Always-on-Recording

Always-on-recording means that a device records continuously – without anyone pressing a button or issuing a command. The microphone, camera, or another sensor simply runs permanently in the background. Well-known examples are voice assistants like Alexa or Siri: they listen constantly and wait for a specific wake word, for example “Hey Siri”. Dashcams in cars work similarly – they film the drive without interruption so that video evidence is always available in the event of an accident. The principle itself is therefore not new, but it is gaining significant importance through AI devices that must constantly gather context about their surroundings.

Why always-on-recording is indispensable for AI devices

Many AI features require the device to recognize the right moment on its own. An assistant that only reacts when a button is pressed misses spontaneous situations. An AI glasses prototype that tells you who is standing in front of you must keep the camera running constantly – otherwise it has no data when you need it.

The same applies to health wearables. A smartwatch that only measures heart rhythm or oxygen levels when prompted to do so will miss anomalies that occur precisely between two measurements. Always-on-recording is therefore not a comfort feature here, but a fundamental technical prerequisite for the function itself.

At the same time, this raises significant questions around data protection and privacy. Who processes the data, where it is stored, and for how long – none of this is always transparent. Regulators in the EU have therefore begun to address exactly this issue.

How always-on-recording is implemented technically

The obvious problem: a device that continuously records data and uploads everything to the cloud consumes a great deal of energy and generates enormous amounts of data. The practical solution is called local pre-processing. A small, energy-efficient chip in the device – often referred to as a “Neural Processing Unit” or NPU for short – analyzes the data stream directly on site. Only when it detects something relevant, for example the wake word or an unusual heartbeat, is data forwarded or stored.

This principle is called edge processing: the actual computing work takes place at the edge of the network, meaning in the device itself, not in a remote data center. Most of the recorded data is not permanently stored at all, but is overwritten after just a few seconds. What sounds like “continuous recording” is technically often a rolling short-term buffer.

This principle is especially evident with dashcams: the camera constantly overwrites the oldest recordings once the storage is full. Only when the impact sensor registers a collision does the device automatically save the relevant section before it can be overwritten.

Always-on-recording in products and current debates

In everyday life, one encounters always-on-recording on many levels today. Voice assistants in smart speakers are the most widespread example. In addition, there are video doorbells, baby monitors, AI glasses like the Ray-Ban Meta, and increasingly also laptops that use software to detect presence and automatically lock the screen when you get up.

In the tech press, the term mainly comes up when new devices with AI context features are introduced. In 2024, the startup Humane presented the so-called AI Pin – a brooch that continuously analyzes conversations and surroundings and summarizes them on request. Similar concepts, for example from Rewind AI, permanently recorded screen content in order to make the user’s digital memory searchable.

These products have sparked a broad debate: what happens if someone wears such a recorder without informing the other people in the conversation about it? In Germany, this is simply illegal without consent. Being technically possible and being legally permitted are two very different things – and this exact tension is at the center of the discussion around always-on-recording.