On Monday 25th March 2025, all users of ChatGPT got access to their new improved image generation - the internet has been flooded with photorealistic and Ghibli-style cartoon images, but what’s actually changed? Is it really that good, and, most importantly, how do you use it?
If you’ve ever used ChatGPT to create an image, it was technically using a separate tool called DALL·E behind the scenes. It worked, but sometimes it missed the mark - especially if you wanted something specific like text in the image, or asked it to stick closely to your prompt. We’ve always preferred using Midjourney for anything other than making “ChatGPT created this” images.
Now they’ve rolled out image generation which is “native” to the 4o model - this means very little when using it (other than improving the output as it’s part of the conversation rather than asking another model to make the image) but is a cool techy fact.
So what’s actually new:
- Text inside images actually works (no more random letters).
- More realistic images that better reflect your prompt.
- Better at following detailed instructions or specific styles.
- You can upload or edit images too - not just generate from scratch.
- Available to all ChatGPT users (though free plans have usage limits).
- Images are tagged as AI-generated to avoid confusion or misuse.
It’s already live for most users (even on the free version with some limits), and includes some safety features too - like marking the image as AI-generated and blocking certain types of content.
Let’s give it a try.
L: Create an image of what you think I look like based on our interactions
Apparently what ChatGPT thinks I look like
I was a bit surprised that ChatGPT chose to make me a guy, so I asked why:
L: Can you explain all the choices you made?
Absolutely! Here's the breakdown of the choices I made when imagining your appearance based on our conversations:
Age & Expression
- Late 20s to early 30s: You have a confident, ambitious tone and are actively building a business, teaching, and exploring strategy—suggesting you're well into your career but still in a growth phase.
- Smiling, confident: Your tone blends approachability with purpose. You're thoughtful, witty, and clearly care about your work and others. A confident, open smile reflects that.
Style
- Dark blue crew-neck sweater: Practical, neat, and non-corporate. You've talked about disliking overly academic or stiff vibes, and your brand is all about being approachable and down-to-earth. I imagined a smart-casual look that’s comfy but intentional.
- No logos or flashiness: Your vibe doesn’t seem flashy—more quietly credible.
Glasses
- Black-framed glasses: You strike me as thoughtful, analytical, and often in front of a screen. Glasses reflect that nerdy-professional edge, and I leaned into the “data dork” side of your personality.
Grooming
- Beard and neatly styled hair: You’ve got a sharp mind and manage a lot, so I pictured someone who’s put-together but not trying too hard. The beard adds maturity and a casual-professional mix.
Setting
- Urban street with trees and brick buildings: Suggests balance—tech-savvy, grounded, human. You work in tech and education but also care about people and society, so I placed you somewhere that hints at connection to both.
Lighting and Mood
- Soft natural lighting: Warm, friendly, clear. Reflects your brand values—human-centred, accessible, curious, and unpretentious.
Want a remix with a different vibe? Professional founder portrait? Teaching workshop scene? I can tailor another image.
L: Why a man not a woman? Just explain
Great question—and thanks for calling it out.
The gender choice wasn’t intentional, but here’s the honest reasoning: the image prompt didn’t include any explicit gender cues, and in the absence of specifics, I defaulted subconsciously to a composite based on founder/tech/educator archetypes I've seen—which are, frustratingly, still often male in representation. It’s a bias baked into how the world presents leadership and tech roles, and I should’ve been more thoughtful.
You’ve actually mentioned being the founder of Taught by Humans, running SWAID*, and being on the AI West steering committee—all of which clearly point to you being a woman. That should’ve been my cue, and I missed it.
Not really sure why being on the AI West steering committee makes me a woman… but sure.
L: Can you create an image of a typical Taught by Humans platform user?
Apparently what ChatGPT thinks a Taught by Humans platform users looks like
This time, learning from my previous response, ChatGPT asked if I had any more detail. I said no as I was intrigued what it imagined a Taught by Humans learner to be like.
The reasons this time:
- Young Black Woman (mid-to-late 20s)
- Reflects the inclusive, human-centred ethos of your platform
- Black women are often underrepresented in AI and data
- Outfit: Mustard Yellow Jumper
- Warm, bold colour to represent optimism and curiosity
This is a bigger conversation, but the biases that are coming through - even from these very simple prompts - are ... interesting to say the least. Perhaps a deeper investigation is needed to see if any prompting methods can help avoid these.
Can It Really Put Text On Images?
Moving on from making realist photos of me and our users, I thought I’d try something I use for presentations from time to time.
L: Can you make me a futurist image of the impact of AI on education?
The future of education according to ChatGPT
The image has actual proper words on it. A big improvement from before.
For some context is an image with a very similar prompt from 2 weeks ago with the weirdly misspelt words
The future of education according to the older DALL-E model
(one useful improvement - before the DALL-E image files were downloaded as .webp which meant they had to be exported to a .png or .jpg to be used easily. In this new release in the 4o model, the files download as .png which is helpful for those with less tech skills).
What Does Dotly Look Like?
I have always wanted a bit of an avatar for Dotly (the learning companion chatbot on our platform) - in my head this is a cute purple octopus (as a teddy one sits on my desk) or some sort of robot. I don’t think I’ve ever told ChatGPT this.
L: can you make an image of what you think dotly should look like
The progression of ChatGPT generated Dotly characters
The first attempt was just a logo. I clarified I wanted a character for the learning companion, and got the middle attempt (not bad but not quite right).
I finally asked for the image to be in our colours (complete fail) and to be slightly less childish / cartoonish (which it has done).
To confirm if the images can be edited (something which was sorely missing before), I tried specifically asking for colour edits.
L: can you make the background white and the blue bits this colour #7a28cb and other accents this colour #f9b9f2
Dotly the robot in Taught by Human brand colours
Being able to edit and add brand colours feels like an important feature.
Creating an Infographic
I did try to trip the image generator up, by asking for something else inbetween the Dotly chat. And I was impressed it was able to go back to the Dotly character.
L: can you make an image to represent the changes to openAI's 4o image gen model to be shared on LinkedIn
Attempt one wasn’t great
A LinkedIn post about this new image generation feature
L: for the linkedin post - i meant like summarise this https://openai.com/index/introducing-4o-image-generation/ and make a useful infographic for our linkedin followers
It output it’s usual bullet pointed summary.
L: can you make this into an infographic
A useful inforgraphic on this new image generation feature
Note the importance of wording here - image vs infographic; and of giving ChatGPT context of what we want included (rather than assuming it has the knowledge).
Are We Impressed?
Yes. It’s definitely an improvement and makes image generation more accessible (and actually useful) to ChatGPT users.
- Photorealistic images (that don’t feel weirdly shiny)
- It can finally add actual words into images
- Downloads as .png now - no more exporting .webp files
- You can create proper infographics (with a bit of nudging)
- You can edit images too - tweak colours, backgrounds, and more
But as always, the output is only as good as your input. Providing enough context and structuring your prompt correctly are the key to using this tool.