Lyon also said AP is increasingly handling and sourcing images from new kinds of sources, though it represents a small percentage of the photos it distributes.
Contacted in Paris, Cozien said the high end version of his analysis software takes roughly thirty minutes to process an image, and that this must be done by a trained analyst. (An extremely important image requiring the best possible evaluation can take twice that length, he said.) Roughly eight AFP employees in different parts of the world have been trained to use the software, according to Cozien.
“The software gives lots of results, but an operator has to analyze the results,” he said, comparing it to way a trained technician is required to operate an x-ray machine. “Anyone who buys the software needs to have the training and it [takes] at least one full week.”
A related product from Cozien’s company can batch process images and flag suspect files, which then need further examination. But the batches aren’t yet large enough to accommodate all of the images from a service like AFP, or the tens of thousands per day that flow through AP. At this moment in their evolution, photo analysis programs can only be used for spot checks or special situations. (On the bright side, at least that’s a more common use than what news organizations do when it comes to the widely-available plagiarism detection services.)
One key factor in detecting image manipulation is the human element, which is another reason why we’re unlikely to soon see wide scale use of photo analysis by news organizations.
“It’s beyond just the technical aspects of a photo,” Lyon said when explaining how his organization determines the authenticity of images sourced from citizens or those unfamiliar to AP.
Lyon listed a range of questions asked of the source of a photo when evaluating an image of unknown or undetermined origin: “Who is the photographer? Where were they? Why were they there? Do they have other images they can show us taken before and after the image we’re interested in? And what can we do to get certain questions answered about the provenance of the image that put us more at ease?”
Another reason the human factor is essential - and why asking questions of a source can often trump technical analysis - is the images themselves. Lyon said low resolution images, like those taken by cameras inside basic cell phones, are harder to analyze using software.
“The minute an image is lowered in resolution, that it falls under a certain threshold, it becomes almost impossible to run any kind of program to detect manipulation,” Lyon said. “A big image with a lot of detail can be more closely analyzed than an image of 60 or 70k and so sometimes what happens, and this is a confusing aspect of it, on the Internet you get a lot of low res images floating around and it’s difficult to understand what’s been done to them and where they come from.”
The old cliché is that a picture is worth a thousand words. In today’s world of ubiquitous cameras and tools for photo manipulation, it seems a picture is also often worth a thousand questions.
“In the face of the ease of the scam there is a need on the part of all journalists to never assume anything and to always cross-check and verify in order to remain trusted sources of news and information,” Lyon said.
Correction of the Week
An item in the Extra Bases baseball notebook last Sunday misidentified, in some editions, the origin of the name Orcrist the Goblin Cleaver, which Mets pitcher R. A. Dickey gave one of his bats. Orcrist was not, as Dickey had said, the name of the sword used by Bilbo Baggins in the Misty Mountains in “The Hobbit”; Orcrist was the sword used by the dwarf Thorin Oakenshield in the book. (Bilbo Baggins’s sword was called Sting.) -The New York Times
[Update: the headline of this post originally read “Notes on Photo Fraud.” After some objections, we have changed it. —ed.]