We present a new method for image taxonomy. Our formulation shows how the combination of automated metadata-based analysis and knowledge-acquisition tools can help build better image search applications. We build on top of existing metadata-based techniques to image taxonomy and combine them to human assistance, in order to guarantee some minimal level of semantic soundness. Knowledge from humans is acquired thru their interaction with CAPTCHAs, which application is shown here in an innovative way. We believe this new application of CAPTCHA is a promising opportunity for knowledge acquisition because it takes advantage of a very special moment: since humans are willing to take a CAPTCHA test to access some resource, it might be possible to take advantage of this moment to extract meaning from human’s response to the CAPTCHA test, in a collaborative fashion.