How The Daily Beast Is Taking A Page Out of The Gateway Pundit’s Playbook
Journalistic Integrity Takes a Backseat as The Daily Beast Chooses to Skip Fact-Checking A Known Conspiracy Theorist.
Part One
On April 12th, The Daily Beast published an article by Jake Lahut. The article aimed to draw a comparison between Ron DeSantis’ campaign and the playbook of Richard Nixon. While there may be some truth to this comparison, folks in the disinformation community were immediately alarmed by the primary source of the article, Steven Jarvis.
I’ve learned quite a bit about Steven Jarvis recently from his conspiracy theories regarding many subjects and personalities in the disinformation community (including writers at The Daily Beast) to a long history of coordinated harassment against his assumed opponents, and a proclivity for risqué artificially generated art.
The alarm that rippled through the disinformation community wasn’t just focused on Steven Jarvis’s history. Many were outraged by an apparent lack of fact-checking that, if completed, would have disproven the core anecdotes of the article.
I plan to use this space to fact-check a particular anecdote from the article centered around the use of an image.
Jake Lahut writes that Steven Jarvis’s child, whose image was not on social media, had their likeness transformed into a “sexually graphic meme” in July of 2022. In the next paragraph, Jarvis is quoted as saying the image appeared with “basically a porn-style message”. Quite a significant change in language. No additional details or evidence are provided to assert the veracity of the claim. Most importantly, Jarvis says “They altered that image, but it was clearly my child”.
After reviewing the image in question in detail and reverse-engineering the process by which Jarvis claims the image was created, my assertion is that the above statement is highly implausible. I can say with near certainty the alleged image of Jarvis’s child was artificially generated using a service called Generated.Photos. This service can customize thousands of StyleGAN images (hyper-realistic images of people who do not exist) based on a set range of selected options including age, sex, hair length, emotion, head position, eye color, hair color, and ethnicity. Its results are random but one may be inclined to find a reasonable facsimile of an existing face if they were motivated to do so.
Compare the above image sets with the image that Steven Jarvis claims is his child. We see the same front-facing head posture, similar hairstyle, the same non-descript grey background, and several StyleGAN artifacts that allow us to conclude the image is artificially generated including render errors in the teeth and ears. Any assertion that this is an altered image of an actual living child is false.
Further, if Jake Lahut had access to the image and had done a reverse image search, he would have immediately seen a link to Generated.Photos, the service that created the image. It’s also surprising that a disinformation hobbyist wouldn’t have completed a reverse image search on an alleged photo of his child to see if it had been posted elsewhere online.
Part Two
The above provides ample evidence to question the veracity of the anecdote presented at the beginning of Jake Lahut’s article, but I’ve decided to push further into this story by examining some of the claims made in a blog post written by Steven Jarvis in August 2022, shortly after the image was shared to social media.
In the underlined portion, Jarvis asserts that StyleGAN images can be customized to match the characteristics of other real images. To some extent, this is true. StyleGAN image libraries require thousands if not millions of images to return hyper-realistic images of people who do not exist. Before its acquisition, the website ThisPersonDoesNotExist.com used a dataset of 70,000 images. Generated.Photos, the service believed to have been used to create the StyleGAN image that Jarvis claims is his child, uses a dataset of 2.7 million images and allows for some image customization as was previously noted. Designing a StyleGAN image around one image is possible but will likely be inaccurate due to the amount of data necessary.
Further in the article, Jarvis shares an anecdote about a haircut incident with another of his children. He insists that the artificially generated image is an altered image of his child because the hair on the image matches that of his child's hairline from the time of the haircut.
In line with what was previously said, StyleGAN image generators require thousands of images to produce hyper-realistic images. It’s been established that Jarvis’s daughter had no images on social media outside of an image on a family member's Facebook page (according to his blog post). Even a handful of images would not be enough to reproduce someone's likeness, let alone the reproduction of a hairline after a specific event. This is highly implausible.
Jarvis asserts that the creation of his daughter’s likeness was a two-step process. I’ll examine both steps in detail. First, he suggests that his opponent “used an Ap” [sic] to change her sex from female to male. According to the screenshot, FaceApp was used.
In reverse engineering this process, I needed to see what happens when FaceApp modifies the characteristics of an image as it switches between female and male. I used two sets of images, one set of three female celebrities of varying ages and one set of three StyleGAN images that present as young girls.
A byproduct of changing characteristics from female to male within FaceApp is that, regardless of age, the app will make the new image more masculine and appear slightly older. The ears are more exposed, the jawline is strengthened, and the hair is both thickened and darkened. Had Steven Jarvis’s child’s image been run through this image processor, it would have also been aged and experienced similar changes to its ears, jaw, and hair, yet Jarvis claims the specific haircut of the four-year-old was preserved.
When StyleGAN images are transformed via FaceApp we notice less significant aging, though some evidence of aging remains. What is most notable, however, is that rendering abnormalities from the image’s creation are corrected once the image has been transformed. This is particularly important as the artificially generated image that Jarvis claims is an image of his child shows abnormalities in the ears as well as the teeth.
The last claim made by Jarvis is that the image, which had been allegedly modified through FaceApp, was then “run through a GAN program” to arrive at the final result. Reiterating from earlier, StyleGAN programs like the former ThisPersonDoesNotExist and Generated.Photos require libraries of images from the tens of thousands to the millions. It is certainly possible to train a program on a library of specific images, but the individual would require many, many images and a computer with enough processing power to acquire the hyper-realistic likeness of that person. It’s been established that virtually no images of the child exist online.
Steven Jarvis makes many assumptions in his August 2022 blog post with little evidence to verify his claims. In The Daily Beast article, he’s quoted as saying “They altered that image, but it was clearly my child”. The preponderance of evidence proves this to be highly implausible. I do not know what Steven Jarvis’s child looks like, but I am aware of the process and hardware necessary to create StyleGAN images and I’ve shown the effect FaceApp has on gender characteristics.
Unfortunately, The Daily Beast and Jake Lahut have now validated these implausible claims, the result of which will likely empower and inspire other bad actors and ultimately damage the reputation of The Daily Beast.
