Generative AI has moved from making ad images faster to deciding which creative deserves the budget. A new study in Marketing Science offers one of the clearest field signals yet: an AI system that combined image generation, performance prediction, and brand-standards screening beat both a professional designer and an aesthetics-only AI model in a live Instagram campaign. For marketing leaders, the lesson is not to replace creative judgment with prompts. It is to build a measurable learning loop around creative.

The research, “Leveraging Generative Artificial Intelligence to Create Visual Content in Digital Advertising”, was authored by Remi Daviet of INSEAD and Yohei Nishimura of the Wisconsin School of Business at the University of Wisconsin–Madison. According to the INFORMS research release, the system was tested in a live campaign for an outdoor-activities company, where AI-generated nature backgrounds were vetted for brand fit and then field-tested.

The result is about optimization, not image generation

In the head-to-head comparison, the AI-generated portfolio produced a mean click-through rate of 0.98%, versus 0.65% for the professional human designer’s batch and 0.78% for an AI benchmark optimized only for aesthetics. Put differently, the reported AI portfolio’s average CTR was about 51% higher than the human batch and 26% higher than the aesthetics-first model. The system’s creative beat the human designer’s batch in 99.59% of samples, while also showing lower variability.

That distinction matters. A standard image generator is rewarded for producing plausible, attractive output. The study’s framework used Bayesian active learning: generate candidates, predict which ones are likely to perform and remain on-brand, test them, and use the results to choose the next candidates. The goal was not “make a pretty picture.” It was “find a brand-acceptable visual that earns a response.”

As Daviet put it, “When performance prediction and brand alignment are jointly optimized through active learning, the resulting visuals can deliver higher average returns and lower creative risk than traditional processes.”

The most important finding is durability

Many AI marketing demos look impressive until the novelty fades. This study included a follow-up test 18 months later, during a high-stakes booking season, without retraining the model. The AI portfolio recorded a 3.38% mean CTR compared with 3.24% for the company’s contemporary human-designed campaign, again with lower variance, according to the study announcement carried by EurekAlert.

The follow-up is not proof that AI will beat human creative in every category. The release does not report the sample size, impression count, confidence intervals, or detailed variance values, and the field setting involved one outdoor-activities brand on Instagram. Those limits should temper sweeping claims. Still, a persistent advantage after 18 months is a stronger business signal than a one-off benchmark or a synthetic preference test.

What this changes for marketing teams

First, creative operations need a performance data layer. Store the inputs and outcomes for every meaningful variation: hook, image composition, offer, audience, placement, spend, CTR, conversion rate, and downstream revenue. Without clean feedback, generative AI only increases the volume of guesses.

Second, turn the brand guide into testable rules. Some standards can be machine-checked—logo placement, prohibited claims, color ranges, product visibility, and required disclosures. Human reviewers should retain the final say over taste, cultural context, and whether a concept feels true to the brand.

Third, use small, controlled experiments rather than flooding the account with hundreds of near-duplicates. Keep a human creative director in charge of the brief and the guardrails; let the system explore the visual space and prioritize the next test. This protects brand coherence while making learning faster.

The strategic takeaway: build a creative learning loop

Nishimura warns that teams treating generative AI as “a simple production tool” can miss the performance gains and consistency of a purpose-built search-and-alignment system. That is the practical dividing line for the next phase of AI advertising. The winners will not necessarily be the brands generating the most assets. They will be the brands that connect generation, approval, experimentation, and revenue measurement into one repeatable loop.

Want to turn AI experimentation into accountable marketing performance? Real Internet Sales helps businesses build data-informed digital marketing systems that connect creative, search visibility, and conversion. Call 803-708-5514 or visit realinternetsales.com to start the conversation.

Research cited: Marketing Science paper; INFORMS release; EurekAlert release.