Four Milliseconds of Face

This is a design note, not legal advice. The cases below are summarised from published commentary and from the regulator’s own pages, and anyone building on them should take proper advice.

In October 2024 the Privacy Commissioner found that Bunnings had breached the Privacy Act by running facial recognition in its stores. On 4 February this year the Administrative Review Tribunal set that determination aside in part. It is easy to read that as a win for the retailer. The detail is more useful.

What the Tribunal kept

The Tribunal still found breaches of Australian Privacy Principles 1 and 5. The notices to customers were too generic, and there was no documented privacy risk assessment. It set aside the finding on APP 3. The most striking part for a designer is a technical one: the Tribunal accepted that “collection” can include a face that is held for about 4.17 milliseconds before being discarded.

Read that twice. Four milliseconds is not a moment any shopper will ever perceive. The law treats it as an event that needed notice.

The December deadline

The obligation to be transparent about automated decision-making under APP 1.7 to 1.9 starts on 10 December 2026. The OAIC published fact sheets on 30 September to help organisations get ready. Put the two together and the direction is clear: specific notice, written down, before the system runs.

The design problem hiding in the law

An ambient system is built so that nothing announces itself. That is the point of it, and it is exactly what a notice-based regime does not like. Helen Nissenbaum’s idea of contextual integrity is the best frame we have: information should flow in line with the norms of the context it came from. A shopper walking into a hardware store has a norm about what a hardware store knows. The four milliseconds broke it, and a sign by the door did not repair it.

A banner cannot fix that, and a link in a footer certainly cannot. What the law is asking for, in practice, is a consent surface that works in a room: something a person can see, understand and refuse at the moment it matters. That is an object, a light, a sound, a threshold. It is design work.

Who authorised the agent

There is a second front. In a Harvard JOLT Digest piece from 28 August, Zefang Wu argues that software agents acting for a user deserve a presumption of authorisation, drawing on the Perplexity cases, which have so far produced an injunction in March and a Ninth Circuit ruling in August. We take no view on the law. The design point is that “who said this agent could do that?” is now a live question in court, and an ambient system acts without being asked more often than any agent does.

What we would test

The Threshold Study we offer asks when a system is noticed. The same method can ask a sharper question: at what point does a person notice that they are being sensed, and does that point line up with where the law now expects notice? If a consent surface only works once the person is already uncomfortable, it has failed. If it works long before, it may be too loud to be ambient. The interesting design lives between those two.