An Inspection Exception Log Can Teach Judgment Under AI

A machine vision station rejects a machined metal housing over a bright streak that might be a scratch. The new inspector reads the reject signal. The experienced colleague asks a better question: is the streak on the part, or is it an illumination artifact? Same classification, two entirely different responses. That gap, argues behavioral scientist Gleb Tsipursky, is exactly what an inspection exception log can close as AI-assisted inspection spreads across factory floors.

Tsipursky, CEO of Disaster Avoidance Experts and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026), proposes turning disputed classifications into structured training records. The goal is operators who evaluate evidence, not operators who obey verdicts.

Why the definitions matter first

A false reject is acceptable material failing inspection. A defect escape is nonconforming material passing. Between them sits a third category the industry rarely names: inconclusive cases, where the image data cannot support any disposition. In those situations, neither the system output nor the operator's opinion counts as ground truth. Final disposition stays tied to documented quality criteria and independent review. That principle anchors everything else.

Many inspection problems originate upstream of classification, in image acquisition. Lighting, gain, focus, temperature and calibration all shape what the model sees, as a Vision Systems Design analysis of image-quality tuning noted. National Instruments' machine vision lighting guidance adds illumination geometry, part presentation and surface reflectivity to the list. The best algorithm cannot detect a defect the image never captured. That's physics, not software.

This doesn't license operators to tweak settings. Their job is to flag questionable results, preserve evidence and trigger approved review procedures. Investigation and system adjustments belong to engineering and quality staff.

What goes in the log?

One logging process covers suspected false rejects, potential escapes and inconclusive inspections alike. Each record should include:

• Part and lot identification
• Station and camera information
• Timestamp
• Inspection recipe or software version
• Original pass/fail result
• Original image and the relevant region of interest
• Confidence score or classification metric, if reported

Where a score is logged, operators should document its meaning and threshold, not assume it represents a probability of correctness. Reasonable caution. Plenty of scores don't.

The most valuable field may be the operator's initial assessment. Before learning the verified disposition, trainees record a brief observation and a proposed next action — for example, "bright reflection extends across multiple edges; surface damage not confirmed." Options include holding the part, comparing against a reference image, requesting a repeat capture or escalating. The final review documents the disposition, the supporting evidence and the responsible reviewer. Keeping the original assessment on file shows how judgment evolves as evidence arrives.

Examples beat warnings

Consider three similar housings producing a bright streak near the same polished corner. A trainee calls all three scratches. Review shows two are reflections from part orientation; the third is a physical defect confirmed in secondary inspection. The lesson isn't who was right. It's separating an observation from a defect claim, and knowing what evidence supports which.

A weekly review session led by quality and vision engineering makes this repeatable. Present a small set of anonymized cases — false rejects, confirmed defects, unresolved inspections — showing the original image and system output before revealing the outcome. Trainees state a preliminary assessment, identify missing evidence and choose the next step. Then reveal the disposition and discuss disagreements. The objective is evidence-based reasoning, not consensus.

Research on expert decision-making and AI delegation supports the structure: experienced professionals tend to retain authority over complex judgments while delegating routine tasks to AI. Those studies examine knowledge work, not factory inspection. The parallel still holds.

The log as an engineering feedback loop

Repeated disputes in the same field of view may indicate lighting, positioning or imaging problems. Basler's guidance on machine vision illumination covers glare, shadows and surface reflectivity effects on feature visibility. Other patterns in the log can expose unclear acceptance criteria, thin training data or weak validation procedures. The log stops being a training artifact and becomes a diagnostic instrument for the inspection system itself.

How to measure the pilot

Start small: one station, one defect category. Sensible metrics include:

• Correct identification of inconclusive cases
• Appropriate use of escalation procedures
• Accuracy on independently adjudicated sample sets
• Time required to resolve exceptions

Track false rejects and defect escapes separately, because aggregate accuracy can conceal meaningful shifts between them. An analysis of AI vision performance metrics between Vision Systems Design and Cognex CEO Matt Moschner keeps throughput, yield, user interaction and ongoing KPI review in the picture; training outcomes complement those numbers, they don't replace them.

The clearest sign of progress isn't a metric. It's an operator who recognizes insufficient evidence and knows what to do next. The exception log builds that judgment — and hands engineering a structured record of where the system, not the human, needs work.