You Can’t Identify AI Images as Well as You Think
This is not a real person, and it’s not even a photo, but an AI generated image. This is from the website This Person Does Not Exist, which was created over 5 years ago. AI has improved a lot since then.
In case you haven’t noticed yet, artificial intelligence has taken the world by storm. I’ve been extremely fascinated with the various facets and applications of AI since early 2019, when I saw the website This Person Does Not Exist, which upon every refresh of the page, generates a hyper realistic portrait of a human that does not actually exist. But that’s the AI of 5 years ago. As time marches forward, more and more AI images that we see online are indistinguishable from real photos. That website was one of the earliest examples of AI being widely publicly accessible, and since then, countless websites, apps, and integrations have popped up, from virtual assistants, to summaries of lab results for medical patients. It’s clear to me that AI falls into the same realm as the internet and smartphones - it's one of the most important modern technological advancements with a huge impact on culture and society that we've only just begun to see. Whatever your current view on the topic is, AI is already a part of most of our everyday lives, and it will only become more so as it continues to be baked into our phones, computers, and other electronics.
These AI images were shared on various Facebook pages which feature nature and landscape “Photos”. They are clearly different versions of the same concept. While these could be based on an actual place and rock formation, I was not able to find an original photo from which these were derived, so if anybody does know of a photographer with a similar photo, please let me know.
While there are many fascinating aspects on the topic of AI, today I’m going to keep a narrow scope. My main purpose is to show you that you can’t identify AI images as well as you might think you can. Along the way, I’ll be pointing out a practice I’m seeing that I consider a dark side of AI, where photos are being stolen and regenerated, with credit rarely given to the photographer or artist. As you scroll on social media, you’re likely swiping past many of these AI images without even raising an eyebrow. If you want to test your identification skills before proceeding, check out this challenge on my Instagram. So far, only a few people out of the nearly 300 participants have correctly guessed which images in the series are AI.
Typical comments on AI generated landscape images shared on nature focused Facebook pages. Many have hundreds or even thousands of comments from people that apparently have no idea the image they are viewing is not real. This is not limited to landscape photography - regardless of niche, the sharing of these types of AI images is rampant.
A few important notes before we dive in deeper. None of the images shown below are mine. Many Facebook pages are pushing out a large volume of AI images where a single actual photo was used for reference, but credit is rarely given to the original photographer. I found these AI versions of landscape photos which I recognized or was eventually able to track down where the original came from, and I’ve spoken with all of the original photographers for permission to use these examples. You’ll find links to their Instagram pages in the caption for each one. Though below I only compare the original photo to one AI version, in some cases, I found several different versions, similar to the balanced rock above. The AI copies often go viral, since they utilize photos which are already known to be viral, and they’re widely viewed and praised as real photos, as seen in the screen grab above. Of course, it’s hard to say how many of the comments are bots versus real people, but what I tend to see is there are a handful of bot comments, which get the algorithms to give the post more attention, then real people start seeing it and commenting on it as well, so there are certainly plenty of people that are not questioning the authenticity of these images. Occasionally, I see some comments pointing out that it’s fake or AI, but those are few and far between, and more often than not it seems people don’t even second guess them.
You Can’t identify AI images as well as you think
If you’re thinking, ‘I definitely notice what’s AI and what's not’, keep in mind that you never know for the ones you don't question, or the ones you get wrong. As long as you’re reading a blog post or article about AI images, you might think you can tell the difference, but had you seen the same image on a Facebook profile you’d be none the wiser. Of course most people have no problem identifying the obvious and easy ones, such as where people have 6 or 7 fingers on one hand, but there are plenty which lack any clear indication that the image was AI generated. My point is that as you casually scroll on social media, you likely are seeing AI images and not realizing it. The reason that some of these images look like real photos, is because they are taking a real photo and simply regenerating a new version of it, as we’ll see in 4 examples below.
Use the slider above to compare the original photo and the AI regenerated version, which appeared on a Facebook page which consistently posts these types of AI copies of nature and landscape photos, without credit to the original photographer. Notice the change of texture of the rocks, removal of backpack, and different hair and shoes. If you’ve had a photo go viral or even semi-viral before, it’s quite possible that there already exists at least one AI version of it out there somewhere.
Original Photographer: ashley.goes.hiking
Most of us have at least a general idea of how AI image generation works that goes something like this:
A process uses machine learning to create new images based on text prompts, drawing upon a library of existing images to inform its understanding of what different objects and concepts look like. What many people don't know though, is that you can feed an image generator a single photo or image, and have it recreate a new version, as opposed to providing a text prompt. Sometimes this only introduces minor differences, such as the hair color of a person, but it can also make major updates. In these cases, we can essentially think of it as an ‘AI filter’, where textures and some objects change, but the overall concept of what’s there is the same. When this method is used, as opposed to generating an image “from scratch” using a text prompt, it often looks just as real as the original photo, with differences so subtle that even the original photographer may not notice.
A young girl is swapped out for an older woman facing the opposite direction, and you can see different textures and terrain.
Original Photographer: jess.kesti96
Who cares? Why does this matter?
My focus today has been on landscape AI images, however there’s a much bigger picture to look at here. Regardless of your specific niche, there already exists a mountain of AI generated content within it, and beyond fake images, it seems that the share of the content we see on the internet that is no longer created by humans is instead generated by AI will continue to grow. With the rise of powerful AI models, it’s becoming increasingly easy to generate convincing text that’s almost indistinguishable from human writing. This technology has been used to create fake news articles, social media posts, and even entire websites. Where this all ends up we can only guess, but many people, including myself, are certainly finding the internet at large as a less and less meaningful “place” to spend our time. At the very least, this is something we need to raise our awareness on.
At first glance it may just appear to be a sky swap, but upon closer examination, it’s easy to see that the entire image was regenerated using AI.
Original Photographer: marco.miglionico
Just like the internet and smartphones, AI can be used for good, and it can be used for bad. I’m excited for all of the useful applications which increase efficiency and save us time, as well as utilization in health care and other important areas of our lives, but I think it’s good to be aware and cautious of the downsides. It kind of seems pointless to try and ‘train our eyes’ to detect AI, because it will only get harder and harder to discern real photos from fake ones. The share of AI images that we can detect will shrink and shrink until it’s 1% or less. If I believe a landscape photo is real and I like it and comment on it, does that hurt me at all? Maybe not, but AI is already being used for more nefarious purposes, so I should be wary of the dark side of it, and the ways it can be used for scamming, which is already being done at an alarming rate. As the internet is increasingly becoming a vast graveyard of AI-generated content, where human voices are drowned out or even mimicked by AI, it’s less a place for authentic human interaction and creativity, and more a platform dominated by machine-generated content.
If you were scrolling social media and saw this AI recreation, how would you know that it was AI?
Original Photographer: yuri_vantowski
