AI-Generated Ads: What Buyers Actually Notice, and How to Keep the Work from Looking Generic

The question of whether people notice AI-generated ads is often framed as detection: can a viewer tell a machine made it? In practice, the more useful question is what a viewer notices and how it affects their trust and willingness to click. This page is an analysis of the common failure patterns, not a report of a study. We do not have a verified dataset on how buyers respond to AI-generated ads, and we are not going to invent one. [NEEDS VERIFICATION: any published research on consumer response to AI-generated advertising before citing numbers.]

What marketers report, and what you can test in your own account, is that viewers react to specific symptoms rather than to the production method. Those symptoms include sameness (the ad looks like many others), product inaccuracy (the item shown is not quite the item sold), garbled or inconsistent text in the image, faces and hands that look slightly wrong, and copy that makes sweeping promises without evidence. Each of these can appear in human-made ads too; AI tends to make them cheaper to produce in volume.

Infinall AI generates campaign drafts from a product URL: research, strategy, ad copy, and creative such as image ads, carousels, and video drafts. They land in an Approval Queue where you review and edit before export. The tool does not guarantee that an ad will perform or that viewers will find it authentic, and it does not buy media or place ads. The review step exists because these symptoms are visible to you before they are visible to your customers.

The rest of this analysis takes each symptom in turn, explains why it happens, and says what to do about it.

The most common problem with AI creative is convergence. Image models trained on similar data, given similar prompts, produce similar results: the same soft lighting, the same composition, the same gradient background, the same cheerful tagline structure. When many advertisers in a category use the same tools with minimal direction, their ads start to blur together, and a viewer scrolling past learns to ignore the whole look.

This is a structural risk rather than a flaw you can see in one ad. It shows up when you compare your ad with the competitors' in the same feed. That is why research matters before generation. In Infinall, Library includes Ad Spy for browsing real competitor ads. Looking at what is already running tells you which visual and verbal conventions are saturated, so you can deliberately depart from them.

Practical countermeasures:

• Start from something specific to your product: real photography, real screenshots, real customer language, a distinct brand colour. Distinctiveness comes from inputs that only you have.
• Give the tool a point of view. Generic direction yields generic output; specific direction about the audience, the objection, and the tone yields more particular results.
• Use editing tools to adjust colour grade, lighting, and composition so a draft stops looking like a default.
• Vary format as well as image: a carousel that makes an argument step by step feels different from a single hero image.

Sameness is also a copy issue. Headlines built from the same templates ("Discover the secret to...") are as recognisable as stock imagery. Reading your copy aloud next to a competitor's is a quick test of whether it has any voice.

Where sameness is a soft risk, product inaccuracy is a hard one. Generated or edited imagery can alter the item: the wrong number of buttons, a logo that is slightly off, text on packaging that is gibberish, a colour that does not match, or a feature that does not exist. A viewer who clicks through and receives something different from the ad does not blame the tool; they blame you, and they ask for a refund or leave a review.

Visible artifacts matter too. Common ones include warped lettering inside images, hands and fingers with the wrong shape, reflections that do not make sense, and objects that merge into each other. These are less frequent than they were in earlier generations of models, but they have not disappeared, and they are the details people comment on.

Review habits that help:

• Compare every product image to the real product, zoomed in. Check logos, labels, ports, controls, and proportions.
• Prefer putting text (headlines, prices, calls to action) in the layout layer rather than relying on the image model to draw it, and proofread every word.
• For app or software ads, use real screenshots or recordings, and make sure they show the current version.
• Look at people in the image closely, including hands, eyes, and teeth, and replace or crop what looks wrong.

Infinall's image editing tools (erase, resize, enhance, colour grade, remove background, style transfer, lighting) exist for this fix-up work. The Approval Queue is where you decide what ships. None of it replaces your own check, since the tool does not know what your real product looks like beyond what you supply.

Copy is where AI-assisted advertising can create the most serious problems, because language models generate fluent, confident statements whether or not they are true. A draft might say a product is "clinically proven," "the number one choice," or "guaranteed to save you hours," none of which the model has any basis for.

Buyers notice this in two ways. Experienced buyers distrust superlatives and vague promises on sight. And platforms and regulators notice claims that are unsubstantiated: health, financial, income, and performance claims are restricted on major ad platforms and in many jurisdictions. [NEEDS VERIFICATION: current Meta, Google, and TikTok ad policies on health, income, and performance claims.] Infinall does not guarantee compliance and does not replace legal review.

A claim checklist before approval:

• Is every factual statement something you can document today? If not, remove it or soften it to what you can show.
• Are numbers (discounts, trial lengths, user counts, delivery times) accurate and current? Enter these by hand.
• Are testimonials and ratings real, current, and used with permission?
• Does the ad promise something the landing page and the product deliver?

Tone matters as well. Copy that feels written for no one in particular reads as machine-made even when it is grammatical. Specific details such as the actual use case, the real objection, and the customer's own phrasing do more to make copy feel human than any stylistic tweak. If you have reviews or support messages, mine them for language and use it.

Whether you need to disclose that an ad was AI-generated depends on the platform, the content, and the jurisdiction. Some platforms have introduced labels or disclosure requirements for AI-generated or digitally altered media, particularly for realistic people and events, and rules continue to change. [NEEDS VERIFICATION: current AI-content disclosure rules for Meta, Google, and TikTok ads, and any advertising-disclosure rules in your markets.] Check the current policy of each platform where you run ads, rather than relying on a guide, including this one. Approval inside Infinall is not approval by Meta, Google, or TikTok.

Beyond formal rules, there is the question of honesty. Showing a synthetic person giving a testimonial, or a synthetic scene presented as a real customer result, is the clearest way to lose trust and invite complaints. Illustrative imagery that is plainly a creative treatment is different from fabricated evidence. When in doubt, use real customers, real screenshots, and real results.

A short pre-publish checklist:

1. Does the product shown match the real product, zoomed in?
2. Is all on-image text proofread, and is any text in the image intentional?
3. Is every claim documentable, and is every number current?
4. Are any people, testimonials, or results real, or clearly labelled as illustrative?
5. Have you checked the current platform policy for this ad type?
6. Does it look different from the three nearest competitor ads in the feed?
7. Does the landing page deliver on the promise?

Build that checklist into your Approval Queue routine. AI makes drafts cheap, which makes review the scarce resource. Teams that treat review as the main job, and generation as the fast part, tend to publish work that holds up when a real customer looks closely.