Evaluating Images in AI Era: Guide

Whether it's a snapshot of a cute panda on Instagram or a photo of a huge protest on a news site, we're used to trusting that photographs are a special record of reality.

But as AI imagery takes over the internet, it's become harder than ever to distinguish real from fake, fueling the spread of misinformation. People see images of individuals, places or moments that look photographed, even when they're not, and they believe they are authentic because the photos feel real.

In this new landscape, UO media studies expert Maxwell Foxman suggests a new approach to evaluating images in media and news: one that moves away from a binary of real versus fake and toward a more nuanced scale that captures how closely an image represents what it's trying to depict.

"It's hard to break the faith we place in photographs and videos," said Foxman, associate professor of media and game studies at the UO School of Journalism and Communication. "But as generative AI tools get more sophisticated, I think just having a binary - is it absolutely true or not - is going to be too simplistic for being able to evaluate information."

He proposes a concept called fidelity, a term borrowed from the fields of computer science and media production. Fidelity provides a framework that anyone can use when deciding whether an image deserves their trust, especially as AI-generated images are increasingly blurring the boundary between real and inauthentic.

"My hope is to provide a set of conceptual tools that anyone could use to start to ask better questions about the gap between what happened and how it is represented," Foxman said. "This kind of literacy is something any thoughtful media consumer can take up."

Applying fidelity to verify what can be trusted

An AI-generated image can look photorealistic but carry misinformation or disinformation. Fidelity is a measure of how accurately computer-generated media represents reality, which can help people determine whether it deserves to be trusted.

To assess if an image is of high- or low-fidelity, Foxman suggests asking six questions:

Fidelity versus objectivity

Photographs weren't always considered a source of objective truth in journalism, Foxman said. When cameras were invented in the 19th century, photos were initially seen as manipulative compared to other media devices like the printing press, he said.

But Foxman said public trust in photography grew in the 1930s through the 1950s. Journalists were able to establish their authority in opposition to technologies like radio and television by bringing photos of world events - from war to travel - easily and cheaply into people's homes. This helped them distinguish themselves and their work as "objective."

Objectivity has since become a core tenet of journalism, and our current visual intuition searches for whether something looks objectively real or not, Foxman said.

"It's important to remember objectivity is something as a concept that we kind of take for granted," he said. "We expect journalists to be objective, right?"

But AI poses particular challenges for objectivity, Foxman added.

Today, a computer-generated image can look photorealistic but carry misinformation or disinformation, or it can hold the truth but not be a perfect rendering of reality.

Objectivity asks a true-or-false question, Foxman said, whereas fidelity shifts toward a scale: How closely does the image draw from the actual event it is trying to represent?

"The first point of fidelity is that an image will never quite represent an actual event," he said. "But it can get really, really close - or really, really far - and you have to understand where it sits in representing that event."

Fidelity's role is aspirational, Foxman said, adding that representing the full reality may always remain out of reach. The concept also considers that the processes involved in creating the image are just as important as the final product.

"With low-fidelity work, you don't have a sense of how it represents real events," Foxman said. "There's also little to no explanation or audit trails that show where the creators of the image got data, how they used the data and what tools they used to create the image."

"As generative AI tools get more sophisticated, I think just having a binary - is it absolutely true or not - is going to be too simplistic for being able to evaluate information."

Maxwell Foxman

^{Associate Professor in Media and Game Studies}

Objectivity in journalism may need a rewrite

Foxman sees fidelity as both a public visual literacy tool and a professional standard, as generative AI and other computer-generated imagery become increasingly used in news and media. It can help audiences ask sharper questions about what they see, but it can also help journalists decide when and how to use AI and other digital tools responsibly.

Fidelity may be an alternative occupational goal to objectivity that can help strengthen the trust between journalists and their audiences.

"This is a first step in thinking beyond objectivity as a way to discuss the truth," Foxman said. "It is really a challenge to a core ideology for journalists, and it's meant to be."

Each new technology, from the camera to the internet to smartphones to now AI, has come with new tensions, Foxman said, reshaping how objectivity is practiced and perceived and how trust is formed between media makers and consumers.

Foxman urges newsrooms to look outward to other industries for lessons in integrating AI and computer-generated imagery with transparency, accountability and integrity.

"My next stage of the project is to look at what I call examples of 'journalism at the periphery,'" Foxman said, such as news through Twitch, immersive journalism, video games and comics. "People working in those media are also thinking very deeply about the relationship around truth and how to represent it through computer-generated images."

How UO Researchers Put AI to Work