AI-Generated Advertising and Brazilian Consumer Law: Avoid Misleading Ads

AI-Generated Advertising and Brazilian Consumer Law: Avoid Misleading Ads

Advertising made with artificial intelligence follows the same rules of Brazil's Consumer Defense Code (CDC) that apply to any other ad. It cannot lead the consumer into error, and the company answers for what it publishes, even if the image or text came out of a tool. The practical risk sits in images that show a product, result or setting that does not exist, and in copy that promises terms the company does not honor. The fix is a simple review routine, with a named owner, before any piece goes live.

Key takeaways

• The CDC does not ask which tool created the ad. It asks whether the consumer can be misled about the product, service, price or terms.
• AI-generated images and simulations are the most sensitive point, because they look like visual proof of what is being sold.
• Leaving out an essential detail (deadline, condition, limit, what is included) can also amount to misleading advertising.
• The advertiser must be able to prove what it claimed, so the review should leave a record.
• A short review routine with a checklist and human approval reduces more risk than banning AI use.

What the CDC says and why AI does not change the rule

Brazil's Consumer Defense Code covers advertising in a set of articles that every company that advertises should know. Under the text of the law, advertising must be identifiable as such, misleading or abusive advertising is prohibited, and a sufficiently precise offer binds whoever made it. The burden of proving that the information is true and correct also falls on the party sponsoring the advertising.

The central point is the definition of misleading advertising: any advertising information or communication that is entirely or partly false, or capable of leading someone into error in any other way, including by omission, about the characteristics, quality, quantity, origin, price or other data of the product or service. Nothing in that definition depends on how the piece was produced. A hand-retouched photo, a stock image and an AI-generated image all go through the same analysis.

What AI changes is scale and ease. Before, creating a photo of a product that did not exist took a designer and time. Today anyone on the team can generate dozens of variations in minutes, and each one can carry an error nobody noticed. More volume means a higher chance that a problem slips through without review.

A note on sources: what we describe here comes from the text of the CDC (official documentation) and from a practical reading of how companies use AI in marketing. It is not legal advice. For specific cases in your industry, consult a lawyer with experience in consumer law and advertising.

Where AI-made ads tend to mislead

The same problems keep showing up in a few places. Knowing them helps you build the review.

• Product images that do not match the real product: finish, color, size, number of items, or accessories that do not come with the delivery.
• Idealized settings or results: a facade, a view, a kitchen or a before and after that the customer will not receive.
• Nonexistent people presented as satisfied customers, which suggests a real testimonial when there is none.
• Copy with absolute promises such as "no fees", "guaranteed delivery" or "results in a few days", written by the AI because it sounds persuasive, with no basis in what the company actually does.
• Numerical simulations, such as savings, installments, timelines or returns, generated without the assumptions and conditions in view.
• A price or campaign condition that has changed but still appears in an old piece that was reused.
• Omission of what is essential: the offer is highlighted while the limit, deadline or extra cost is hidden or missing.

A language model tends to produce fluent, confident text. That is useful for drafting but risky for factual claims, because fluency does not mean the claim is true. The same logic applies to images: the result looks photographic, and that is why it convinces.

Illustrative scenario: the simulation that looked precise

Illustrative scenario: a home renovation company asks an AI tool for a social media ad. The tool generates an image of a spacious living room with a high-end finish and a line of copy saying "complete renovation, no surprises in the quote". The marketing team likes it, adjusts the color and publishes.

What can go wrong: the image shows a finish standard that is not in the advertised package. The phrase "no surprises" is an absolute promise, and the company knows some items are only identified after the technical visit. A consumer who hires based on this piece can argue they were misled, and the company will have to show that the information was correct.

What a review would have done: replace the image with one from a real project or label it clearly as an illustration, rewrite the line into something verifiable ("detailed quote after the technical visit"), and record who approved it. The change takes a few minutes and removes the biggest risk.

The same reasoning applies to proposals with savings simulations in sectors such as solar energy. If that is your case, read how to separate proposal, assumptions and after-sales support in practice, because a simulation without visible assumptions is exactly the kind of piece that leads to disputes.

Do I have to say the ad was made with AI?

Under the text of the CDC, the criterion is not the tool used but whether the consumer is misled. That means saying "made with AI" does not fix a misleading image, and not saying it does not make an ad illegal by itself. What matters is what the piece communicates.

That said, transparency is good practice in specific situations:

• When the image is illustrative and does not represent the real product, make that legible on the piece itself, not only in a hidden footer.
• When there is an AI-created character, avoid presenting it as a real customer or as a testimonial.
• When the piece involves an altered or generated image of a real person, check image-use authorization and, if personal data is involved, the obligations under LGPD, Brazil's data protection law. The national data protection authority (ANPD) publishes official guidance on data protection.

