Users adjust their behaviour to fit how automation judges them

What it is

Automation changes people in more than one direction: users don't just use automation, they adjust their own behaviour to fit how it judges — waving at the sensor so it registers presence, stepping closer to the speaker, issuing commands in a fixed phrasing, beating the automation by acting manually just before its window opens. This orbiting adjustment around algorithmic systems is the real shape of ambient automation in deployment: what gets installed is not just a system, but a set of daily routines reorganised around it.

It tends to be filed as anecdote ("my mother thanks the speaker by name"); it is in fact a serious design signal: workarounds are the most honest requirements document a system's defects will ever get — every obstacle users route around is a spot the recognition, interaction, or feedback design failed to cover.

Why it happens

How does adaptation grow? Three steps.

Opaque judgement boundaries → trial-and-error probing. The system's optimal operating zone (pickup range, recognised grammar, sensor angle, time windows) is written nowhere, so users probe. What the probing yields — "from that corner you must repeat it three times", "don't address it in that room" — settles in as personal routine. The routine organises itself around what I believe it does, not what it actually does, and the gap between the two is precisely the folk theory: naive explanation filling the vacuum a black box leaves.

Variance between success and failure → users learn to avoid the failure zones first. Sensors and recognisers have working-zone boundaries, and users wander in and out, generating variance. Humans outpace the machine at this learning: once the failure zones are reliably avoided, only successes remain — which looks like "mastered it", but is in fact the user absorbing the system's defects into muscle memory.

Hidden cost → the adaptation goes unnoticed. Each small adaptation is cheap (one step closer, one repetition, one rephrasing); together they form a standing behavioural tax. And once adapted, behaviour becomes the new "normal" — users no longer notice they are accommodating. Asked "anything inconvenient?", the answer is "no, not really" — the tax has been internalised.

Studying it

• Bucher (2017), the algorithmic imaginary: users' imaginings of algorithms — accurate or not — genuinely regulate their behaviour. This supplies the theoretical frame and empirical base for "adaptation organises around the imagined, not the actual".
• Workaround records in smart-home field studies: taping over sensors, "feeding" devices fixed phrasings, racing the automation — ethnographic studies repeatedly document these and read them as dual strategies: reclaiming control and absorbing defects.
• Method: longitudinal before/after comparison of routines (ethnography or diary studies); cue-based recall interviews — ask users to explain "why do you use it this way", exposing folk theories; log-side workaround indicators (repeat-command rate, retry-after-failure rate, fixed-phrasing share).

Where it stops holding

• Not all adaptation is bad — the test is what is being adapted to. Any good tool reshapes its use; learning to program a thermostat is skill growth. Repeating every command three times because recognition is flaky is absorbing a defect. The criterion: is the adaptation to a legitimate product boundary (a microphone has a range; stepping closer is reasonable) or to a defect (same distance, sometimes works, sometimes not)?
• Adaptation rates vary sharply across populations. Early adopters treat adaptation as fun and bragging rights ("I can make it understand me"); mainstream users as burden; the excluded give up entirely — evaluation must report adaptation by segment, or the average will hide the third group's total loss.
• "Users will learn" is not a licence for poor quality. It is the most common substitution in product arguments: the existence of user adaptation does not reduce the system's responsibility for recognition and feedback — it merely relocates the cost from the system onto the human.

Applying it

• Collect workarounds as demand signals: repeat-command rate, retry rate, fixed-phrasing clusters, manual races in specific windows — build these diagnostic metrics into the product and feed them to iteration.
• Lower the adaptation cost: make the system's input requirements explicit (pickup-range diagrams, supported phrasing examples, sensor-coverage maps) — turning dark probing into open learning, folk theories into official documentation.
• How to check: track longitudinally the count of extra actions to complete the same task — repetitions, distance walked, phrasing variants tried. A decline across product versions means the system is improving; stability means users are still paying the tax.

Related

• Same group: Z3.07.2 Automation keeps adjusting its strategy from observed behaviour · Z3.07.3 Mutual adaptation can entrench a wrong initial assumption · Z3.07.4 Long co-evolution needs periodic review, not unchecked drift
• Nearby: Z3.05.4 Behavioural drift in learning automation keeps eroding predictability · Z1.04 Interaction without interfaces
• Search terms: user adaptation · workaround · algorithmic imaginary · folk theories