# AI Image Ads That Don't Look AI: A Prompting Guide for Performance Creative > Why AI images in ads feel fake, and how to fix it: a seven-part prompt, a two-minute quality check, Meta's AI labelling rules and a test to prove what works, backed by research on AI faces and ads. - **Author**: Marxx - **Published**: 2026-10-24 - **Category**: Production - **URL**: https://marxx.ai/posts/ai-image-ads-look-real --- **How do you make AI images for ads look real?** Replace the model's defaults with specifics: a real person doing a real action, in an ordinary, slightly messy setting, with one light source, a phone-style camera and natural texture. Leave text and logos out and add them yourself. Then check hands, product, text and physics before publishing. The aim isn't to hide AI. It's to remove what distracts. ## The paradox: AI faces fool people. AI ads often don't. Two findings seem to contradict each other. In a 2022 study in *PNAS*, Sophie Nightingale and Hany Farid showed people a mix of real faces and AI-generated ones. Participants identified which was which with **48.2%** accuracy, slightly worse than a coin flip. They also rated the synthetic faces as slightly *more* trustworthy ([Scientific American](https://www.scientificamerican.com/article/humans-find-ai-generated-faces-more-trustworthy-than-the-real-thing)). Yet in December 2024, NielsenIQ found that consumers **intuitively identified most AI-generated ads** without being told. They described them as "annoying", "boring" and "confusing". EEG measurements showed **weaker memory activation** than for traditional ads, even for AI ads viewers rated as high quality ([NIQ, 2024](https://nielseniq.com/global/en/news-center/2024/niq-research-uncovers-hidden-consumer-attitudes-toward-ai-generated-ads/)). The AI image paradox: people identify AI faces with 48.2% accuracy, below chance, yet consumers picked out most AI ads and remembered them less **How both can be true:** A single, carefully selected face is a narrow test. An ad is a whole scene: hands, a product, a room, light, text, and a story that has to make sense. Each element is another chance for something to feel slightly wrong. **The face can pass while the ad fails.** That's the problem this guide solves. It explains why AI images feel "off", gives you a prompt structure that replaces the model's defaults, a two-minute quality check, the rules on disclosure, and a way to test whether it's working in your account. Six tells of AI images and the fix for each: skin, light, hands, text, background, camera --- ## Part 1: Why AI images feel "off" Why AI images feel off: they are too typical, they contain small errors in hands, text and logic, and low-quality visuals take more effort to process ### 1. They're too typical Image models learn from vast numbers of photos. Asked for "a woman holding skincare, studio lighting, high quality", a model tends to produce something close to the *typical* version of that scene: even light, flawless skin, an empty background. Real photos are the opposite. They're **specific**: one light source, a stray towel, a slightly crooked frame. When an image has none of those specifics, viewers may not know why it feels staged, but they feel it. ### 2. They contain small errors Hands with an extra finger, labels whose letters almost spell something, a reflection that doesn't match. These are the classic giveaways. They're improving with every model generation, but they still appear, especially in hands holding products and in text. ### 3. They cost attention NIQ's study found that low-quality visuals required **more cognitive effort** to process, which distracted from the message. That links to a broader finding in perception research: people tend to prefer what's easy to process (see [visual ad creative principles](/blog/visual-ad-creative-principles)). An image that makes viewers think "is this real?" is spending attention on the wrong question. > **The quotable version:** AI images don't look fake because they're bad. They look fake because they're average. --- ## Part 2: The six tells, and how to fix each | # | The tell | Why it happens | The prompt fix | The check | |---|---|---|---|---| | 1 | **Skin that's too smooth** | Models favour flawless, "beautiful" defaults | "Natural skin texture, visible pores, shot on a phone" | Zoom to 100% | | 2 | **Light that's too perfect** | Studio glow from everywhere is the average | Name one source: "morning window light from the left, soft shadows" | Do shadows fall one way? | | 3 | **Hands that don't work** | Hands are complex and vary hugely | Describe the action: "holding the bottle by the neck, label facing the camera" | Count fingers; check the grip | | 4 | **Garbled text** | Models draw letter *shapes*, not spelling | Don't ask for text: "no text, no logo" | Delete and replace all text | | 5 | **A background that's too tidy** | An empty room is the typical room | Specific clutter: "a towel on the rail, a mug on the counter, a slightly messy shelf" | Would someone live here? | | 6 | **A camera that's too good** | Catalogue framing is the default | "Shot on a phone, slightly off-centre, natural framing" | Does it look like someone took it? | **When the fixes go too far:** "imperfect" isn't the same as "bad". A blurry, badly lit image fails for a different reason. Aim for **specific and believable**, not deliberately poor. --- ## Part 3: The anatomy of a realistic prompt Every vague word in a prompt gets filled with the model's default. So the method is simple: **replace each default with a specific.