Can AI spot fake content? Test results of over 12 AI models published
As AI becomes able to generate images and videos that are comparable to reality, it is becoming more difficult than ever to distinguish whether the content is genuine. There are AI services that can distinguish such content, but how accurate are they? The New York Times has examined 12 types of AI services.
Can AI Detection Tools Really Spot Fake Images and Videos? - The New York Times
The New York Times examined the following 12 AI services:
・AI or Not
・BrandWell
・ChatGPT
・Claude
・Copyleaks
・Gemini Pro
Hive Detect
・Reality Defender
・Resemble.ai
Sensitivity
・Sightengine
・SuperGrok
The New York Times used AI services like ChatGPT and Grok to generate images and videos, then had them check whether they were AI-made. In the example below, all AI services except BrandWell, Claude, and Resemble.ai correctly identified the images as AI-made.
The image below was generated by ChatGPT, but ChatGPT itself was unable to determine whether it was AI-generated.
The image below was generated by Gemini. These images were considered 'complex,' and only four of them—AI or Not, Gemini Pro, Hive Detect, and Resemble.ai—were able to determine they were AI-generated.
There are pros and cons when it comes to videos. Many AI services don't even support video detection. Below are some videos that many AI services have mistaken for real.
In addition, the AI is excellent at detecting 'real' images, and all AIs unanimously judged the following images, which were taken of actual plants with a camera, to be real.
The New York Times stated, 'Many see these AI detection tools as 'powerful judges of truth,' deployed by banks and insurance companies trying to spot AI-enabled fraud, teachers looking for plagiarism, and online investigators verifying images and videos circulating on social media. Overall, however, the conclusions of these tools should be corroborated by other research, such as official photos and news reports.'
Related Posts:
in AI, Posted by log1p_kr