On this pageJump to a section
Making a decent ad image used to take a designer, a brief and a day. As of this week it takes a sentence and about a minute.
That is a big shift, and it will feel like a gift for about a month. Then every competitor will be able to do it too, and the market will fill up with competent, forgettable images. The bottleneck moves. It stops being "can we make it" and becomes "should we run it, and is it true."
I run ads for a living, so I am delighted by the tools. I am also wary of what unlimited volume does to an account without someone steering.
What did OpenAI release on March 25, 2025?
An image generator built into GPT-4o, and rolled out as the default image generator in ChatGPT.
What it can do
OpenAI says it accurately renders text, follows detailed prompts, and lets you refine images through natural conversation, keeping consistency across iterations. It says other systems struggle with about 5 to 8 objects while this one can handle 10 to 20 different objects. Users can specify things like aspect ratio, exact colors using hex codes, or a transparent background.1
What it cannot do yet
OpenAI lists its own limitations: cropping, hallucinations, high binding problems, precise graphing, multilingual text rendering, editing precision, and dense information with small text.1
Occasionally crop longer images, like posters, too tightly.
It also says images take longer to render, often up to one minute, and that all generated images come with C2PA metadata identifying them as coming from GPT-4o.1
Is Meta already doing this inside Ads Manager?
Yes, for a while.
Meta's tools
In May 2024 Meta announced generative AI features for advertisers, including full image variations based on an original creative, text overlays with a dozen typefaces, image expansion to different aspect ratios, and text generation for headline variations. It reported advertiser-run results such as a 13% ROAS increase for Casetify testing background generation. That is one advertiser's reported result, not a promise.2
The volume is already huge
Meta's Andromeda post (which I broke down in Meta Andromeda) says more than a million advertisers used its GenAI tools to create more than 15 million ads in a month, and expects the number of creatives in its systems to grow significantly.3
15M+
If anyone can make images, what becomes scarce?
Judgment. Four kinds in particular.
| Cheap now | Still scarce |
|---|---|
| A hundred image variations | The one idea worth making |
| Polished visuals in seconds | A visual identity that is recognizably yours |
| Rewriting copy at scale | Claims that are true and provable |
| Fast iteration | Taste about what to cut |
| Looking professional | Earning trust |
Direction
Someone has to say which problem to lead with, which audience to speak to and which feeling to create. A model will make anything you ask. It will not tell you what to ask for.
Dara Denney, who runs creative strategy for brands on Meta, framed the same thing in a video she posted on January 23, 2025:
format matters but messaging is more important than ever
A model can give you ten formats in ten minutes. It cannot tell you which message is worth saying, and that is the part I would never hand over.
VideoDara Denney
10 Meta Ads Creative Formats That Convert Like Crazy in 2025
Her fifth format covers AI ads and the five buckets of AI tools she sees. The video is sponsored by Storyblocks.
Brand consistency
Run fifty AI images with fifty slightly different looks and you have no brand, the same sameness problem I described in The AI Copy Tell. A style guide, reference images and a clear standard for color, lighting, composition and tone are what make generated work look like yours.
What is a five-gate review for AI-generated ads?
Run every AI image through these gates before it goes live. Fail one, and it goes back.
- Brand gate. Does it match your look, voice and the page it leads to? An ad that looks nothing like the landing page creates a jolt.
- Claim gate. Is every statement in the image and copy true, and can you prove it? Do not let a model invent a result, a number or a testimonial.
- Defect gate. Zoom in. Check faces, teeth, hands, text, edges and cropping. OpenAI itself notes cropping and small text as weak spots.1
- Authenticity gate. Does the image suggest a real customer, a real result or a real place that does not exist? If so, change it.
- Test gate. Run it against human-made or earlier creative. Let results, not novelty, decide.
On fake people and fake proof
The FTC's August 2024 reviews rule bars testimonials from people who do not exist, AI-generated ones included.4 A synthetic person presented as a satisfied customer sits right in that territory. Use AI images for illustration, mood and product visuals, and keep real testimonials real.
What does a good brief for an AI image look like?
The same as a good brief for a human designer, only more specific. OpenAI's own examples show how much detail the model can follow, so give it some.
- The job. What this image must make the viewer feel or understand in one second.
- The subject. Who or what is in it, and what they are doing.
- The setting and light. Where it takes place, and what the light looks like.
- The framing. Aspect ratio, where the subject sits, and how much empty space to leave for text.
- The palette. Your brand colors, by name or hex code.
- The exclusions. What must not appear: logos you do not own, brand names, claims, anything that implies a result.
Then iterate in conversation. Vary a single element per test so you know what caused what. Save the prompts that produce on-brand results in a shared file, so the next image starts from a proven base instead of a blank box.
How do you use AI creative without wrecking your account?
Treat it as a production tool inside a system, not a replacement for one.
Write the standard first
Before generating anything, write down your visual rules: colors, lighting, framing, what people should look like, what is off limits. Keep three or four reference images that define the look. Every prompt starts from that standard.
Generate for concepts, not just variations
The temptation is to produce a hundred versions of one idea. Resist it. Use the speed to explore ten different ideas, then pick the best three to develop. Different reasons to buy, not different filters, and a steady testing cadence tells you which ones work.
Keep a human on the last mile
Someone with taste should approve every image that runs. That person is not slowing you down. They are the quality control that keeps cheap volume from becoming cheap-looking work.
When anyone can generate an image, creative direction becomes the product. The tools are about to be everywhere. The operators who stand out will be the ones with a point of view and the discipline to check their work. That standard is how I approach creative in the system I build for clients.
Keep building.Ekai
Questions people ask
Can AI image generation replace ad designers?
It can replace a lot of first-draft production. It does not replace the person who decides which ideas are worth making, keeps the brand consistent, checks claims and spots defects. That role becomes more important as volume gets cheap.
What are the known limits of 4o image generation?
OpenAI lists cropping, hallucinations, high binding problems, precise graphing, multilingual text rendering, editing precision and dense information with small text. It says the model can occasionally crop longer images too tightly.
Is AI-generated creative safe to run in ads?
It can be, with review. Check brand fit, make sure every claim is true, look for visual defects, avoid fake people presented as real customers, and test performance like any other creative. The FTC's 2024 rule bans fake reviews and testimonials, including AI-generated ones.
Does Meta offer its own AI creative tools?
Yes. Meta announced generative AI features for ads in May 2024, including image variations, text overlay, image expansion and text generation, and has said more than a million advertisers used its GenAI tools to create more than 15 million ads in a month.
Sources
- 1
- 2
- 3
- 4
- 5
