We study portfolio diversification in an experimental decision task, where asset returns depend on a draw from an ambiguous urn. Holding other information identical and controlling for the level of ambiguity, we find that labeling assets as being familiar or from the homeland of subjects increases portfolio weights by around 25%, respectively; although the return-generating process remains unaffected. Importantly, we only find these effects when the returns of assets are highly ambiguous. Our ambiguity robust mean-variance model accurately predicts benchmark portfolio weights of the experimental control group, where assets are not labeled: subjects allocate more wealth to assets with low ambiguity. For treatment group portfolios, which show a bias towards assets with a familiar or homeland label, the model does not hold. This misdiversification against the benchmark portfolio can be rationalized via the concept of source dependence of uncertainty attitudes.