This study examines whether training artificial intelligence systems on large-scale datasets constitutes copyright infringement and how legal outcomes differ across the United Kingdom, European Union, and United States. Using a comparative doctrinal methodology, it analyses statutes, case law, and regulatory instruments alongside the technical stages of scraping, tokenization, and parameterization to identify where acts of reproduction arise. The findings show that AI training inherently involves copying, but the legality of that copying varies: the UK maintains the strictest regime with narrow exceptions, the EU permits training through structured TDM rules with opt-outs, and the US provides the broadest protection under fair use. This fragmented landscape creates significant uncertainty and compliance burdens for developers while offering limited clarity for creators seeking compensation or control. The study concludes that harmonized reforms, improved transparency, and clearer statutory definitions are essential to balance innovation with the rights and economic interests of creators.