** The anatomy of a realistic prompt: subject, action, setting, light, camera, texture, and what to leave out | Part | Default you're replacing | Specific example | |---|---|---| | **Subject** | "Beautiful model" | "Woman in her 30s" | | **Action** | Posing | "Holding the bottle by the neck, label towards the camera" | | **Setting** | Empty studio | "Small bathroom, towel on the rail, mug on the sink" | | **Light** | Even studio glow | "Morning window light from the left, soft shadows" | | **Camera** | Perfect catalogue lens | "Shot on a phone, slightly off-centre" | | **Texture** | Airbrushed | "Natural skin texture, slight grain" | | **Leave out** | Model-written text | "No text, no logo" (add them in your design tool) | ### Before and after The same prompt before and after: a generic studio prompt versus a specific phone-style scene **Before:** > Beautiful woman holding a skincare bottle, studio lighting, high quality, 8k **After:** > Woman in her 30s in a small bathroom, holding a skincare bottle by the neck near her face, label towards the camera, morning window light from the left, soft shadows, natural skin texture, shot on a phone, slightly off-centre, towel on the rail and a mug on the sink, no text, no logo **Why the second works:** "Beautiful", "studio", "high quality" and "8k" all push towards the polished average. Every phrase that replaced them is a specific detail a real photo would have. --- ## Part 4: Check before you publish Generate several versions and pick the best. Then run every image through a two-minute pass. The two-minute QA pass: hands, product, text, physics, phone size, side by side 1. **Hands:** Count the fingers. Does the grip make physical sense? 2. **Product:** Is the shape, colour and label *exactly* right? Product accuracy matters more than anything else in the image. 3. **Text:** Delete anything the model wrote and add real text yourself. 4. **Physics:** Do shadows, reflections and liquids behave? 5. **Phone size:** Shrink it. Does it still read as a photo? 6. **Side by side:** Put it next to a real photo from your brand. Does one look obviously different? ### Worked example: Fixing a skincare image *Hypothetical example.* **First attempt** (default prompt): a flawless face, studio glow, a bottle with a melted-looking label, and a blank grey background. It looks like a stock photo, and the label is unreadable. **Diagnosis against the six tells:** Skin (1), light (2), text (4), background (5) and camera (6) all fail. **Second attempt** (seven-part prompt, "no text"): a bathroom scene in window light. The label area is blank. Of eight generations, two have a hand gripping the bottle oddly. Pick one of the others. **Finishing:** Replace the blank label with the real product label in a design tool, add the headline, then run the QA pass. **What to test:** This version against your existing real product photo, with the same offer and budget (Part 6). --- ## Part 5: Looking real isn't the same as hiding AI It's tempting to treat "make AI look real" as "make sure nobody can tell". That's the wrong goal, for practical as well as ethical reasons. Looking real isn't hiding AI: Meta labels ads with significant AI edits or photorealistic AI people, says it will detect third-party AI tools, and shows AI info in About this ad **Meta labels AI-made ads.** Under its 2025 approach, ads created or significantly edited with Meta's own generative AI tools, or with an AI-generated photorealistic person added, get an "AI info" label. In a June 2026 update, Meta said it will automatically detect ads made with third-party AI tools using industry-standard signals and label them, with "AI info" shown in "About this ad" ([Meta, 2025, updated 2026](https://about.fb.com/news/2025/02/gen-ai-transparency-metas-ads-products/)). **So the realistic goal is:** **real enough not to distract, honest enough to label.** Realism removes the friction NIQ measured. It shouldn't be used to pass off something false as real. **Where AI images shouldn't be used at all:** - **Customer results.** Before-and-after images should show real customers. - **Testimonials.** Don't put an AI face next to a real review. - **The product, when accuracy matters.