This illustration has been attributed to the American artist David Suter. It explains the concept of “framing” very well.

A cameraman is filming with a camera a scene happening just in front of him. In the scene, a man runs away, followed by another man carrying a knife. But the tight framing of the camera, eliminating the context, reverses the threat: the fleeing man’s shoe looks like a knife; therefore, in the camera viewfinder, the pursuer seems threatened.

We don’t know if the cameraman has a zoom and voluntarily zoomed. We don’t know if there is a director that tells him what to frame. We don’t know if the two people crashed into the scene by chance, even if they’re acting. We don’t know anything except that the very presence of a technological means affects reality: it frames it, and, by framing it, in some way, it alters it.

For those who do journalism, it means that every time we use a tool (even if it were only our interpretation, our brain), we should remember that we are altering reality by using it. We should not forget how we inform and accidentally misinform while informing: what we choose to say and what we omit – not out of malice but because it is not part of our point of view, or maybe it does not enter the viewfinder through which we observe the reality, or it doesn’t fit our story – shape the story we are telling.

While informing the audience, we always run the risk of misinforming.

This simple, highly effective image should have always been a warning to journalists worldwide: if I am in the presence of a photo or video, I cannot use that photo or video as a source. Ever. Unless I can prove that the photo or the video is true and accurate. This was true even before the emergence of AI generative tools.

Fake is everywhere. Knowing technology and its history

Verifying the authenticity of a photo or video is a must: we need to ensure the image is not doctored, manipulated, or misrepresented. We also need to double-check many things: we need to be sure that a person pictured in a photo is not a double, for example. We need to know that a place exists. Cross-checking the image’s source, date, and location is part of this process, but more is needed: we must adapt the approach to the state of the art of technology.

We need to know all the existing possibilities to alter an image for starting.

Since 1800, artists and photographers manually painted or retouched photographs, often to improve the subject’s appearance or remove unwanted elements.

Another technique to manipulate pictures in the XIX century was multiple exposure and combination printing: exposing photographic plates multiple times or combining multiple negatives during printing was a technique to create composite images.

Airbrushing is a technique that involves using an airbrush — a small, air-operated tool — to spray paint, ink, or dye onto a surface, such as canvas, paper, or photographs. The airbrush mixes the pigment or colourant with compressed air, creating a fine mist that can be applied with precision and control. This technique allowed artists to alter, retouch, or manipulate the appearance of images.

In the darkroom, several techniques – dodging, burning, and masking – allowed photographers to manipulate the exposure and contrast of images during the printing process since the 1900s.

Then, photo montage emerged, and photographers started creating composite images by cutting and pasting multiple photographs or parts of pictures together.

In parallel, several video manipulation techniques emerged too: chroma-keying, video feedback, and time base correction enabled filmmakers to manipulate video signals to create special effects and alter the appearance of recorded footage.

Finally, we entered the digital and video manipulation age: software like Adobe Photoshop has made editing, manipulating, and altering images easier than ever. Various techniques, including cloning, healing, warping, colour correction, and content-aware filling, became popular and accessible to almost anyone.

At the same time, digital video editing software has enabled filmmakers to create realistic visual effects, motion graphics, and composite scenes using techniques like green screening, masking, and rotoscoping. These techniques became quickly available to anyone with a basic technological background, too.

And then, starting in the 2010s, deepfakes and Generative Adversarial Networks (GANs) came out.

Machine learning algorithms, such as deep neural networks, can now generate hyper-realistic fabricated images and videos by superimposing one person’s face onto another’s body or synthesising entirely new content: the so-called deepfakes.

Generative adversarial networks (GANs), a type of machine learning technique, can generate realistic images, videos, or other media by training two neural networks to compete against each other. This technology has enabled the creation of realistic, high-quality fabricated content. Moreover, using software like Midjourney, you can do this simply by prompting a text. Even Bing with its AI copilot set in the creative mode has this capability.

For example, here I wrote to Bing “Can you create for me the picture of a hyper realistic horse riding in an abandoned modern town?”

The result is four images that I can modify again and again through the machine, powered by Dall-E. As you can see the machine has got its own bias. For example, it added someone riding with the horse. Moreover, this is not the best we can do. While I’m writing, for example, Midjourney is better in creating hyper-realistic photos. So much better that it can be used to create pictures that are very difficult to distinguish from real photos shot in real life, unless you know where to look and what to do (for now. It will soon become nearly impossible just from looking).

Trump’s fake images are just the beginning

On March 20, 2023, Elliot Higgins, founder of Bellingcat, a well-known and award-winning investigative journalistic project, used a generative AI program, Midjourney, to generate a series of fake images showing Donald Trump being arrested. He made a thread about it on Twitter.

Higgins had done the same thing on March 16, 2023, with fake images of Putin being arrested.

