6 Of The Most Accurate Ways To Spot AI-Generated Pictures
The line between AI and human art gets blurrier by the second. Many AI-generated art pieces can pass as human, for better or worse. Here's a few of my best tips and tricks for figuring out if something was written with AI or not.

John Angelo Yap
Updated November 1, 2023

Reading Time: 7 minutes
If I gave you two pictures to choose from, how certain are you that you can identify which is AI? It's quite hard. I even failed our very own test (but we don't talk about that).
But that's the problem with images today.
With the release of DALL-E 3 and Midjourney V6 probably around the corner, many of the popular AI image generators are becoming smarter and more powerful. Technology has sort of caught up with humans and reached a point where it can create images that make you question who made it.
With issues such as deepfakes becoming increasingly rampant, how can you protect yourself from falling victim to AI images?
Don’t worry. I got your back. Here are six easy and some of the most accurate methods of spotting AI-generated images.
How Can You Tell If It’s AI?
Look For Identifying Information
The quickest and easiest way to spot AI-generated images is by looking for watermarks. You can usually find this in the bottom side of the image. This might be a dead giveaway though.

This doesn’t work all the time. For one, Midjourney and the new DALL-E 3 model don't output images with watermarks. Users can also get rid of these watermarks by getting a paid subscription.
What you can do instead is to download the image. There are times when uploaders couldn’t be bothered to change an image’s original file name. The thing is, popular AI image generators often have a file name pattern that they follow, which makes them easy to spot if you’ve familiarized yourself.
- DALL-E 3: “DALL·E [Generation Date] [Generation Time] – [Original Prompt]”
- Adobe Firefly 2: “Firefly [Original Prompt] [Prompt UID]”
- Midjourney: “[Discord Username]_[Original Prompt Separated By Underscores]_[Prompt UID]”
Here's an example of each:
- DALL-E 3: DALL·E 2023-10-23 10.47.46 - a woman with striking blue eyes
- Adobe Firefly 2: Firefly a woman with striking blue eyes 71649
- Midjourney: discorduser_a_woman_with_striking_blue_eyes_eb1cf9-3a-4b7-8a2e
Google’s New “About This Image” Feature
Google just rolled out a new feature called “About This Image” that lets you check the origin of an image. To use this, select any image from Google Images and click the three vertical dots next to the close button. You should see this option now.

You’ll be directed to a separate tab containing every web result that includes the image. From there, you should now have a better idea of the image’s source and whether or not it came from AI.

Unfortunately, this is only available for Google Images. For third-party images, you can use the always reliable Google Lens instead.

Check For Glitches
We recently published an article about identifying which between two images are AI and human. The following image is an excerpt from that article. Can you spot which isn’t real?

You’re getting a gold star if you said the one on the left. But how can you tell exactly? Well, it’s easier just to show you.

Notice how the edges of some petals get distorted? Some even blend together. This is common in AI images. You can also see that the blurring of the foreground and the background isn’t consistent, particularly on the daisies on the bottom left.
This symptom is most apparent in faces. Take this image from Midjourney as an example:

You can count on one hand how many of these faces aren’t distorted or completely warped, and they’re usually the ones on the foreground. This phenomenon isn’t just limited to faces too. Sometimes, even other body parts can be affected by rendering issues, commonly manifesting as extra fingers or missing limbs. There are times when the results are nightmarish and borderline body horror images though.

Generative models still also register letters as shapes and couldn’t comprehend their meaning. This results in bad text generation, something that no AI model can fix at their current stage.

Focus On The Supporting Details
When you daydream, do you spend a lot of time thinking of the background details? What do the people look like? What language do they speak? What’s the weather like?
The same goes for AI. They don’t dwell on the supporting details. Instead, they focus their computing power in the foreground, which is also why you don’t see much face distortion in the main subjects of the image earlier.
To demonstrate this, let me give you an example:

Most people would immediately glance in the middle of the image. However, when you look closely, you’ll start to notice that some things aren’t adding up.

