Meta is editing your ads, and nobody checked what actually shipped

If you buy Meta ads from a LATAM account, there is a good chance the ad you approved is not the ad your audience is seeing. That is not a bug, it is the product design. Meta ships a bundle of automatic creative enhancements turned on by default: it crops, expands, rewrites, translates and scores your creative with music. Most accounts we audit have them on, and nobody on the team ever made that call.

This is not an argument against AI in creative. It is an argument against signing off on something you cannot read.

What is on without you choosing it

The enhancements sit at the ad level, inside the creative section, and arrive switched on when you build the campaign through the recommended Ads Manager flow. It is a block of toggles almost nobody expands. What is inside:

• Image expansion: the model generates new pixels to fill vertical formats. Your 1:1 asset becomes 9:16 with an invented background. If your logo sat near the edge or your framing was deliberate, neither survives.
• Text improvements: headline, description and primary text get reordered, and a line from the body can get promoted to the headline. That line was not written to be a headline.
• Generated text variations: new copy built from yours plus whatever is on your landing page.
• Catalog overlays: badges like "Free shipping", "Best seller" or a price stamped on the product image, pulled from your feed.
• Music, filters, vignettes and automatic cropping on Reels and Stories.
• Automatic translation of ad text into the viewer language.

What breaks in LATAM specifically

In a single-language, single-currency market with one promo running, these do little harm. In LATAM three things break at once.

Price. A catalog overlay showing a stale price, or the right number in the wrong currency, is not a cosmetic problem — it is misleading advertising, and depending on the country it is a consumer-protection matter. Run Guatemala, Mexico and Panama from one account with three currencies and three catalogs, and the odds of a bad badge are high. You will find out from a customer comment, not from a report.

Promotion. Text improvements do not know your commercial calendar. They promote whichever line performs best statistically, which is sometimes the one naming the discount that ended last Friday. Nobody catches it, because the headline you wrote is still sitting right there in the field, untouched.

Voice. This is the one that stings if you write in neutral Spanish on purpose. Auto-translation and rewriting inject regional turns you did not pick, or flatten locally-written copy into textbook Spanish. If your brand speaks to Venezuelans like a Venezuelan and the machine sands that down, you just lost the only thing that made you sound different. No client has ever asked for that.

The real issue: you cannot audit what shipped

The editing is not the dangerous part. The reporting is. Ads Manager shows you the ad performing as one unit, not the version that actually served. When an ad with enhancements on performs well, you do not know whether it was your copy or the model's. When it tanks, same blind spot. Which means your creative testing process — the thing you spend weeks on — is measuring variables you do not control.

It is the familiar trade with platform automation: it is not that it performs badly, it is that it takes away your ability to learn. And learning is the only asset you still own when CPMs climb.

How to audit it in one afternoon

This is not a project. It is an afternoon, in this order:

• Open your ten highest-spending ads from the last month and expand the automatic enhancements panel on each. Write down what is active. You will find switches nobody flipped by hand.
• Use placement-level preview, not the default one. Feed, Reels, Stories and Audience Network will each show you a different crop. Screenshot every one.
• Look up your own brand in Meta Ad Library. It is public and it shows what is genuinely running, not what you uploaded. Fastest check available, and the one that produces the most surprises.
• In the asset-level breakdown, read performance image by image. If a variant you do not recognise is soaking up delivery, that is your answer.
• Kill anything touching price, promos or legal text first. Catalog overlays and text improvements come off any campaign with a live offer or a verifiable claim.

When leaving them on is the right call

Switching everything off by reflex is its own kind of lazy. There are real cases: cold prospecting with evergreen creative, low-production accounts that need variety and have no designer, vertical-format tests when the 9:16 asset simply does not exist. There, image expansion and text variations give you creative inventory you would not otherwise have.

The working rule: leave them on where the cost of being wrong is low and variety is scarce. Turn them off wherever there is a number, a date, a legal condition or a brand voice you care about.

Judgment, not configuration

A year from now these toggles will have different names and there will be three more. The principle holds: do not publish what you cannot read first, and do not celebrate a result you cannot attribute. If the platform edits your ad, let it be because you decided it should, in that campaign, for that reason.

Run the Ad Library check today. It is free, it takes ten minutes, and it is much better that you find it than a customer in the comments.