A couple of weeks ago, the Greenwich Skeptics in the Pub played host to Sophie Nightingale, who studies the psychology of AI deepfakes. The particular project she spoke about was an experiment in whether people can be trained to be better at distinguishing them from real images.
In Nightingale’s experiments, she carefully matched groups of real images to synthetic ones, first created by generative adversarial networks (GANs), later by diffusion models (GeeksforGeeks raters’ demographics.
Then the humans were given some training in what to look for to detect fakes and the experiment was rerun with new sets of faces. The bad news: the training made a little difference, but not much. She went on to do similar experiments with diffusion images.
Nightingale has gone on to do some cross-modal experiments, including audio as well as images, following the 2024 election incident in which New Hampshire voters received robocalls from a faked Joe Biden intended to discourage voters in the January 2024 primary. In the audio experiment, she played the test subjects very short snippets. Played for us in the pub, it was very hard to tell real from fake, and her experimental subjects did no better. I would expect longer clips to be more identifiable as fake. The Biden call succeeded in part because that type of fake had never been tried before. Now, voters, at least in New Hampshire, will know it’s possible that the call they’re getting is part of a newer type of disinformation campaign aimed at
In another experiment, she asked participants to rate the trustworthiness of the facial images they were shown, and was dismayed when they rated the synthetic faces slightly (7.7%) higher than the real ones. In the resulting paper for Journal of Vision, she hypothesizes that this may be because synthetic faces tend to look more like “average” faces, which tend to be rated higher in trustworthiness, even if they’re not the most attractive.
Overall, she concludes that both still images and voice have “passed the Uncanny Valley“, and video will soon follow. In the past, I’ve chosen optimism about this sort of thing, on the basis that earlier generations have been fooled by technological artifacts that couldn’t fool us now for a second. The Cottingley Fairies looks ridiculous after generations of knowledge of photography. On the other hand, Johannes Vermeer’s Girl with a Pearl Earring looks more real than modern deepfakes, even though the subject is generally described as imaginary. So it’s possible to think of it as a “deepfake”, painted in oils in the 17th century.
Fakes have always been with us. What generative AI has done to change this landscape is to democratize and scale their creation, just as it’s amping up the scale and speed of cyber attacks. It’s no longer necessary to be even barely competent; the tools keep getting easier.
Listening to Nightingale it seems most likely that work like that in progress by an audience member on identifying technological artifacts that identify fakes will prove to be the right way forward. If those differences can be reliably identified, they could be built into technological tools that can spot indicators we can’t perceive directly. If something like that can be embedded into devices – phones, eyeglasses, wristwatches, laptops – and spot and filter out fakes in real time, and we should be able to regain some ability to trust what we see.
There are some obvious problems with this hoped-for future. Some people will continue to seek to exploit fakes; some may prefer them. The most likely outcome will be an arms race like that surrounding email spam and other battles between malware producers and security people. Still, it’s the first approach that seems to offer a practical solution to coping with a vastly diminished ability to know what’s real and what isn’t.
On the Internet your home always leaves you, part 4,563. Twenty-two-year-old blogging site Typepad will disappear in a few weeks. To those of us who have read blogs ever since they began, this news is shocking, like someone’s decided to tear down an old community church. Yes, the congregation has shrunk and aged, and it’s drafty and built on creaking old technology (in Typepad’s case, Moveable Type), but it’s part of shared local history. Except it isn’t, because, as Wikipedia documents, corporate musical chairs means it’s now owned by private equity. Apparently it’s been closed to new signups since 2020, and its bloggers are now being told to move their sites before everything is deleted in September. It feels like the stars of the open web are winking out, one by one.
On the Internet everything is forever, but everything is also ephemeral. Ironically, the site’s marketing slug still reads: “Typepad is the reliable, flexible blogging platform that puts the publisher in control.”
Illustrations: “Girl with a Pearl Earring”, painted by Johannes Vermeer circa 1665.
Wendy M. Grossman is an award-winning journalist. Her Web site has an extensive archive of her books, articles, and music, and an archive of earlier columns in this series. She is a contributing editor for the Plutopia News Network podcast. Follow on Mastodon or Bluesky.
