Why AI game ads look generated, and how to fix it
Looking like AI is what costs AI-generated ads their clicks. The five tells that give a game ad away, the research behind them, and the production decisions that remove each one.
People spot AI in ads, and being spotted is what costs the clicks. In a working paper on 4,633 sibling ads from 202 brands, Exner and colleagues at Columbia and Harvard Business School found no average click-through penalty for AI images (MediaCat). Split by how each image read, the result changed. AI that passed as human-made had the best click-through rate of the four groups. AI that read as AI fell below human work. So the useful question for a game studio isn't whether to use AI in ads. It's whether the finished ad shows it. The tells are specific, and each one has a production fix.
Do AI-generated ads perform worse?#
Not on clicks, in a 4,633-ad field study. On Kantar's pre-tests they score lower. The studies that seem to disagree mostly measure different things.
Start with the result that cuts against the worry. Lee, Todri, Adamopoulos and Ghose at NYU Stern tested product ads on the Google Display Network. Fully AI-generated ads raised click-through rate "by up to 19%" over human-made ones, while human ads modified by AI "showed no significant improvement" (NYU Stern). These were product ads, not game ads. Still, it's a fair challenge to anyone who says AI creative is worse by default.
Kantar's data points the other way. GenAI-created ads in its LINK database average the 54th percentile on overall impact, against the 65th for non-AI ads (Kantar). But ads where the AI was used unobtrusively did better, with over 40% in Kantar's top tier for branded cut-through. Kantar calls AI "obvious" when the average viewer would notice it, and its GenAI ads also scored lower on branding (Kantar).
The Exner study suggests a way to reconcile them (SSRN). It sorted every image by its real origin and by how it read.
| AI that looks human-made | 0.79% |
|---|---|
| Human-made | 0.67% |
| AI identified as AI | 0.62% |
| Human-made mistaken for AI | 0.55% |
Look at the lowest column. Human-made images that read as AI did worst of all, and 24.87% of human images were misread as AI. The penalty follows the look, whoever made the picture. Advertisers optimising for app installs also adopted generative AI significantly more than others in the sample.
Your audience is already looking for it#
Most US adults say they've seen AI ads recently, and many don't like it. Gallup asked 3,270 US adults in May 2026 how they view businesses using AI in advertising (Gallup).
| Negative | 49% |
|---|---|
| Neutral | 32% |
| Positive | 19% |
79% of the same adults said they had seen AI ads in the past 30 days, and 62% called AI-created people or voices unacceptable. Among 18 to 29 year olds, 66% were negative. The IAB, surveying 505 consumers aged 16 to 43, found that 71% think they've seen an AI ad, up from 54% in 2024. It also found 82% of ad executives believe Gen Z and Millennials feel positive about AI ads, against 45% who actually do (IAB).
Spotting doesn't need a trained eye. In a NielsenIQ study of more than 2,000 people, about 150 of them on EEG, consumers "intuitively identified most of the AI-generated ads" and found them annoying, boring and confusing. Even high-quality AI ads produced weaker memory activation (NielsenIQ).
Game makers are sceptical too. In GDC's 2026 survey, 52% of game professionals said generative AI is having a negative impact, up from 30% in 2025 and 18% in 2024. Only 7% called it positive (Game Developer). Players are warier still: Quantic Foundry found 85% of gamers negative about generative AI in games, though its sample leans towards core PC players (GamesMarket). These figures come from different questions, groups and years, so read them as a direction, not one number.
Players check, too. When Activision announced Guitar Hero Mobile with an AI-generated image on Instagram, players picked out arms blending into guitars, buttons drawn as "nondescript blobs" and drums as "a mess of sticks and floating cymbals". The image was hidden (TheGamer). Wizards of the Coast denied, then admitted, that the background of a Magic: The Gathering marketing image had AI components (Vice).
The five tells that give a game ad away#
Each tell is either a known limit of image models or a habit of how generated ads get assembled.
Text rendered inside the image#
Image models still get lettering wrong. Diffusion models "continue to struggle to generate consistent and legible text within images", according to the STRICT paper at EMNLP 2025 (arXiv). Google listed long-form text rendering among the things it was still working on when it launched Gemini 2.5 Flash Image (Google Developers Blog).
The failure is often subtle: a word that's almost right, or a button label in a font the game never used. Advertisers told Business Insider that cleaning up after Meta's AI tools had become routine, garbled text included (The Next Web). And a word baked into a flat image can't be retyped. You regenerate or edit the pixels, and the rest of the image can shift with it.
A hero who drifts off-model#
Consistency is the other open problem. Google puts it plainly: "A fundamental challenge in image generation is maintaining the appearance of a character or object across multiple prompts and edits" (Google Developers Blog). An opinion piece in The Drum on Coca-Cola's 2025 AI holiday ad counted the cost: "Only two of the 11 truck scenes feature the same truck" (The Drum).
