Donald Trump may have coined a new term in his latest false attack on Kamala Harris’ presidential campaign. In a pair of posts on Truth Social over the weekend, the former president said that Vice President Kamala Harris “A.I.’d” photos of a huge crowd that showed up to see her speak at a Detroit airport campaign rally last week.
“There was nobody at the plane, and she ‘A.I.’d’ it, and showed a massive ‘crowd’ of so-called followers, BUT THEY DIDN’T EXIST!” Trump wrote. “She’s a CHEATER. She had NOBODY waiting, and the ‘crowd’ looked like 10,000 people! Same thing is happening with her fake ‘crowds’ at her speeches.”
The Harris campaign responded with its own post saying that the image is “an actual photo of a 15,000-person crowd for Harris-Walz in Michigan.”
Aside from the novel use of “AI” as a verb, Trump’s post marks the first time, that we know of, that a US presidential candidate has personally raised the specter of AI-generated fakery by an opponent (rather than by political consultants or random social media users). The accusations, false as they are, prey on widespread fears and misunderstandings over the trustworthiness of online information in the AI age.
It would be nice to think that we could just say Trump’s claims here are categorically false and leave it at that. But as artificial intelligence tools become increasingly good at generating photorealistic images, it’s worth outlining the many specific ways we can tell that Harris’ crowd photos are indeed authentic. Consider this a guide for potential techniques you can use the next time you come across accusations that some online image has been “A.I.’d” to fool you.
Context and Sourcing
By far the easiest way to tell Harris’ crowds are real is from the vast number of corroborating sources showing those same crowds. Both the AP and Getty have numerous shots of the rally crowd from multiple angles, as do journalists and attendees who were at the event. Local news sources posted video of the crowds at the event, as did multiple attendees on the ground. Reporters from multiple outlets reported directly on the crowds in their accounts: Local outlet MLive estimated the crowd size at 15,000, for instance, while The New York Times noted that the event was “witnessed by thousands of people and news outlets, including The New York Times, and the number of attendees claimed by her campaign is in line with what was visible on the ground.”
Suffice it to say that this mountain of evidence from direct sources weighs more heavily than marked-up images from conservative commentators like Chuck Callesto and Dinesh D’Souza, both of whom have been caught spreading election disinformation in the past.
When it comes to accusations of AI fakery, the more disparate sources of information you have, the better. While a single source can easily generate a plausible-looking image of an event, multiple independent sources showing the same event from multiple angles are much less likely to be in on the same hoax. Photos that line up with video evidence are even better, especially since creating convincing long-form videos of humans or complex scenes remains a challenge for many AI tools.
It’s also important to track down the original source of whatever alleged AI image you’re looking at. It’s incredibly easy for a social media user to create an AI-generated image, claim it came from a news report or live footage of an event, then use obvious flaws in that fake image as “evidence” that the event itself was faked. Links to original imagery from an original source’s own website or verified account are much more reliable than screengrabs that could have originated anywhere (and/or been modified by anyone).
Telltale Signs
While tracking down original and/or corroborating sources is useful for a major news event like a presidential rally, confirming the authenticity of single-sourced images and videos can be trickier. Tools like the Winston AI Image Detector or IsItAI.com claim to use machine-learning models to figure out whether or not an image is AI. But while detection techniques continue to evolve, these kinds of tools are generally based on unproven theories that haven’t been shown to be reliable in any broad studies, making the prospect of false positives/negatives a real risk.
Writing on LinkedIn, UC Berkeley professor Hany Farid cited two GetReal Labs models as showing “no evidence of AI generation” in the Harris rally photos posted by Trump. Farid went on to cite specific portions of the image that point to its authenticity.
“The text on the signs and plane show none of the usual signs of generative AI,” Farid writes. “While the lack of evidence of manipulation is not evidence the image is real. We find no evidence that this image is AI-generated or digitally altered.”
And even when portions of a photo appear to be nonsensical signs of AI manipulation (à la misshapen hands in some AI image models), consider that there may be a simple explanation for some seeming optical illusions. The BBC notes that the lack of a crowd reflection on the plane in some Harris rally photos could be caused by a large, empty area of tarmac between the plane and the crowd, as shown in reverse angles of the scene. Simply circling odd-looking things in a photo with a red marker is not necessarily strong evidence of AI manipulation in and of itself.
Trust No One?
At this point, AI watchers are used to the prospect of AI “deepfakes” and their potential to mislead. But Trump’s latest false claims about getting “A.I.’d” reflect the other side of the deepfake coin, where people can start to doubt authentic information just because of the potential for AI to generate a similar fake. We’ve recently seen this same “deep doubt” phenomenon manifest in conspiracy theorists who insist President Joe Biden has been replaced by an AI-powered hologram.
By encouraging his followers to doubt images that are clearly authentic (and that can be trivially confirmed as such), Trump appears to be trying to create a media environment where AI’s mere existence means even multiple reliable sources of corroborating information can’t necessarily be trusted. For a candidate who continues to falsely claim that the 2020 election was stolen from him, that could set a dangerous precedent for November’s upcoming election.
This story originally appeared on Ars Technica.
