Several studies have found that collaboration networks are scale-free, proposing that such networks can be modeled by specific network evolution mechanisms like preferential attachment. This study argues that collaboration networks can look more or less scale-free depending on the methods for resolving author name ambiguity in bibliographic data. Analyzing networks constructed from multiple datasets containing 3.4M ~ 9.6M publication records, this study shows that collaboration networks in which author names are disambiguated by the commonly used heuristic, i.e., forename-initial-based name matching, tend to produce degree distributions better fitted to power-law slopes with the typical scaling parameter (2 < α < 3) than networks disambiguated by more accurate algorithm-based methods. Such tendency is observed across collaboration networks generated under various conditions such as cumulative years, 5- & 1-year sliding windows, and random sampling, and through simulation, found to arise due mainly to artefactual entities created by inaccurate disambiguation. This cautionary study calls for special attention from scholars analyzing network data in which entities such as people, organization, and gene can be merged or split by improper disambiguation.