On this pageJump to a section
Meta will now write your headline, repaint the background behind your product and stretch your image to fit Reels. The announcement describes what those tools produce. It says nothing about who checks the output.1
That gap is yours. Since my first Meta campaign in 2014, the rule has never changed: the platform makes the ad possible, and you own what the ad says. AI just raises the volume. Fifty variations in a minute can mean fifty liabilities in a minute if nobody reads them.
What can Meta's generative ad tools actually change?
Per Meta's May 2024 announcement, the generative features sit under Advantage+ creative in Ads Manager (here is which enhancements to turn off) and cover a handful of jobs.1
- Image generation: variations of an original creative, including new scenery.
- Background generation: new backgrounds built around a product image.
- Image expansion: adjusting an asset to fit different aspect ratios on Reels and Feed.
- Text overlay: adding text to the image with a choice of fonts.
- Text generation: headline variations in addition to primary text, with Meta testing text that reflects a brand's voice.
Adoption is not small. Meta said its GenAI tools produced over 15 million ads in one month, made by more than a million advertisers, and it estimated a 7% conversion increase for businesses using image generation.3 Those are Meta's own figures, so read them as the platform's pitch, not as proof for your account.
15M+
The announcement page I read does not mention human review or labeling. That is not a criticism. It is a reminder that review is not built in.
Dara Denney, who works with brands on Meta creative, reviewed the ad formats she expected to convert in a video posted on January 23, 2025, and opened with a line that fits this problem:
Format matters but messaging is more important than ever.
That is the AI era in one sentence. The tools will happily change the format. The words still carry the claim, and the words are yours.
VideoDara Denney
10 Meta Ads Creative Formats That Convert Like Crazy in 2025
Her AI ads segment sorts the tools into five buckets (note the video is sponsored by Storyblocks).
Who is responsible when the AI writes the claim?
You are. Meta's advertising standards prohibit ads that use deceptive or misleading practices.4 The FTC's standard is the same idea from the regulator's side: ads must be truthful and not misleading, and you need adequate substantiation before the ad runs.2
Two details from the FTC guidance matter more once AI is writing copy.
Implied claims count
The FTC judges the net impression of the whole ad, including images and graphics, not just the words. If an ad suggests a benefit, it needs substantiation even if the benefit is never stated.2 A generated background of a sunlit gym next to a "fast results" headline is a claim, whether or not any sentence says so.
Testimonials do not rescue a claim
The same guidance says a testimonial cannot support a claim you could not substantiate directly, and a "results not typical" line alone does not fix a dramatic result.2 If a tool makes a testimonial-style variation more dramatic, the ad is not safer. It is riskier.
How do you build a claim ledger?
A claim ledger is the single most useful document you can write before turning on generative features. It takes an hour and it replaces a hundred arguments.
- List every claim your offer is allowed to make: outcome, speed, price, guarantee, credentials, comparisons.
- Next to each claim, write the proof: the document, study, policy page or data that backs it, and the date you checked it.
- Write the approved wording. Lock the exact sentence, plus two safe alternates.
- Write a banned list: phrases and promises the tools must never produce, such as guaranteed outcomes, timeframes you cannot back, and anything that sounds like a diagnosis.
- Date the ledger and give one person ownership of it.
Now every variation has a referee. Does the new headline say something that is not in the ledger? It does not ship.
What is a sane review process for AI variations?
Review in layers, cheapest first. Do not read every ad for every issue at once.
- ClaimsDoes the takeaway match the ledger?
- BrandProduct shown accurately, tone sounds like you
- DetailsHands, text, logos and faces
- ShipApproved, with the reason written down
| Let the tools do | Keep a human on |
|---|---|
| Resize and reframe for each placement | Every claim, number and promise |
| Alternate backgrounds for the same product | Whether the scene implies a result or a setting you cannot deliver |
| Headline phrasing variants from approved wording | Any new benefit the variant introduces |
| Font and overlay layout | Logo, colors and tone against your brand rules |
| Volume of options | Final approval and the reason you approved it |
Layer one: claims
Read the words, and watch for the generic phrasing that buyers can smell instantly, then read the picture. Ask what a reasonable person would take away from the whole ad. If the takeaway is not in the ledger, reject it.
Layer two: brand
AI backgrounds drift. Check that the product is shown accurately, that nothing is warped, and that the tone sounds like you. If the tool changes how a person or a product looks, compare it to the original.
Layer three: the weird stuff
Look at hands, text, logos and anything with a face. Generated images get small details wrong (which is why creative direction matters more now), and small details are what screenshots get shared for.
Where should AI creative start and stop in your account?
Start where the downside is small and the volume helps: resizing, background tests on products that do not make performance claims, and headline rephrasing inside approved wording. Meta says its retrieval systems now match ads to people at huge scale, and more distinct creative gives that system more to work with.3 That is a good reason to use the tools. It is not a reason to skip the ledger.
Stop where a mistake costs you an account, a refund wave or a regulator's attention: results claims, before-and-after style imagery, anything medical, and any ad where a real customer or a real person appears to speak.
The platform can make a thousand ads. It cannot take responsibility for one.
A simple weekly rhythm
- Monday: update the ledger if the offer, price or proof changed.
- Before launch: run every variation through the three layers.
- Midweek: check live ads for the same drift. Placements can crop things in surprising ways.
- Friday: archive winners with the reason you approved them, so the next batch learns from them.
None of this slows you down once it is a habit. Small teams and solo operators feel it most, because there is no second pair of eyes unless you build one in, which is why review steps are part of the systems I build for clients. It stops the one bad ad from costing more than the whole month of tests.
Keep building.Ekai
Questions people ask
Do I have to review ads that Meta's AI generates?
Yes. You are the advertiser, so the claim is yours no matter who or what wrote it. The FTC judges an ad by the overall impression it creates and expects you to hold substantiation before the ad runs. Meta's announcement describes the tools, not any review step, so build your own.
What does Meta's generative AI actually change in an ad?
As announced in May 2024, it can create image variations, generate backgrounds around a product, expand an image to fit different placements, add text overlays, and write headline and primary text variations, all inside Advantage+ creative in Ads Manager.
Can an AI-written headline create a false claim?
Yes. A headline that implies a result you never promised, such as a speed, a guarantee or a medical benefit, is still a claim. Implied claims need substantiation just like express ones.
What is a claim ledger?
A one-page list of every claim your brand is allowed to make, the proof behind each, and the exact words approved to express it. Anything an AI generates is checked against the ledger before it ships.
Sources
- 1
- 2
- 3
- 4
- 5
