False wakes are the main destroyer of trust
Aliases: unsolicited activation · trust-breaking false accept · unexpected speech
What it is
Household trust in an always-on speaker is a prediction: “unless I call it, it stays quiet and does not treat this talk as a session.” A false wake falsifies that prediction in front of the room: the device speaks, lights, maybe even answers when no one called. As an input event, the same trigger may also start capture. As a felt event, it is the main way trust is taken apart — not because it is the most frequent failure, but because it is one of the few pieces of evidence that the whole room can hear: “so it was listening.” The stake here is felt: will people still leave it powered and unmuted.
Why it happens
Trust is an expectation about the idle state. A true wake matches the expectation and can even strengthen “it works when I call.” A false wake is a prediction error, and the error is public: the whole room hears “I didn’t catch that.” A silent buffer overrun has no audience, so it does not update the model. An audible false wake has an audience, and the model shifts from “it listens when called” to “it might be listening to this conversation.” Once that update lands, every later private topic is re-priced, even when the engineering false-accept rate is not high. Cute apology copy cannot repair it: the copy admits “I interrupted” while proving “I can enter the conversation without being called.” What people then do — unplug, mute, move it out of the bedroom — repairs feeling, not the detection threshold.
A collision during talk about money, bodies, or relationships updates harder than one while music is playing in the living room. Sensitivity of content sets the size of the update, not the raw count of triggers.
Studying it
Take critical incidents: when did the first false wake after install happen, what was being said, did the device then get muted or moved. Dependent measures are days still powered and unmuted, and whether the interview model of “is it listening” changed — not lab false-accept counts. Counts are a detector metric; the changed model is the trust metric.
In diaries, mark false wakes as spoken / lamp-only / found later in history. The three do not destroy trust equally: spoken is worst, history-only with no one in the room to hear it is weakest. Score “startled” and “then changed placement” separately, so fright is not treated as the inverse of success while quiet removal is missed.
Where it stops holding
Demo modes and retail-floor chatter do not enter the household trust ledger. The user said the wake word and then changed their mind: that is not a false wake. Strictly local spotting that never uploads still interrupts by speaking, but the half-update “the words already went to the cloud” is weaker. A dictation keyboard with no speaker can rarely falsify the prediction in public; the path of damage differs. For someone who already distrusts the device, a new false wake is confirmation, not a turning point.
Applying it
- When a false wake speaks, offer a stop as strong as a true wake, and say the buffer will be discarded. Do not recast the falsifying event as personality with a witty apology.
- Treat mute, unplug, and relocation in the seven days after an audible false wake as a trust incident, not only as a quality blip on the false-accept chart.
- In rooms where money, health, or family conflict are talked, default to a stricter wake threshold and accept more missed wakes. Do not copy the living-room threshold into the bedroom.
- How to check: post-install visits ask “what did you do after it last spoke on its own,” not “do you think it is accurate.” Mute or move is handled as trust destruction, not as one recognition miss.