Ad platforms also have their own policies on synthetic content and misleading claims, and those policies change. Check each platform's current rules at the time of the campaign. Rules specific to AI in Brazil are still under discussion, so it is worth following what changes in Brazil's AI regulation and what companies should do now. This is a working hypothesis, not a certainty: the trend is toward higher expectations of transparency, so using a clear label when an image is illustrative costs little.

A review routine before you publish

An effective routine is short enough to be followed on busy days. This step-by-step works for most small and mid-sized companies.

• Step 1: record the origin. Note which tool generated the piece, which instruction was used and who requested it. This helps if you need to explain the piece later.
• Step 2: list every factual claim in the ad (price, deadline, delivery, result, warranty, what is included) and confirm each one with the person responsible for that area, not with the AI itself.
• Step 3: compare the image with the real product or service. If it does not match, replace the image or mark it visibly as an illustration.
• Step 4: look for absolute promises such as "always", "free", "guaranteed" or "immediate". Replace them with verifiable conditions.
• Step 5: check that the essential conditions are visible on the piece itself or behind a clear link: offer deadline, limits, extra costs, eligible audience.
• Step 6: verify testimonials, reviews and people. Use only real, authorized accounts. Generated characters should not appear as customers.
• Step 7: read the text the way a consumer would, in a hurry, on a phone. If a reasonable person could understand something different from what the company delivers, rewrite it.
• Step 8: named human approval. A responsible person marks the piece as approved, with a date, and keeps the final version.

For the copy itself, this checklist for reviewing AI-generated text before you send it complements the steps above with checks for facts, tone and personal data.

How to organize responsibilities and limits

The routine only works if someone owns it. In smaller companies it is common for marketing to create, the owner to approve at a glance, and nobody to check the commercial terms. The safest split is usually this: the person who creates the piece is not the one who approves the factual claims, and whoever knows the offer (sales or operations) confirms what is written.

It also helps to define in writing what AI can and cannot do in marketing. A short policy, like the one described in this guide to a company policy for employee ChatGPT use, can establish, for example, that AI drafts but does not publish on its own, that customer data stays out of the prompts, and that every numerical claim needs an internal source.

For those who want to think about risk in a more structured way, the NIST AI Risk Management Framework offers a public reference on how to map, measure and manage the risks of AI systems. It is not an advertising standard, but it helps organize the internal conversation.

The method we use in AI projects fits well here, without overdoing it:

• Map: identify where AI already enters ad creation, which tools are used, who uses them and which types of pieces come out of it.
• Limit: define what AI may generate (drafts, headline variations) and what requires confirmed data (price, deadline, result, testimonial).
• Pilot: apply the review routine to one type of piece for a few weeks, recording the corrections made and the cause of each one.
• Transfer: hand the routine to the team with clear owners, and decide whether to expand, adjust or stop based on what the record showed.

Decision guide: publish, adjust or block

When doubt appears at publishing time, a simple guide prevents improvised decisions.

• Publish if every claim was confirmed by someone who knows the offer, the image matches what will be delivered or is marked as an illustration, and the approval is recorded.
• Adjust if the piece has an absolute promise, a hidden condition, an idealized image or a simulation without assumptions. Fix it and return to review.
• Block if the piece uses a nonexistent testimonial or person as if real, shows a product or result the company does not deliver, or depends on data nobody can prove.

For campaigns already running, repeat the review whenever the price, stock, deadline or terms change. Reused old pieces are a common source of error, because the copy still looks good but is no longer true.

Frequently asked questions

Is the company responsible for an ad created by AI? Yes. As we read the CDC, whoever advertises answers for the advertising they run, regardless of the tool used to produce it. Saying "the AI got it wrong" does not usually remove responsibility from whoever published. That is why human approval before publication is the most important part of the routine.

Can an AI-generated image be used to show the product? It can, as long as it does not mislead the consumer about the characteristics of what they will receive. If the image is illustrative, say so clearly and legibly on the piece itself. For physical products, a real photo of the item is usually the safest option.

Is it mandatory to state that the ad was made with AI? We did not find a general requirement of that kind in the text of the CDC, and the central criterion remains whether the consumer is misled. Even so, labeling illustrative images and created characters is a prudent practice, and platform rules or new regulations may require it. For a specific case, consult a lawyer.

How do we start without slowing marketing down? Choose one type of piece, such as promotional ads, and apply the review steps for a few weeks, recording what was corrected. That record shows where the risk really is and lets you adjust the routine before extending it to other campaigns. If the review feels heavy, simplify the checklist instead of dropping it.

Sources

• Código de Defesa do Consumidor (Lei 8.078/1990), texto compilado
• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework