In March 2023, a claim that the former U.S. president, Donald Trump, had been arrested surfaced online. The images shared alongside the claim show Trump resisting an arrest and another one shows him being chased by the police. Due to the virality of the claim and the images attached, one would have thought the news was true. But fact-checkers would later find out that the images were generated using Artificial Intelligence also known as AI.
The ability of Artificial Intelligence to generate unreal images has sparked concerns about its possible misuse to spread disinformation. This may further compound the challenges faced by fact-checkers and derail their efforts in combating the spread of manipulated visuals meant to dissuade people online.
The term AI encompasses all technologies capable of completing tasks that traditionally require human intelligence. Recent advancements in machine learning and related fields have enabled AI to automate a broad range of activities, from image recognition to language processing.
While the technology can be used in an ethical way, it can also be used to generate false images and videos for the purpose of manipulating or deceiving people. This advanced technology works by using machine learning algorithms to study media and imitate it, producing convincing content that seems so real.
For example, AI could be used to create convincing footage of political figures saying or doing things that they never actually did. This could be used to sway public opinion or to discredit political opponents during electioneering.
Though The FactCheckHub had earlier outlined various ways to verify digitally manipulated images, this piece is important due to the constant advancement in AI tools and software development. Unlike other forms of deep fake technology that rely on already existing visuals, advanced Artificial Intelligence is capable of creating fake visuals of people, places and events that do not exist anywhere.
Identifying AI-generated images can be a challenging task due to the advanced technology employed in generating them. While arriving at a definitive conclusion regarding such images may not always be possible, combining the following techniques can aid in recognizing images that are probably AI-generated.
1. Critical Examination
Like digitally manipulated images, when AI-generated images are examined critically, it could have features that points to the fact that it was artificially created. Though there is no hard and fast rule in identifying them, because different AI images exhibit different characteristics, close observation could reveal possible errors or manipulation.
AI-generated images often contain subtle differences from natural images. By analyzing the key features like facial features in the image, it may be possible to detect whether it was generated by AI. For instance, some experts have said AI is not yet perfect enough in designing some key features in humans such as eyes and hands. Paying attention to some of these key features may reveal some dubious details that can indicate manipulation.
For instance, this photo which shows police officers chasing Trump seems to be very real at first glance. However, look closely and you will see that the face of Trump and the cops look blurry. The officers appear to be chasing Trump – but they are looking in a totally different direction.
2. Reverse image search
One of the simplest ways to detect AI-generated images is to perform a reverse image search. This involves using a search engine, such as Google or TinEye, to search for the image online. If the image is a deep fake or otherwise manipulated, it may not appear anywhere else online, or it may appear in multiple places with different contexts. It may also lead you to the actual context behind the image.
For instance, when Twitter appointed Linda Yaccarino as its new Chief Executive Officer, a photo showing a woman wearing a navy blue suit with her cleavages exposed surfaced online with a claim that it shows the Twitter’s new CEO. Our findings using Google reverse image search tool however revealed that the pictures were created by a platform that produces adult images with Artificial Intelligence (AI). But this method is limited to images which have never been used before or if the AI-generated image has been modified to look different from the original.
3. Analyse image’s metadata
Images created by AI often exhibit metadata that differ from those obtained through a camera or smartphone devices. This metadata may encompass knowledge of the image’s composition, as well as the specific software and methodologies employed to produce them. Analyzing this information can aid in verifying an image’s authenticity or uncovering whether it was computer-generated.
EXIF.tools and Forensically are tools that can examine image metadata. Scrutinizing metadata provides insights into whether an image is probably AI-generated or not. Limited situations arise where an image’s metadata is intact.
4. Analyze the patterns
Images generated by AI frequently display distinct characteristics not typically found in natural images. These distinct characteristics can include repeated patterns or noticeable effects such as pixelation or blurriness. Analyzing these patterns may enable one to easily identify AI-generated images.
A prime example of this phenomenon appeared on social media, where an image depicting Donald Trump being apprehended by law enforcement was widely circulated. However, a noticeable feature on the photo was the blurred depiction of the characters, suggesting they were artificially rendered.
5. Use AI detection tools
Some researchers are developing AI detection tools that can verify if an image has been generated by AI. These tools use machine learning algorithms to analyze various features of the image, such as noise patterns etc. While these tools are still in development and may not be 100% accurate, they can be useful in gathering a plethora of evidence that can aid verification.
6. Check image’s Error Level Analysis
This helps to detect potential manipulation or alteration on an image. By examining error levels, ELA can pinpoint areas that may have been tampered with or generated, revealing any significant deviations.
7. Comparison with known AI-generated images
Due to the increased production of AI-generated images and their virality online, it is now easier to identify new AI-generated images by comparing them to existing ones. By comparing the image under observation with a collection of known AI-generated images, one may discover whether it is AI-generated or not.
8. Consult Experts
If you want to determine whether or not an image is real, it may be wise to consult experts in the related fields. Specialists in photography, computer graphics, or image processing are knowledgeable about this subject and can help determine if the image was made using artificial intelligence. Researchers and professionals who specialize or have done research works on AI-generated image detection might have access to the latest techniques and tools that can aid in the verification process.
As advancements in AI technology continue to evolve, new methods for detecting AI-generated images will likely be developed. The ongoing challenge of detecting these images necessitates a combination of various analytical methods and tools.
Nurudeen Akewushola is a fact-checker with FactCheckHub. He has authored several fact checks which have contributed to the fight against information disorder. You can reach him via [email protected] and @NurudeenAkewus1 via Twitter.