What Is a False Positive and Why It Matters

The false positive problem

A false positive occurs when an AI system incorrectly identifies normal behaviour as suspicious. In theft detection, this means a staff member receives an alert for something that isn't actually concealment — a customer reaching for their phone, adjusting their bag, or simply trying on a product.

False positives matter because they erode trust. If staff receive too many incorrect alerts, they stop responding. The system becomes background noise, and real events get missed. This is called alert fatigue — and it's the reason many AI security systems fail in practice despite working in demos.

Why false positives happen

AI detection isn't a binary yes/no decision. Models assign confidence scores to observed behaviour. A person reaching into their pocket might be:

• →Checking their phone (innocent — confidence: 20%)
• →Retrieving a shopping list (innocent — confidence: 25%)
• →Concealing a product (suspicious — confidence: 85%)

The challenge is where to set the threshold. Too low, and you get flooded with false alerts. Too high, and you miss real events.

The calibration approach

Rather than a one-size-fits-all threshold, effective systems calibrate per store. Factors that influence detection include:

• →Camera angles — overhead vs eye-level capture different perspectives
• →Store layout — narrow aisles create more occlusion
• →Lighting — artificial vs natural light affects visual clarity
• →Product placement — items near the body vs on high shelves
• →Customer behaviour norms — some products naturally involve more handling

During calibration, the AI learns what "normal" looks like in your specific environment. This dramatically reduces false positives compared to a generic model.

Multi-model verification

Modern systems don't rely on a single detection signal. IntelliGuard uses a multi-model cascade:

1. 1.Pose estimation — understanding body positioning
2. 2.Action recognition — classifying the type of movement
3. 3.Object detection — tracking items in the frame
4. 4.Temporal analysis — looking at behaviour over time, not single frames
5. 5.Context scoring — considering location, time, and prior behaviour

An alert only fires when multiple models agree at high confidence. This layered approach suppresses the vast majority of false positives while maintaining detection sensitivity.

The metrics that matter

When evaluating AI detection systems, ask for:

• →Precision — what percentage of alerts are real? (Higher = fewer false positives)
• →Recall — what percentage of real events are caught? (Higher = fewer misses)
• →False positive rate trajectory — does it improve over time with calibration?

A system with 90% precision means 1 in 10 alerts is false. At 95%, it's 1 in 20. The difference in staff trust is enormous.

Continuous improvement

The best systems improve over time. As more data is collected from each store, models can be refined to understand that environment better. Managed services provide regular model updates that improve accuracy without requiring any action from the pharmacy team.

This is why "AI as a service" models outperform one-time installations. The AI gets better every month — without the pharmacy needing to do anything.

IntelliGuard uses a multi-model cascade calibrated per store to minimise false positives. Learn more about how it works.