Resources

The rapid advancement of images, videos and audios created with generative artificial intelligence makes it necessary to train our gaze and make it critical to detect its weak points (which they have) and learn about tools that can help us with this content

Generative artificial intelligence is advancing at breakneck speed and is now capable of creating images, videos, or audios so realistic that they can easily pass for real content. This technology presents advantageous aspects such as creation speed, the generative capability of algorithms, and cost reduction… but aside from the important debates about the dangers to creativity or the destruction of human labor, its realistic capacity makes it a very useful tool for generating misinformation.

False images about the DANA in Valencia, the war in Gaza, the US elections candidates, child pornography with AI… There are thousands of examples. Even in the world of art, ‘fakes’ have slipped in. In 2023, the jury of the prestigious photographic prize World Press Photo awarded the first prize in one of its categories to an image generated by artificial intelligence. Not even the most expert eye can avoid being fooled.

This is why it is necessary not only to raise awareness about the existence of these images and the confusion they can generate, but also to train the eye and make it critical to detect their weak points (which they have) and to know some tools that can help us detect them. How do we do it?

First of all, it’s important to know that the artificiality of an image is not always hidden. Often, the authors themselves indicate that it is artificial intelligence. To identify it, you need to locate keywords like “AI,” “DALL-E,” “Midjourney,” or other names of generative artificial intelligence tools, which can be difficult for those not familiar with this technology. On the other hand, some image generators add watermarks to their creations. These are usually subtle and placed in one of the corners of the image, so you have to be vigilant.

The critical and reflective eye and thought are our main tools for defending against misinformation in general. Therefore, before accepting any content as valid, it is necessary to ask questions. Is what we see or hear credible? Is it plausible that two antagonistic politicians like Pablo Iglesias and Santiago Abascal would be seen holding hands smiling as they walk down the street? Is it reasonable that the Pope would appear dressed in a flashy Balenciaga parka? Do you think it’s possible that Donald Trump was arrested by dozens of police officers on the streets of New York, or that Kamala Harris dressed in a communist uniform? Common sense is our main ally for detecting the first red flags.

But just like there are methods to verify written lies or falsified images, the content generated by artificial intelligence or deepfakes (content created by AI from real content) requires attention to some specific details because, for now, we can always find some ‘defect’ that reveals its artificiality.

• Imperfections in faces: Errors in ears, misplaced eyes, missing teeth, strange reflections in glasses, or excessively symmetrical faces can be indicators of AI generation.
• Errors in hands: Creating realistic hands is still a challenge for current AIs. Check for the number of fingers or the shape of the hands, as they often display extra fingers, deformities, or strange proportions.
• Lip-syncing and audio: For videos, check if the movement of the lips naturally matches the sound they are producing.

Artificial intelligence can also develop technologies capable of detecting fake images. These tools, which are still in testing and therefore not infallible, analyze the likelihood that a content is synthetic. In some cases, it may be necessary to pay a subscription to access advanced features.

The most well-known ones are Hive Moderation (which offers a percentage indicating the probability that an image was created by AI after analyzing it) and AI or Not (using advanced algorithms and learning techniques to analyze whether images were created by various AI models). Other tools include FotoForensics, Illuminatry, Fake Image Detector, and Winston AI Image Detector.

For the identification of fake audio, there are also specific resources like Loccus.ai, a detector created by a Barcelona-based company that has been purchased by the U.S. technology company Hiya, a leader in global voice security services.

In addition to these specific tools, there are other general resources that can help us verify AI-generated content. For example, we could use the magnifying glass feature of the InVID WeVerify Vera.ai image and video verification tool to zoom in on certain parts of an image without pixelation to detect errors. Even reverse image searches on TinEye or Google Images can help discover if other media and verification organizations have identified this content as AI-generated.

One of the latest advances in the identification of AI-generated images could come from Google. In October, it was revealed that the internet giant is preparing a new version of Google Photos that will integrate a feature capable of showing the information of an AI-generated photo. The new system, the launch date of which is unknown, will not only catalog images generated by its own AI, where it will be indicated as “made with Google AI,” but it will also apparently be able to detect and catalog real images altered with Google Photos AI and even those created by other platforms. This information will be displayed in the image’s metadata under the title “AI Info.”