You can’t hide your lying AIs.
Not only is AI slop taking over the internet, but it’s becoming indistinguishable from the real deal. Scientists have found that people can’t tell the different between human and AI-generated faces without special training, per a dystopian study published in the journal Royal Society Open Science.
“Generative adversarial networks (GANs) can create realistic synthetic faces, which have the potential to be used for nefarious purposes,” wrote the researchers.
Recently, TikTok users blew the whistle on AI-generated deepfake doctors who were scamming social media users with unfounded medical advice.
In fact, these faces from concentrate have become so convincing that people are duped into the thinking the counterfeit countenances are real more than the genuine artifact, Livescience report.
To prevent people from being duped, researchers are attempting to design a five-minute training regimen to help users unmask the AI-mposters, according to lead study author Katie Gray, an associate professor in psychology at the University of Reading in the UK.
These trainings help people catch glitches in AI-generated faces, such as the face having a middle tooth, a bizarre hairline or unnatural-looking skin texture. These false visages are often more proportional than their bonafide counterparts.
The team tested out the technique by running a series of experiments contrasting the performance of a group of typical recognizers and super recognizers — defined as those who excel at facial recognition tasks.
The latter participants, who were sourced from the Greenwich Face and Voice Recognition Laboratory volunteer database, had reportedly ranked in the top 2% of individuals in exams where they had to recall unfamiliar faces.
In the first test, organizers displayed a face onscreen and gave participants ten seconds to determine if it was real or fake. Typical recognizers spotted only 30% of fakes while super recognizers caught just 41% — less than if they’d just randomly guessed.
The second experiment was almost identical, except it involved a new group of guinea pigs who had received the aforementioned five-minute training on how to spot errors in AI-generated faces.
The test takers were shown 10 faces and evaluated on their AI-detection accuracy in real time, culminating in a review of common rendering mistakes.
When they participated in the original experiment, their accuracy had improved with super recognizers IDing 64% of the fugazi faces while their normal counterparts recognized 51%.
Trained participants also took longer to examine the faces before giving their answer.
“I think it was encouraging that our kind of quite short training procedure increased performance in both groups quite a lot,” said Gray.
Of course, there are a few caveats to the study, namely that the participants were put to the test immediately after training, so it was unclear how much they would’ve retained had they waited longer.
Nonetheless, equipping people with the tools to distinguish humans from bots is essential in light of the plethora of AI-mpersonators flooding social media.
And the tech’s chameleonic prowess isn’t just visual.
Recently, researchers claimed that language bot ChatGPT had passed the Turing Test, meaning it is effectively no longer discernible from its flesh-and-blood brethren.