Some of these images then ended up, accompanied by explanations, on the homepages of some newspapers. In some cases, the newsroom made specific choices like, for example, showing the picture of a smartphone with Higgins’ tweet (see this example from the Washington Post). In other cases, they used those fake images to illustrate their articles (see this example from la Repubblica, Italy).

I’m a big fan of Bellingcat’s work – if a journalist could be defined as a fan. Higgins made his joke clear. And some of those images are not realistic at all, if you look closely.

But those images started circulating decontextualised, as to be expected. And even if they could have an educational result – as Higgins himself argued –, helping the audience think twice before sharing a picture, I argued that, as journalists, we should not manage fake images this way at all. Even if it’s a joke.

Producing and spreading hyper-realistic images generated by AI is confusing: it increases the pollution of the infosphere. We should carefully consider the consequences of our approach to these technologies in a context where trust in newspapers is already at an all-time low.

This Trump case is just the beginning, as we are seeing. The Belanciaga Pope images came out as viral content. On several social networks, you can find any hyper-realistic photos from the fictional universe (for example, selfies from the Old Testaments) or from our universe (for example, historical facts never happened, like this series on Twitter about the never-happened “Great Cascadia,” an imaginary 9.1 earthquake plus tsunami in 2001. Please, share with caution in both cases!).

Photographs and videos are definitively losing their documentary value.

It will be necessary to have irrefutable proof of their authenticity. And we must be prepared for the future – for example, hyper-realistic and believable live broadcast videos.

What should we do as journalists?

First of all, we must consider all the problems related to the use we make of images of any kind. We must reflect on the consequences of the images (and videos) we spread. Then we have to make decisions.

A sensible decision would be to write directly on the image – so that it can hardly be used out of context – a sentence like “This image is fake and was generated with an AI”

In this way, even if anyone cannot be prevented from freely producing and distributing this kind of images, as journalists, we do not contribute to the pollution of content.

Then, we must decide once and for all that in the absence of obvious anchors to reality and three independent sources, an image or video is not a source at all: this may be consolidated in some journalistic traditions but not others (as in Italy, for example). Moreover, haste has often reversed the burden of proof: in some cases, people hurriedly publish, assuming the image or video is true, and then eventually correct it later. It would be better, of course, never to do this.

What if I can’t prove in any way that a photo or a video is real? Simple answer: I don’t publish them.

Finally, we must commit ourselves to disseminate and make this method popular with techniques and tools to track and spot fakes.

The Invid project

“As the power of AI rapidly advances,” Billy Perrigo wrote for Time, “it will only get harder to discern whether an image or video is real or fake. That could have a significant impact on the public’s susceptibility to foreign influence operations, the targeted harassment of individuals, and trust in the news.”

I’m not sure that it’s getting harder (I mean: we have to contextualise what “harder” means; it’s always been hard, depending on how familiar you are with the technologies and techniques I’ve described), but we definitely have to deal with this.

Perrigo chose to explain how to spot fake images in his article for Time. And this is a great idea. Much of the work could be done with tools like the INVID Project (now Vera.ai).

INVID Project developed free tools and technologies to assist journalists, news organisations, and other stakeholders in verifying user-generated videos and images shared on social media platforms, particularly in the context of breaking news.

INVID comes with several options that guide you to discover what you can and should do with videos and photos.

Source first – for example, you have to reverse search the picture (which means search the web through search engines using the picture itself). INVID can do that for you.

The point is that in any case the first thing we have to do with photos or videos is finding the original source. This could be difficult, but without a source we don’t have anything in our hands.

Then, you can analyse the picture – making sure that you’ve found the best possible high-resolution image for the analysis – looking closing or, for example, using the magnifier.

Time did this, choosing to follow the Balanciaga Pope trend explaining how to spot a deepfake simply with a magnifier tool and reasoning.

By the way, this is another possibility to spread the image by adding context and knowledge, helping ordinary people dealing with these pieces of content.

You have to look at possible inconsistencies of the image (fingers, reflections, lights, strange details, physical impossibilities).

You can use the forensic analysis provided by INVID to see if something strange is detected on the pixels’ structure.

Some software producers argue they are capable of spotting video deepfakes in real time, but we weren’t able to test those softwares yet.

The problem is that AI is changing so fast that tips and techniques to spot these images will have to quickly improve, and so you’ll always need to know more and more about context, history and other details about a picture, contacting people, finding different sources or videos or photos shot from other points of views, and so on.

This means, again, that we need to slow down, because it requires time to do this job. It will be more and more necessary.

And I believe we should decide once for all that we will never use AI generated hyper-realistic images without a clear statement printed on the image itself, to avoid any kind of possible confusion.

Deepfakes, GANs and how to deal with them – lessons from the fake Trump arrest pictures

Alberto Puliafito

Ideas, methods, techniques to verify photos and videos and to manage AI generated images in the newsroom

Author

Alberto Puliafito

‌

‌‌‌

‌‌‌