Ipsos blind test: 40% can't spot AI ads, but human-made ones still drive 14% more short-term sales
Ipsos and the Newhouse School at Syracuse University ran a controlled experiment comparing AI-generated ads with human-made ones: they took real 30-second spots from 10 brands including Cheerios, Fiat, H&M, and Visa (all pre-2021), used Gemini to reverse-engineer each ad's creative brief, then had Sora 2 fully regenerate the ads from those briefs. The resulting 20 ads (half human, half AI) were shown in a blind test to 3,000 US viewers.
Confirmed
- About 40% of viewers could not tell which ads were AI-made, meaning AI ads have passed the "indistinguishable" bar
- Even when they couldn't identify the source, viewers still rated human ads higher: more eye-catching, more imaginative, funnier, more distinctive, and more conversation-worthy
- On quantitative measures, human ads scored 14% higher on Ipsos's short-term sales metric and 17% higher on long-term brand equity
- 30% of the AI remakes used a narrative approach, differing from the composition of Ipsos's larger ad library
- The findings were summed up as "AI ads are good enough — and that's exactly the problem," with the brief identified as the key area for improvement
Why it matters
- Around the same time the study was released, a startup raised $3.5 million to mass-convert website content into thousands of AI ads; the Ipsos experiment provides the first serious controlled data on this "AI advertising at scale" approach
- The data suggests detectability is no longer the main bottleneck — the real gap lies in sales lift and brand building, a key warning for cost-efficiency-focused advertisers: being able to produce an ad doesn't mean it can sell
2026-11-04 ~ 2026-11-04 · 6 related posts
Primary sources
- Ipsos: 40% can't spot AI ads, but human-made ones still sell 14% better — petesena · 2026-11-04
- [source] How the Test Worked: Gemini Reverse-Engineered Briefs, Sora 2 Rebuilt 10 Brands' Ads — petesena · 2026-11-04
- [source] 3,000 Viewers Blind-Test AI Ads: 40% Can't Tell, Yet Still Prefer Human Work — petesena · 2026-11-04