Abstract The history of attempts at automatic recognition of images derived from microscopic and other biological material has now about 15 years to it; but there are as yet few working systems in which recognition tasks of any degree of complexity are done by computers or allied hardware. It is, nevertheless, now possible to obtain commercially means for carrying out certain screening and counting assignments, including devices for the classification of curvilinear patterns such as e.c.gs (though the acceptability of the performance offered often depends on the astonishing inconsistency displayed by humans carrying out the same work!). Generally speaking, in the case of the classic two-dimensional optical pattern, methods of field localization, data capture, scene segmentation, determination of geometry and eventual classification are still largely experimental, and the optimal balance of special hardware, conventional computing and operator action for a particular problem has seldom been seriously investigated. Except in cases involving only the very simplest types of pattern discrimination, and those in which statistically correct results can be obtained notwithstanding gross errors in the classification of individual pattern components, the viability of systems often depends crucially on the ergonomics of the division of function between operator and machine. Indeed, one of the most promising developments is the advent of arrangements in which a technician takes most of the pattern recognition decisions but a machine carries out measurements — which can be of arbitrary complexity — on objects and parts of objects selected by the operator; and of course attends to all the housekeeping and statistical manipulation of the data. These remarks are illustrated by reference to current commercial systems and research projects in radiology, cytology, bacteriology and cytogenetics.