These are all just hidden in plain sight. Here’s another example featuring cars on the sidewalk, distorted background buildings, and missing heads.

Apply Some Logic
What's wrong with this image?

If the watermelons are inside the house, how could they get wet from the rain? Oh, and who the hell cuts watermelons like that?!
That’s the problem with AI image generators: they lack common sense. Here’s another one:

There shouldn’t really be a chair atop another chair. The shadow of the table’s centerpiece also isn’t consistent with the placement of the lights.

Why is the train in the passenger waiting area instead of the tracks?
AI lacks the nuance that we have. They don’t see the world for what it is, but they see it as an aggregate of elements. They don’t have an exact idea of how they fit together, but they know some objects and styles belong together.
Here’s a short list of logic errors that’s common in AI images:
- Extra or missing fingers.
- Excessive repetition of elements.
- Inconsistent lighting and shadows.
- Lack of flaws and imperfections in a face.
Use AI Image Detectors (I did some testing too!)
Lastly, you could also use detectors to find out if an image came from AI or not. There are some that I believe deserve a mention, starting with Optic’s AI Or Not: a web application that identifies the origin of an image. This is the most popular option, with media outlets such as New York Times and the Wall Street Journal highlighting its effectiveness.

If you like statistics, you’re going to love Hive Moderation. It’s a complete image suite which includes features such as visual moderation, demographic attributes, and reverse image search. And get this: it doesn’t just give you a likelihood score for AI, it even outputs a confidence score for each generative model.

Another tool you can try is Illuminarty. Like AI Or Not, this one’s pretty simple. The biggest difference is that you can get an AI probability score even with their free plan, which isn’t the case for AI Or Not.

The most important aspect of these detectors is their accuracy. So, I went ahead and did my own testing. One big benefit of writing about Midjourney, DALL-E, and Firefly for months is that I have my own collection of AI images, some of which I used here.
Optic's AI Or Not | Hive | Illuminarty | |
Test #1: Midjourney | |||
Test #2: Midjourney | |||
Test #3: Midjourney | |||
Test #4: Midjourney | |||
Test #5: Midjourney | |||
Test #6: DALL-E 3 | |||
Test #7: DALL-E 3 | |||
Test #8: DALL-E 3 | |||
Test #9: DALL-E 3 | |||
Test #10: DALL-E 3 | |||
Test #11: Firefly 2 | |||
Test #12: Firefly 2 | |||
Test #13: Firefly 2 | |||
Test #14: Firefly 2 | |||
Test #15: Firefly 2 | |||
Overall |
I also wanted to see how effective these tools are at detecting human images. It’s an effective way of determining if you’re at risk of false positives when using them.
Optic's AI Or Not | Hive | Illuminarty | |
Test #1: Human | |||
Test #2: Human | |||
Test #3: Human | |||
Test #4: Human | |||
Test #5: Human | |||
Overall |
You can verify these if you want to. For transparency, here’s a Google Drive folder containing all images I used for these tests.
From these results, I’d say that Hive is the most effective AI image detector in the market. Of course, this is just a small sample size so be diligent in researching which one you’ll trust.
In A Nutshell
So, what’s the most effective way of spotting AI images?
You can always look for watermarks, but they’re not always there. Some images won’t appear using Google’s “About This Image.” Looking for inconsistencies, distortions, and logic errors is a bit like a game of Where’s Waldo and could take some time.
I recommend going straight to AI image detectors if there are no watermarks. Not only is this the quickest, but it’s also the most accurate.
However, the most important thing is to keep a sharp eye. Once you get used to seeing AI images, I promise you that identifying them will be as easy as pie.
Want to Learn Even More?
If you enjoyed this article, subscribe to our free newsletter where we share tips & tricks on how to use tech & AI to grow and optimize your business, career, and life.