In a game ad, the character is the brand. Ehrenberg-Bass says distinctive assets need fame and uniqueness, and adds: "The more attention-grabbing the creative, the harder the branding has to work" (Ehrenberg-Bass Institute). A hero whose armour changes between the 1:1 and the 9:16 is spending that branding. The same Coke ad shows the flip side. System1 scored it 5.9 Stars and credited "the same iconic assets" Coke has built over decades (System1).
A scene from a different game#
In our experience, a generated backdrop drifts towards a generic genre scene: the epic castle, the neon city. If your game is flat and cartoonish, a painterly vista reads as someone else's game. Your players know your art better than anyone.
There's a compliance angle too. The UK's ASA says game ads must represent the core gameplay experience, and that a line like "Not all images represent actual gameplay" is not enough (ASA). The ASA wasn't ruling on AI. Our reading is that a generated scene the game doesn't contain carries the same risk.
The sheen#
In our experience, many generated images share one finish: soft rim light, glossy surfaces, smooth gradients, airbrushed faces. Game art is a set of deliberate choices, from cel shading to pixel art to hand-painted. A generic gloss over one of them is the kind of thing an average viewer notices, which is where Kantar draws the line for "obvious".
Things drawn that shouldn't exist#
Extra limbs, props melting into each other, objects that couldn't be built. Guitar Hero's arms merging into guitars is the game case. Outside games, the advertisers Business Insider spoke to included REI, whose generated image showed a bike with two handlebars. Meta's response: "it is the advertiser's responsibility to review the AI outputs" (The Next Web).
When we built our ad composer, which lays ads out in HTML and CSS, we met the same failure in another medium. A language model composing an ad tries to draw things out of boxes. We once got the Colosseum built from twenty stacked arch divs, and a red rectangle placed behind a gladiator who already had a cape. So the composer has a rule. Markup may carry flat fields, gradients, frames, panels, buttons and text. Anything depicted (architecture, characters, props, symbols) must be real artwork, or it doesn't appear.
Fix the tells in production#
Most tells trace back to one decision: something was generated that didn't need to be. So the fixes are about what a model makes at all.
| The tell | What the viewer sees | The production fix |
|---|---|---|
| Text in the image | Garbled or almost-right words, off-brand lettering | Headlines, prices and buttons as live text layers |
| Off-model hero | The character changes between frames and sizes | The game's own character art; a missing pose generated with it as the reference |
| A scene from another game | A generic backdrop in a style the game doesn't have | The game's backgrounds first; one missing scene generated from them |
| The sheen | One glossy finish on everything | The game's art sets the finish; an artist signs off |
| Things that shouldn't exist | Extra limbs, props melting together | Depict only what you have art for; a person checks every ad |
Start from the game's own art#
Your characters, backgrounds, logo and symbols already exist, on-model and in style, and your players know them. They are your distinctive assets. In our pipeline, the game's art is read from its store page or the team's uploads, and cut-outs with real transparency are used as they are.
The vendor evidence is consistent with this. Meta reports 11% higher CTR and 7.6% higher conversion rates for campaigns using its generative ad features, a correlational comparison with no published method (Meta). The features it describes build on the advertiser's own image and copy rather than inventing the ad.
Keep text as live text#
Headlines, prices and buttons belong in text layers, never in pixels inside a generated image. In our pipeline they stay live text, so a wrong word is typed, not regenerated. A new headline, a translation or a legal line is an edit to one layer.
- background
- props
- character
- logoClashRoyale
- headline · copy
Clash RoyaleStrategy · Free3 MINUTES.ONE CROWN.- Real-time duels
- 100+ cards
- Clan wars
- button
- the adClashRoyaleClash RoyaleStrategy · Free3 MINUTES.ONE CROWN.
- Real-time duels
- 100+ cards
- Clan wars
Play free→
That's how Vidscape builds ads: from the game's own art, as editable layers with live text, so art direction stays with your team. Our page for creative directors covers how that fits a studio's review.
Generate only what's missing#
Most ads need one or two things the game doesn't have: a scene for a seasonal event, a pose the character was never drawn in. Generate those and nothing else. In our pipeline, generation fills only the gaps, anchored to the game's art as reference images, so a new pose still looks like your character.
A background scene is generated at the ad's own aspect, as a full scene, rather than stretched from another size. A stretched scene is its own small tell. Our post on mobile game ad sizes covers why each size needs its own layout.
One more lesson from building this. When we turned flat generated images into editable documents, isolating each element from the source image worked better than re-imagining it from a text description. In our tests, a description dropped the details that make an element on-model. The picture kept them.
Put a person between generation and launch#
Meta's line about reviewing AI outputs holds on every network. Before a batch goes live, someone who knows the game should:
- Read every word at full size, small print included.
