, which is the most common reason deployments underperform and the thing least often in a proposal.

Detection is substituted for the physical control it was meant to supplement, and then the barrier is never installed.

Most emotion and intent products rest on contested science, and several are prohibited outright.

Legal material

Descriptions of the AI Act, data protection, biometric statutes and workplace rules are general and jurisdiction-dependent. They are not legal advice and the notes say so.

Commercial position

No vendor material. No sponsored content. No affiliate links. No product rankings.