When OpenAI teased a surprise livestream on Tuesday, I was certain it wasn’t exactly a surprise. Instead, we were probably looking at some sort of pre-planned damage control. The entire world knew DeepSeek V3 was going to be released this month, and the Chinese AI firm made it available this week to challenge ChatGPT and other competitors.
I expected OpenAI to mention release dates for GPT-4.5, which is now being tested, and GPT-5, which should be available next month. This would be enough to steer the conversation away from DeepSeek, at least for a while. While OpenAI did not end up talking about its next-gen AI models, I was right about the distraction.
ChatGPT’s new GPT-4o image generation is available around the world, and ChatGPT has gone viral on social media as a result. You can find countless examples online of AI-generated images with a single prompt in ChatGPT, and the results are insane.
These look to be image-generation capabilities that exceed what Dall-E could already do inside ChatGPT. The AI can finally generate written text, which used to be a telltale sign of AI-generated content, whether it was photos or videos.
The AI produces fast results, and, as you’ll see below, it has no problem using the faces of celebrities or copyrighted content. What I’m getting at is that the ChatGPT image powers are too good for its own good. Prepare for an internet where fake images can go viral overnight and where bad actors will abuse these AI features to mislead the masses.
As you might have guessed by now, the ChatGPT-created images have no visible signs that they were generated with AI. There’s no watermark, so most people will not be able to tell them apart from real photos. OpenAI marks them in metadata, but that’s useless. Here are some examples.
Text generation is awesome
I started my day testing the GPT-4o image generation abilities in ChatGPT. There was no new model to choose from, so I feared EU ChatGPT Plus users would not get the new image generation models for a while. I asked ChatGPT if it got new generation powers and then asked it to give me sources for its claims that it could indeed create images better than ever.
Once ChatGPT delivered accurate information, I asked it to rewrite its entire response in chalk on a blackboard in a classroom full of students. Here’s the image ChatGPT returned after a few minutes of thinking:
The text on the board lacks a short paragraph in the original response, but otherwise, it was flawless. It proved to me that the new 4o image generation powers were working as intended.
Drawing iPhones the right way
The next thing I did was to tell ChatGPT to draw me an iPhone 17 Air. The result was much better than any iPhone ChatGPT produced with Dall-E for me before yesterday.
Hilariously, the iPhone 17 “Air” in the image below looks just like the current iPhone 16 models. There’s nothing thin about it in ChatGPT’s design. But it proves a point we’ve made countless times before, one that even AI understands: All iPhones look similar from one year to the next.
ChatGPT gets many things right in the iPhone image. The phone has an Action button on the right side, the Dynamic Island at the top, and ultra-thin bezels. The antenna lines are visible, and the text and symbols are all up to what you’d expect from an official Apple iPhone marketing image, aside from the “iPhone 17 Air” marking on the wallpaper. Even the 9:41 time is correct.
Now that I’ve tried my hand at ChatGPT image generation, I went online to see what others did with the tool. I wasn’t surprised to see tons of examples.
Making iPhone wallpapers from photos
ChatGPT can create AI images from text prompts, but it can also adapt photos if you upload yours. Redditor CrypticKlii made an iPhone wallpaper with ChatGPT, telling the AI to turn a photo of a dog into a drawing. The result is amazing:
Lifelike fake images
Redditor dennislubberscom uploaded two separate images to ChatGPT and asked the AI to merge them. Specifically, he wanted a photo showing him (from picture A) sitting in the chair from picture B in the same landscape as picture A.
The result is a photo that looks almost lifelike. There’s not a doubt in my mind that the image ChatGPT created would pass as a real photo when shared online.
The creator of AGI
One Redditor gave the AI a simpler prompt, telling ChatGPT to “generate an image of the company
which will produce AGI first.”
ChatGPT promptly complied, producing this incredible Google logo. Yes, it’s hilarious that ChatGPT thinks Google will get there first. But the AI’s ability to create logos belonging to actual companies is easily something that can be abused.
Putting South Park characters in memes
I’m pretty sure you shouldn’t be able to use copyrighted content like South Park characters in AI-generated images, but ChatGPT doesn’t care because OpenAI doesn’t care. Here’s what Redditor Ivan_el_grande did with the AI:
Yes, you can put politicians in ChatGPT images
Redditor OkCrazyBruh used the new AI image generation to draw a painting of two well-known politicians lovingly sharing a plate of pasta in the style of Lady and the Tramp. I won’t mention their names. You’ll recognize them immediately. But, again, this is content ChatGPT should not be able to produce:
Want more Trump… and others?
See if you can name everyone in the ChatGPT-generated pic below and the characters they’re supposed to be. I’m sure none of the real people this image is based on would want to be there:
How about a photo of a Will Smith lookalike featuring… Will Smith?
Nobody is safe, ChatGPT can use anyone’s likeness to create lifelike photos that will be incredibly easy to fake and pass around. Here’s what foodie_geek did with Will Smith:
90’s movies that never existed
My favorite in the series of AI-generated images below is Chunnel. Who wouldn’t want to see Denzel Washington, Alan Rickman, and Bruce Willis in the same action flick?
How about generating original comics?
Rather than put copyrighted characters in ChatGPT images, Reddit user drekmonger used the AI to create comic book covers for made-up superheroes:
A man stealing rotisserie chicken while wearing full armor
Redditor gavinpurcell gave ChatGPT the following complex prompt:
a security cam still from a 1990s grocery store showing a man in full medieval armor stealing rotisserie chickens, frozen in mid-sprint past the dairy section, armor reflecting overhead fluorescent lights, baby blue tiled floors, timestamp reads “08/13/96 04:44 AM”, posters on wall say “NEW! TOASTER STRUDELS!”, motion blur adds chaotic energy, absurd yet intense, low-fidelity with VHS color bleed.
The result doesn’t disappoint:
Instructions to boil an egg
Redditor Ahndarodem created an infographic that teaches you how to boil eggs. What’s impressive here is that the AI-generated the text correctly and the instructions. Well, I don’t know about boiling eggs for 9-12 minutes, but it’s still a fine ChatGPT art result in my book:
NSFW art? ChatGPT got you
Well, how about AI images that aren’t exactly suited for work? Those must surely be banned. ChatGPT’s new image generation powers can’t do that, can they? Yes, they can, as seen below:
ChatGPT image generation is too good
It’s not just me thinking ChatGPT’s amazing image generation results are too good for its own good. Redditor Donastre created the following image, writing: “Seeing the new image generator and not noticing if its Al or not.”
How ChatGPT generates images
One Redditor posted a short clip on the social network to show how ChatGPT goes about generating images. As you can see below, the AI “paints” the image starting at the top, like a printer.
