Bad pixels are generally treated as a loss of useable area and then excluded from averaged performance metrics. The definition and detection of “bad pixels” or “cosmetic defects” are seldom discussed, perhaps because they are considered self-evident or of minor consequence for any scientific grade detector, however the ramifications can be more serious than generally appreciated. While the definition of pixel performance is generally understood, the classification of pixels as useable is highly application-specific, as are the consequences of ignoring or interpolating over such pixels. CMOS sensors (including NIR detectors) exhibit less compact distributions of pixel properties than CCDs. The extended tails in these distributions result in a steeper increase in bad pixel counts as performance thresholds are tightened which comes as a surprise to many users. To illustrate how some applications are much more sensitive to bad pixels than others, we present a bad pixel mapping exercise for the Teledyne H2RG used as the NIR tip-tilt sensor in the Keck-1 Adaptive Optics system. We use this example to illustrate the wide range of metrics by which a pixel might be judged inadequate. These include pixel bump bond connectivity, vignetting, addressing faults in the mux, severe sensitivity deficiency of some pixels, non linearity, poor signal linearity, low full well, poor mean-variance linearity, excessive noise and high dark current. Some pixels appear bad by multiple metrics. We also discuss the importance of distinguishing true performance outliers from measurement errors. We note how the complexity of these issues has ramifications for sensor procurement and acceptance testing strategies.