Deepfakes created by AI are getting so good that many people, even IT literate people, cannot tell the difference. Tidio, a company behind customer service automation, did a survey, and the results were surprising.
Among the more surprising results was that people who were confident they were right were less likely to spot the difference between an AI-generated face and a real one, and those who were less confident scored higher. Overall, deepfakes or AI-generated faces were only correctly identified by a third of the respondents.
Many of the people who said they would prefer to talk to a human when it comes to customer service failed to spot an AI-generated chatbot.
While this is interesting, it is the daily use of deepfakes to deceive that is the worrying trend. We read of public figures having their reputations attacked by their critics using deepfakes, and it can be very nasty.
It is also a nice excuse to label news we do not like as deepfake.
MIT has discovered another flaw in AI-generated images. A significant proportion of images in one of the biggest data sets used in ‘training’ AI are labelled incorrectly. There was an error rate of the sample taken of 5.8%, and for Quickdraw images, the percentage was just over 10%. This may not sound significant, but it means for every 100 images between 5 and 10 are not what they say they are.
Another quirk that works against the success of deepfakes is that most images concentrate on faces. Ask an AI to produce an image of a cat’s face, and it will be pretty believable. Ask it to produce an image of the whole cat and, because of the enormous number of positions legs and feet can be in, the result is something from your worst nightmare.
Deepfakes are being used as a way of subverting reality. Amazon workers are trying to follow Google into creating a labour union, and it turns out certain people created deepfake Twitter accounts to apply extra pressure for the movement.
The good news is that there are plenty of people out there whose work identifies deepfakes and exposes them as such. The bad news is that deepfakes will only get better – as in more convincing. Then we will have a situation where the only thing that can spot a fake will another AI machine.
We already have a survey that shows that AI-generated deepfakes can fool us. Let us hope that we can regain control of the AI train before it goes off the rails entirely.