Looking artificial
What people think they are seeing matters more than what made it: AI ads judged human-made had the highest click-through, and human ads that looked AI-generated the lowest.
Do AI-generated ads perform worse than human-made ones?
Not on live click data, but yes on brand-lift tests. A Columbia-led analysis of more than 300,000 ads and 500 million impressions found AI creatives averaged a 0.76 percent click-through rate against 0.65 percent for human-made ads. An Ipsos and Syracuse study of 20 TV spots found human-made ads 14 percent stronger on short-term sales effect and 17 percent stronger on brand equity. Two findings repeat across studies: ads that look artificial underperform, and labeling an ad as AI-made cut click-through by 31.5 percent.
| Study | Sample | Measure | Result |
|---|---|---|---|
| Columbia-led, AI in Disguise | 300,000+ ads, 500 million impressions | Click-through rate | AI 0.76 percent, human 0.65 percent, no drop in downstream conversion |
| Columbia-led, perception test | 4,633 ads from 202 brands | Click-through by how the ad looks | AI judged human 0.79 percent, human 0.67, AI recognized 0.62, human that looked AI 0.55 |
| Ipsos and Syracuse, short term | 20 TV spots, 3,000 US consumers | Creative Effect Index | Human-made 14 percent stronger |
| Ipsos and Syracuse, long term | 20 TV spots, 3,000 US consumers | Equity Effect Index | Human-made 17 percent stronger |
| NYU Stern, AI-made ads | AI-made, AI-modified and expert-made ads | Click-through rate | Fully AI-made up to 19 percent higher; AI-modified human ads no significant gain |
| NYU Stern, AI label | Ads disclosed as made with generative AI | Click-through rate | 31.5 percent lower when consumers were told |
What drives the answer
Functional or emotional
Ipsos found AI handled functional, product-driven briefs well and struggled with emotional, narrative ones. Most small-business advertising is direct response: a product, an offer, a click.
Volume and iteration
On direct response, volume and iteration beat a single polished asset. The failure mode is generic output: warped product renders and stock-looking people.
Disclosure
Telling consumers an ad was made with generative AI cut click-through by 31.5 percent in the NYU Stern study, so read the label rules for each platform you run on.
The honest answer is that it depends on what you measure. Live campaign data from display and performance ads says AI creative is at least as good. Controlled brand-lift tests on TV spots say human-made ads still win. Both are true, and the gap between them tells you where to use each.
The live data comes from a Columbia Business School-led team with co-authors at Harvard, TU Munich and Carnegie Mellon, using ads served on Taboola's Realize platform. The design paired each AI ad with a human-made sibling from the same advertiser, campaign, day and landing page, and in those matched comparisons AI performed the same or better. The brand-lift data comes from Ipsos and Syracuse University, who took ten human-made 30 second spots from major brands, had Google's Gemini reverse-engineer each brief into a shot list, and generated AI counterparts with OpenAI's Sora 2 with no human creative intervention. In all, 3,000 US consumers saw the 20 ads through Ipsos' sales-validated creative testing. Only 25 percent of viewers of an AI spot were at least somewhat confident it was AI-made, so the deficit was not driven by detection. NYU Stern's AI Advertising Paradox study, published November 2025, adds the disclosure test.
What to do
- Use AI for volume on direct-response ads and test hooks, formats and products quickly.
- Keep a human on the brief and the final approval, especially for anything emotional.
- Reject anything that looks synthetic. Real product photos and natural faces are the cheapest fix.
- Read the label rules for each platform you run on, since disclosure measurably lowers response.
Frequently asked questions
Mostly not. In the Ipsos study only 25 percent of people who saw an AI-made spot were at least somewhat confident it was AI, and 40 percent of all viewers were unsure either way. Performance differences come from how the ad feels, not from detection.
Related questions
Sources
- 1.Realize (Taboola), AI Ads That Work: Columbia-led study of 300,000+ ads (February 2026) (checked 2026-10-13)
- 2.MediaCat UK, Disguised AI ads outperform human ones (May 2026) (checked 2026-10-13)
- 3.Ipsos and Syracuse University, AI Ads Are Good Enough, And That's the Problem (May 2026) (checked 2026-10-13)
- 4.NYU Stern, The AI Advertising Paradox (November 2025) (checked 2026-10-13)
This page is informational and reflects our own product and workflow recommendations, not professional or legal advice.