** Colour, size and texture should come from real photos. AI images work best for **scenes, backgrounds, lifestyle settings and variations around a real product shot**. Our [traffic light guide](/blog/are-ai-generated-ads-effective) covers where AI is green, amber and red. --- ## Part 6: Prove it in your account The research tells you what to watch for. Only a test tells you what works for *your* audience. Prove it in your account: test a real photo, an AI image with a default prompt, and an AI image with a realistic prompt, with the same idea and budget **The test:** 1. **Three versions of the same concept:** Your real photo, an AI image from a default prompt, and an AI image from the seven-part prompt plus QA. 2. **Same offer, copy, audience and budget.** 3. **Run until each has enough results to compare,** usually one to two weeks. | Metric that matters | Why | |---|---| | Cost per result | The outcome | | Early retention (for video), or click-through *with* conversion rate (for statics) | Shows whether the first glance worked | | Comments asking "is this AI?" | A direct signal that realism is distracting | | Metric that misleads | Why | |---|---| | Your team's opinion of the image | You know it's AI; your audience doesn't start there | | Likes alone | Engagement isn't purchase | | Number of images generated | Volume isn't quality | **How to read the result:** If the realistic AI version matches the real photo on cost per result, AI is a cheap way to make more variations of that concept. If it loses clearly, use AI for backgrounds and scenes, and keep real photos for the hero shot. --- ## Six mistakes, and how to spot them | Mistake | Symptom | Fix | |---|---|---| | "Beautiful, studio, 8k" prompts | Stock-photo look | Replace every default with a specific | | Asking the model for text | Garbled labels | "No text"; add text yourself | | Publishing the first generation | Odd hands, wrong product | Generate several; run the QA pass | | Going too imperfect | Blurry, badly lit images | Specific and believable, not poor | | AI faces for testimonials | Trust risk, policy risk | Real customers only | | Assuming realism hides AI | Surprise "AI info" labels | Real enough not to distract, honest enough to label | ## What to take into your next creative review - **A single AI face can pass; a whole AI ad often doesn't.** NIQ found consumers spotted most AI ads and remembered them less. - **AI images look fake because they're average.** Replace every default with a specific. - **Never let the model write your text.** - **Check hands, product, text and physics** before anything runs. - **Aim for credible, not concealed.** Meta labels AI-made ads. ## FAQ **How do I make AI images look real for ads?** Use a specific prompt: a real person, a real action, an ordinary setting, one light source, a phone-style camera and natural texture. Leave text out, then check hands, product, text and physics before publishing. **Can people tell if an ad is AI-generated?** Often, yes. In NIQ's 2024 study, consumers identified most AI-generated ads without being told, even though single AI faces can be hard to tell from real ones. **Do AI image ads perform worse?** They can, when they look fake or distract. NIQ found weaker memory activation for AI ads. Test a realistic AI version against your real photo with the same budget to find out for your audience. **Will Meta label my AI images?** Meta labels ads significantly edited with its own AI tools, or with photorealistic AI people, and has said it will detect and label ads made with third-party AI tools. **Should I add text with AI?** No. Generate the image without text and add it in a design tool. --- ## Make AI visuals around a proven idea **Trial the Marxx generation canvas:** Start from what's working in your category, and get visuals, voice, talent and every placement from one brief, so you're testing variations of a proven idea rather than generating images at random. Marxx talent node with AI avatars and UGC-style creators Marxx video node assembling an ad **[Trial the canvas: Turn one brief into a finished ad ->](https://www.marxx.ai)** *Related: [Should You Put AI-Generated Ads Into Your Account?](/blog/are-ai-generated-ads-effective) - [UGC Video Generators Compared](/blog/ugc-video-generators) - [Visual Ad Creative Principles](/blog/visual-ad-creative-principles)* ### Sources - Nightingale, S. & Farid, H. (2022). AI-synthesized faces are indistinguishable from real faces and more trustworthy. *PNAS*. [Scientific American summary](https://www.scientificamerican.com/article/humans-find-ai-generated-faces-more-trustworthy-than-the-real-thing). - NielsenIQ (2024). [NIQ research uncovers hidden consumer attitudes toward AI-generated ads](https://nielseniq.com/global/en/news-center/2024/niq-research-uncovers-hidden-consumer-attitudes-toward-ai-generated-ads/). - Meta (2025, updated 2026). [Expanding GenAI transparency for Meta's ads products](https://about.fb.com/news/2025/02/gen-ai-transparency-metas-ads-products/). --- - [More Production articles](https://marxx.ai/posts/category/production) - [All articles](https://marxx.ai/posts)