AI Generated Ads Have No Taste. Here Is How to Give It Some

AI generated ads have no taste because taste is not a rendering problem. A model can now draw your bottle perfectly, hold the label steady across fifty frames and light it better than most studios, and still have no opinion about whether the headline should be a warning, a joke or a price. It makes. It does not choose.

So the fix is not a longer prompt. Another thirty words will not hand a model judgement it does not have. The fix is to stop asking it to choose, and to give it choices a person already made. That sounds like a dodge. It is closer to how good advertising has always worked, and below I take a real ad apart to show what those choices look like written down.

The clearest evidence is not in my ad account, it is on television. Two of the largest advertisers in the world put real money behind AI holiday campaigns and both landed badly. Neither failed on rendering.

The Nielsen Norman Group wrote up those campaigns and landed on something I keep seeing in much smaller accounts: audiences can tell when a story has been shaped around what the technology can do rather than what the ad needed to say. McDonald’s Netherlands did not have a rendering problem. It ran a Christmas ad telling people Christmas is miserable, then withdrew it four days later. That was a decision, and no model made it. No model would have caught it either.

This is the part everybody gestures at and nobody writes down. So here is a real ad I pulled out of the AIMS library, from Terrai Beauty, for a serum called One and Done. Read it as a list of decisions rather than as a picture, because that is what it is.

For contrast, here is what came back when I asked an image model for a skincare serum ad and gave it nothing else. No reference, no decisions, which is exactly how most of these get made. This one is generated on purpose and it is meant to be bad.

Nothing in it is wrong. The bottle is rendered cleanly, the light is even, the type is legible. It is a perfectly executed nothing, and that is the problem. It cost me about two cents, which means everyone selling what you sell can make forty of them before lunch.

There is an older framing for this and it is still the best one I know. Dave Trott drew a triangle on a flipchart and said every ad needs three things: impact, communication, persuasion. Get noticed, tell me what, tell me why. I found this version of it in the AIMS library, in a post by Harry Dry, and it is the sharpest picture of the AI creative problem I have seen despite having nothing to do with AI.

I would rather you heard the limits from me than found them on a Tuesday with a launch booked.

• Net new concepts. Distilling a great ad gets you a great execution. It will not invent the idea that makes your category look at you differently, and I have not seen anything on the market that does.
• Offer led wins. If an ad worked because the offer was strong, copying the layout copies the wrong half. Work out what was doing the work before you rebuild it.
• Anything resting on a real person. A founder on camera, a customer holding the thing. A reference cannot lend you credibility you have not earned.
• Brand defining campaigns. The piece that sets how you look for the next three years deserves a person and a budget. This is for the forty ads you need next month.

AIMS is a free searchable library of real ads, emails, videos and pins, indexed across Meta, Instagram, TikTok, Pinterest, Amazon and Etsy, which is where both the Terrai ad and the Harry Dry post above came from. Discover is the browsing half and the free Chrome extension saves anything you run into while you are out on the internet. The part that matters for this article is what happens after you save something: AIMS does the decision writing step itself, reads what the designer actually did, and passes that analysis to the model alongside your product photography and your palette. The teardown page walks through the loop. What comes back is your product, in that structure, with none of their assets in it.