When talking about the dangers of the AI used to create Deepfakesits effectiveness is not usually mentioned, to what extent they manage to go through real. Microsoftwhich defines a Deepfake as an image, video or audio recording generated by AI and used to disseminate false information, launched a year ago a year ago tool that allows to test the visitor’s ability to distinguish the images created by artificial intelligence from which they are real. What we did not know is that this website and its results were going to be used by the company to perform A study on this issue whose results has recently published.
The most 12,000 users who visited the web Real or not Throughout the month of August 2024 they made 17,340 Tests About a total set of 700 images ia and 350 reais. The first were created by tools such as Dall-e 3, Midjourney V6, Amazon Titan V1, Stable Diffusion XL and Stable Diffusion 3.
It should be noted that, over time, all these tools have had new versions that have improved their capabilities, So the results, of doing the study now, would be even more discouraging. With the state of AI of a year ago, the Average hobby was 62 % and the errors of 38 %. As for real images, users successfully identified them a 59 % of the times. Those of AI, a 63 %.
The study indicates that the images that most confuse users are those that ‘They seem realistic, but not professionals’. Also that ‘IA images that show landscapes and objects are more difficult to detect for users, probably due to people inherent to people to identify human faces’. In any case, the chances of a person identifying an image as such are similar to ‘throw a currency in the air’.
Microsoft researchers compared these results with Image detectors AIthese being much better than humans. His Average hobby was 95.4 %maintaining 95 % with both real and artificial images.
Put on Real or Not
If you want to check your ability to distinguish images generated by the real ones, You can still do the tests presented by Real or Notthat is still online.
The operation of the tool is very simple. As the web is accessed, a brief explanation indicates that ‘The growing quality of the images makes them more difficult to detect. Can you say if this image is real or generated by AI? ‘
After clicking on Start Game, Real or not presents the visitor a succession of 15 imagesand in each of them you have to say if it is real or artificial by clicking on the corresponding button. After each choice, the web indicates whether or not you have been right and At the end of the series it facilitates the total percentage of hits and compares it with the average visitors. If the user wants to make another attempt, he will find a new image gallery different from the previous one.