Harvard-backed study of 500 million ad impressions finds AI creativity matches human performance

A large-scale academic study conducted by researchers from Harvard University, Columbia University, the Technical University of Munich, and Carnegie Mellon University has found that generative artificial intelligence (GenAI) performs on par with human creativity in digital advertising, challenging long-held assumptions about the limits of AI-driven content.

The findings, announced by global performance advertising firm Taboola (Nasdaq: TBLA), are based on an analysis of more than 500 million real-world ad impressions and around 3 million clicks. Titled “AI Ads That Work: How AI Creative Stacks Up Against Humans,” the study examined hundreds of thousands of live advertisements served through Taboola’s Realize platform, marking one of the most comprehensive evaluations of AI-generated advertising to date.

According to the research, AI-generated ads delivered performance comparable to those created by humans. Raw data showed AI ads recording a slightly higher average click-through rate (CTR) of 0.76 per cent, compared to 0.65 per cent for human-made ads. However, when rigorous statistical controls were applied, the difference narrowed, indicating broadly similar effectiveness.

One of the most striking findings was that AI ads performed best when they did not appear overtly “artificial.” Ads that blended AI efficiency with human-like creative cues significantly outperformed both purely human-made ads and AI ads that looked obviously machine-generated. Researchers noted that visual authenticity played a critical role in user engagement.

The presence of clear, prominent human faces emerged as a key trust signal. Ads featuring human faces were perceived as more credible and relatable, driving higher engagement. Interestingly, due to Taboola’s internal best practices and policy guidelines, AI-generated ads were more likely to include such trust-building elements than human-designed creatives.

The study also dispelled concerns that faster, cheaper AI production compromises quality. AI-generated visuals were found to maintain or even improve click-through rates without negatively affecting downstream conversions, suggesting advertisers no longer need to choose between speed, scale, and performance.

Early adoption of AI-generated ads was most visible in sectors such as food and beverage and personal finance, indicating that consumer-facing industries with high competition are leading experimentation with GenAI.

Commenting on the findings, Oded Netzer, Vice Dean for Research at Columbia Business School, said the dataset offered rare real-world insight. “By analyzing over 500 million impressions, we were able to move beyond GenAI hype and understand its true impact at scale,” he said, adding that AI can set “a new ceiling for engagement” when combined with human trust cues.

The study employed a quasi-experimental “sibling ads” methodology, comparing AI and human ads created by the same advertiser for the same campaign on the same day, ensuring robust and reliable conclusions.