- Compare the hero with the character's reference art.
- Look where hands, weapons and props meet. That's where things melt.
- Put the ad next to real gameplay and ask whether a player would recognise the game.
AI can do the production. The creative call stays with your team.
Know the label rules, and what disclosure costs#
Meta, TikTok and Google each have a rule for AI in ads:
- Meta labels ads significantly edited by its own generative tools and auto-labels third-party AI content through industry signals, under "AI info" (Meta).
- TikTok requires the AIGC label or a clear disclaimer on AI-generated or significantly edited ad content. Background removal and colour changes are exempt, and undisclosed AI content "will be rejected or restricted" (TikTok Ads Help).
- Google Ads permits AI labels in creatives and added an AI label setting from July 2026 (Google Ads Policy Help).
A label has a price. In the NYU Stern experiment, disclosing AI use cut click-through rate by 31.5%. Across three studies of prosocial ads, Baek, Kim and Kim found disclosure lowered attitudes towards the ad through perceived credibility (International Journal of Advertising). Hiding it isn't an option either, given TikTok's enforcement and the Wizards of the Coast episode. What you control is how much of the ad is generated. Check each network's rule against what you actually generated.
What to change before your next batch#
Audit your last ten AI-assisted ads against the five tells, and note which decision produced each one. Then move those decisions upstream: build from your own art, keep words as text, generate only the gaps, and keep a person at the end.
For where generation fits in game marketing more widely, see our page on AI-generated marketing and the guide to mobile game ads.
Frequently asked questions#
Do AI-generated ads perform worse than human-made ads?
Not on click-through, on average. In a working paper on 4,633 sibling ads from 202 brands, Exner et al. found no average CTR disadvantage for AI images, though Kantar's pre-tests put GenAI ads at the 54th percentile against the 65th for non-AI ads. AI that looked human-made had a 0.79% CTR, while AI that read as AI had 0.62%, below human-made work at 0.67%. An NYU Stern experiment on product display ads found fully AI-generated ads raised CTR by up to 19%.
What makes an AI game ad look generated?
Five things, mostly: text rendered inside the image, a character who drifts off-model between frames or sizes, a scene that doesn't match the game's art style, a uniform glossy finish, and things drawn that shouldn't exist, such as extra limbs or props melting together.
Do I have to label AI-generated game ads?
Often, yes. TikTok requires the AIGC label or a clear disclaimer on AI-generated or significantly edited ad content, and rejects or restricts undisclosed AI content. Meta labels ads edited by its generative tools and auto-labels third-party AI content. Google Ads has an AI label setting from July 2026. Check each rule against what you generated.
- Disguised AI ads outperform human ones (Exner, Hartmann, Ding, Zhang & Netzer) — MediaCat · May 2026
- AI in Disguise (working paper) — SSRN
- The AI advertising paradox (Lee, Todri, Adamopoulos & Ghose) — NYU Stern · Nov 2025
- The state of GenAI in media and advertising — Kantar · Mar 2026
- Rethinking AI-generated advertising — Kantar · Nov 2025
- Americans aren't sold on businesses using AI in advertising — Gallup · Aug 2026
- The AI gap widens — IAB · Jan 2026
- NIQ research uncovers hidden consumer attitudes toward AI-generated ads — NielsenIQ · Dec 2024
- One third of game workers use generative AI, but half think it's bad for the industry — Game Developer · Feb 2026
- Quantic Foundry: most gamers are negative about the use of gen AI in video games — GamesMarket · Dec 2025
- Guitar Hero Mobile's AI-generated announcement art on Activision's Instagram — TheGamer · Mar 2025
- Magic: The Gathering publisher denies, then admits, using AI art in promo image — Vice · Jan 2024
- Brands cleaning up after Meta's AI ad tools (reporting Business Insider) — The Next Web · Jul 2026
- STRICT: text rendering in image generation (Zhang et al., EMNLP 2025) — arXiv · 2025
- Introducing Gemini 2.5 Flash Image — Google Developers Blog · Aug 2025
- The truck is different in every shot of Coca-Cola's AI Christmas ad — The Drum · Nov 2025
- AI or no AI, Coke gets the Christmas love — System1 · Nov 2025
- Brands need distinctive assets — Ehrenberg-Bass Institute
- Misleading gameplay in videogame ads — ASA / CAP · Aug 2024
- Meta's AI product news from Connect — Meta · Sep 2024
- Gen AI transparency in Meta's ads products — Meta · Feb 2025
- Ads policy: misleading and false content — TikTok Ads Help · Apr 2026
- AI labels in Google Ads creatives — Google Ads Policy Help · Jul 2026
- Effect of disclosing AI-generated content on prosocial advertising (Baek, Kim & Kim) — International Journal of Advertising · 2026
