BERLIN: The capabilities of AI chatbots could soon be verging on transgression, according to recent research showing some managing to figure out a person's location from obscure photographs and others now able to generate almost undetectable "deepfake" videos.
In a series of tests, computer virus software provider Malwarebytes found ChatGPT to be "creepy good" at "geo-guessing" locations from photographs that had been scrubbed of their metadata, which often contain details such as location, time and date.
"There are some good reasons to worry about what is visible in background photos, because AI is getting very good at guessing where you are based on the smallest of clues," according to Malwarebytes.
The chatbot was albe to use clues and cues in architecture and environment to narrow down possible locations in photos before either nailing it or coming uncannily close.
"A wheelbarrow of a specific brand or a bird with a limited habitat are enough to provide hints about your location," Malwarebytes warned.
The same week, a tram from Fraunhofer Heinrich-Hertz-Institute and Humboldt University in Berlin published the results of tests that show AI capable of generating "subtle heart-beat-related signals" in so-called deepfake videos.
"The current evolution of image generation techniques makes the detection of manipulated content through visual inspection increasingly difficult," the team said, but some real-life subtleties such as heartbeats were "lost during the deepfake generation process," a limitation that has been "useful for deepfake detection."
Not anymore, however: The team was able to make deepfakes containing what appear to be human pulses. The researchers' findings, which were published in the science and technology journal Frontiers in Imaging, suggest that heartbeat detection techniques are "no longer valid for current deepfake methods."
That said, all hope may not be lost – if clutching at high-tech straws counts: "Analysing spatial distribution of bloodflow regarding its plausibility can still help to detect high-quality deepfakes," the researchers said, in what is likely to be small consolation and even less use to anyone hoping to quickly separate fake from real footage seen on social media. – dpa/Tribune News Service
