In recent years, with the continuous maturation of artificial intelligence technology, the number of AI-generated ad creatives used in advertising production hasgrown explosively. However, with the widespread adoption of AI applications, discussions about whether AI disclosure and labeling are necessary have become increasingly heated. To this end, multiple U.S. states havepassed relevant lawsrequiring disclosure when AI is used in specific ways. The European Union has also, under the Artificial Intelligence Act,mandated labeling for certain AI-generated content.。

Despite concerns that AI disclosure might affect advertising effectiveness, a recent study by MediaScience found that AI disclosure did not negatively impact ad performance, with key metrics including brand recognition, consumer attitudes toward the ad, and brand attitudes.

"We found that if consumers are informed, they are accepting of AI," said Duane Varan, founder and CEO of MediaScience. "Many governments are considering regulating and mandating AI labeling, but the question is how to implement such labeling."

This research, titled the "AI Labeling Impact Study," was conducted by advertising research firm MediaScience in collaboration with its AI video content platform MediaPet, and the Ehrenberg-Bass Institute for Marketing Science at the University of Adelaide in Adelaide, Australia.

Researchers evaluated four labeling methods among 900 U.S. participants, designed to simulate frameworks being considered by EU and U.S. lawmakers. The four methods were: a text label displayed in the first 3 seconds of the ad, a text label displayed in seconds 4-6, a text label displayed throughout, and an icon displayed throughout. A no-label control group was also included.

Effects of AI labeling

In terms of judging whether an ad was AI-generated, the continuous text label was most effective: nearly half of respondents (49%) could identify that the ad was AI-generated. The control group performed worst at 36%, while those who viewed ads with an AI icon were only 2 percentage points higher than the control group, at 38%. In the group that saw text disclosure in the first 3 seconds, AI recognition was 46%, while in the group that saw disclosure later in the ad, it was 40%.

For unaided brand recall, there was only a 7-percentage-point difference across the five study groups, and some AI disclosure methods actually aided recall. The control group and AI icon group both had unaided brand recall rates of 54%, while the seconds 4-6 group and continuous label group both had 61%. The group that saw the label in the first 3 seconds had a recall rate of 60%.

In brand recognition, there was a 5-percentage-point difference among groups. The continuous label group had the lowest recognition rate at 87%, while the first-3-seconds group scored highest at 92%. The control group, seconds 4-6 group, and icon group had recognition rates of 88%, 91%, and 91%, respectively.

The AI icon triggered the worst brand attitude scores, with only 44% of participants expressing a positive attitude toward the brand. The control group and the group that saw the label in the first 3 seconds had the highest brand attitude scores, both at 51%. The seconds 4-6 group and continuous label group were slightly lower, both at 49%.

The icon group also had the lowest ad liking, with 56% of participants saying they liked the ad, compared to 63% for the control group. Groups that saw text disclosure had the highest ad liking. Both groups that saw text for part of the ad had 70% liking, while the continuous text group had 69%.

"At the end of the day, labeling is actually a win-win proposition, as long as the ad itself is of good quality, it's not a problem for advertisers," Varan said.

When to label

As more laws take effect, when and how to label ads as AI-generated is also an issue the industry needs to address. New York state's law took effect in early June, requiring disclosure only when it involves "synthetic performers" (i.e., AI designed to imitate real people).

The study found that consumers largely agree with this approach, with 60% of study participants saying that disclosure should occur when AI is used to simulate humans. However, this percentage dropped by 14 percentage points to 46% when it came to animals. Less than half (45%) of participants said disclosure was needed when AI was used for product placement or voiceover. For color grading and lighting, only 21% of participants thought AI labeling was necessary. Other measured use cases included animation (41%), backgrounds (35%), scriptwriting (33%), music (28%), text/captions (24%), and translation/dubbing (23%).

A key takeaway for marketers is that consumer preferences may not align with actual AI recognition effectiveness. The AI icon was the preferred disclosure method among respondents, leading by 13 percentage points, but it had the lowest recognition rate among the four tested methods at 38%, while continuous disclosure ranked first (49%).

"The problem is that the AI icon doesn't actually improve recognition of AI content, so it's a loss for consumers," Varan said. "This could potentially be remedied by educating people about what the icon means. But first, text labels are indeed needed, and unfortunately, having the label appear later in the ad is not a truly viable option or substitute."