If the training data set is truly open, the raw inputs (or URLs to the sources) should be available. Isn't a direct image lookup on the source data a better way

If the training data set is truly open, the raw inputs (or URLs to the sources) should be available. Isn't a direct image lookup on the source data a better way to do attribution? (For example: using methods similar to a Google Image Search) At least from a legal perspective, protection (i.e., indemnification) should be offered to those who are clearly attributing their sources (which means they shouldn't be violating copyrights in the first place) and if they don't use clear attribution of sources, they should hold the burden of proof to show that they are not violating copyrights.