Systems Get Worked Around

If a camera system affects what happens to people, people respond to it. This is not misconduct; it is the predictable result of a measurement with a consequence.

What adaptation looks like

Standing outside the field of view to do the thing being detected.

Learning the threshold and operating just inside it — moving slowly enough not to trigger, staying just outside the zone.

Obstructing the view, by placing objects or by parking a vehicle.

Timing activity to the periods when the system is known to be unreliable.

Compliance theatre: performing the detected behaviour for the camera while the actual work happens elsewhere.

Each of these tells you something the system was not designed to hear.

Why it is usually a design finding

People work around controls that make the work harder.

A detector that flags a shortcut which exists because the sanctioned route is impractical is measuring a process problem and attributing it to a person.

Investigate why before treating it as a violation, because the answer usually names something fixable that would remove the incentive entirely.

Where the workaround is unsafe, it is more urgent, not less, and the same logic applies: the safe method was slower or harder, and that is the thing to fix.

Deliberate defeat

Distinct from adaptation and worth naming separately.

Cameras covered, turned, or unplugged.

Lens obscured with tape, paint or grease.

Lighting disabled to blind night operation.

These are visible in the image statistics — a sudden change in brightness, contrast or scene structure — and should be alerted on.

Investigate the reason rather than only the act. A camera defeated repeatedly at one location is telling you something about that location.

Adversarial inputs

Research has shown that image classifiers can be misled by deliberately crafted patterns, including printed patterns worn by a person.

In practice this is rare in commercial deployment and much more common as a research finding than as an operational problem.

Which does not make it irrelevant: if a system is the sole control on something valuable, its failure modes matter, and "an unusual pattern makes it not see you" is a failure mode worth knowing about.

The honest position: treat it as a reason not to make a vision system the only control, rather than as an imminent threat.

The measurement consequence

A system being worked around produces data that looks fine.

Detections fall. The chart improves. Nothing has improved.

Which is why detection counts alone are a poor measure of anything, and why the outcome measure — incidents, defects, actual events — has to be tracked alongside.

A falling detection rate with a flat incident rate is a gaming signal, and it is available from data you already have.

What reduces it

Explain the purpose, which is the cheapest intervention and the most often skipped.

Investigate conditions before people, consistently and visibly.

Fix the reason the workaround exists.

Do not attach automatic consequences to detections, which converts a monitoring system into an adversary and guarantees the adaptation described above.

The two-signal check

The measurement that reveals avoidance, and it uses data you already have.

Track detections and the real-world outcome together.

Detections falling while incidents, defects or the underlying event rate stay flat is the signature.

Investigate the location, which is usually where the avoidance is concentrated.

Ask the people who work there, without consequence attached, because the answer names the reason.

Fix the reason. Enforcement produces better concealment; removing the incentive produces the behaviour you wanted.

Detecting deliberate defeat

Cameras covered, turned or blinded show up in the image statistics before anyone reports them.

Sudden change in brightness: covered or blinded.

Sudden change in scene structure: turned or moved.

Frame loss: unplugged.

Alert on these, separately from detections, because they mean the system has stopped working rather than that nothing is happening.

Investigate the reason, not only the act. A camera defeated repeatedly at one location is saying something about that location, and the answer is usually about the work rather than the person.