I heard a presentation on this topic from a computer forensics expert who handles child porn cases at the HOPE conference last summer. I assume that this article, despite being worded somewhat vaguely, operated in the manner that he described.
Essentially: some governmental authority (in the US, I believe the FBI) maintains a database of specific image and video files that are either confirmed or strongly believed to be child pornography. Investigators are able to obtain a database of the metadata and hashes of those files, but not the images themselves. When an investigator is examining an archive of data that may contain child pornography (for example a seized hard drive) they run scripts that hash each file using the same algorithm and compare them with the database of hashes of known child pornography files. The investigator does NOT open the files in any viewer program or examine them directly. If any hashes match, the flagged files are then forwarded to the proper authority (such as the FBI), where the person with the most depressing job ever then opens the files and examines them visually to see if the file with a matching hash is, in fact, the same file as the file in the governmental archive of verified child pornography. If child pornography is then positively identified, the remainder of the archive in which it was located may be forwarded to that office to be visually scanned by child pornography expert in order to determine whether or not there are more images or videos that appear to be child pornography, which have not been previously identified and added to the archive.
Note that the initial forensic investigator will also do other work, such as examining file metadata, system usage logs, browsing history, searching for various worms and other malware, etc. to try and determine if the owner of that archive was, in fact, viewing the images deliberately, or whether it may have been an accidental or incidental download.
One key idea in this system is that the initial forensic examiner will under no circumstance ever VIEW the file, or - in fact - ever see any child pornography whatsoever.
In the scenario described in this article, it sounds to me like Google has implemented some automatic scanning tool that takes the place of this initial investigator, but assuming it is based on the procedure I heard described, there is absolutely zero chance that they are scanning for similar images, since a) the algorithm is only comparing hashes, not actual file data and b) Google would not have the actual images that are being compared against, so they would have no valid data for their image search algorithms to work on.
In fact, it would almost certainly be highly illegal for Google to have such data to work with, although I presume that it would be both legally and technically possible for Google to arrange a system where images were sent to the FBI and scanned on their servers, but that would clearly be an overwhelming amount of data, and an extremely wasteful use of resources for bulk scanning, rather than targeting some particular account, as described above in the procedure for human investigators.
But it does not seem terribly inefficient for Google to have a copy of the hash database, and for them to scan that against hashes of the image files, and then to automatically forward positive hits to the FBI.