避免被AI内容误导的三种方法

Global society makes billions of images and uploads hundreds of thousands of hours of video on the internet every day.

The problem is, some of this content is misleading or downright wrong. And when it’s in visual form, it can be particularly convincing.

Take the Met Gala that happened earlier this month in New York. While photographers snapped photos of Rhianna, Beyoncé and Nicole Kidman as they strutted their stuff, others saw “photos” of celebrities, such as Rosalía, Lady Gaga and Jacob Elordi, who were actually elsewhere (the images in the below Instagram carousel are AI generated) .

While this type of AI slop might seem harmless and can be easily verified, other “media fakery” is becoming far more problematic and demands more robust techniques to verify.

Traditional verification techniques are falling short as AI becomes increasingly convincing and the line between authentic and synthetic blurs. This is true across all content, from still images to moving ones and audio deepfakes.

The volume of content and the speed at which it travels doesn’t help. It also doesn’t help that fact-checking can take hours or days while fakes can be created in seconds.

First, equip yourself

Guides on detecting AI-generated content suggest multiple strategies and acknowledge there are no perfect solutions. But there are helpful things you can do.

Familiarise yourself with examples of fakes and study how they were fact-checked. This helps you understand what is possible and learn how fact-checkers sort real from fake.

Look deeply. Zoom in. Pause the content or watch it frame-by-frame. Inspect the small details. Look out for inconsistencies, textures that are flat when they shouldn’t be, or patterns that are too perfect or are inexplicably off. Does the location shown match with where the scene is purported to be? Do shadows fall naturally and do lines follow the rules of perspective?

Look widely. Are you familiar with the source? What else does it publish and how long has it been around? What do other trusted sources say? How does this depiction compare to others that are available? Or if there aren’t others available, should that give you pause?

Then, apply your learnings

Let’s take an example and work through it together.

This Facebook reel, posted by an account called “Real Talk Hub”, purports to show migrants being stopped and returned by Australian police at an airport.

Before getting too granular, let’s take stock of the opening image.

The video uses scale to show what appears to be a long stream of passengers. Some are moving toward and some are moving away from a plane. It is difficult to identify specifics in the video. The superimposed text blocks almost all of the horizon line. Shallow depth of field makes aspects in the distance blurry and hard to discern.

Many of the passengers have darker skin and are visually coded as “other”. They interact with a light-skinned police officer who takes notes on a clipboard.

The vertical video is framed carefully to not reveal identifiers like the name of the airline that seems to start with the letter “P”. This makes it difficult to search the airline’s name and whether credible sources corroborate the story that’s told.

Even though the people and scenes look realistic at first glance, the video’s integrity unravels when we slow down and look closer. People in the passenger line morph and transform.

The officer is able to single-handedly remove the paper from the clipboard and it appears to inexplicably leave white strips behind. The police vests look different to images you can find in verified media photos of the Australian Federal Police.

Taken together, all these clues suggest the video is AI-generated.

Think like a fact-checker

Many AI-generated videos can trick you and create a very compelling narrative. So, fact-checkers have developed triangulated methodologies that examine elements beyond just what you see in the video.

One way to do this is to systematically check contextual factors – the other things surrounding the content. Our team’s research has found professional fact-checkers usually pay attention to the type of social media accounts or websites distributing suspicious media.

For this AAP verification on a video about banning dogs on the beach, it was crucial to inspect the user’s activity and posting patterns.

In addition to visual anomalies, the fact-checkers also found an invisible watermark that helped them determine the content was AI-generated.

Other things to check are how long a social media account has been operating, how often the social media account posts, and whether the account is transparent about its use of AI.

These aren’t fool-proof indicators of authenticity, though. The migrant example above comes from an account that is about five years old. It also comes from a “verified” account, which might make it feel more credible. But both Facebook and X now let users pay for this verification.

Overall, when it comes to suspect images or video, don’t just look deeply. Also look widely.

AI-generated content can increasingly fool our eyes, so you also have to look beyond what’s in the video. Taking a mixed-methods approach that considers visual and contextual clues can help. By training your ability to think like a fact-checker, you can stay safer online.

Silvia Montaña-Niño is also associate investigator of the ARC Centre of Excellence for Automated Decision-Making & Society and the Fact Check Research Team at this centre.

T.J. Thomson receives funding from the Australian Research Council. He is an affiliate with the ARC Centre of Excellence for Automated Decision Making & Society.