Advertisers have a lot of expectations. They target specific demographics to buy certain products at certain price points, but they’re often wrong. “It’s pretty hard to be world-class at this as a human. So we kind of say, ‘Let the machines do the job‘,” said Unity advertiser solutions VP Julie Shumaker during VentureBeat’s Transform 2018 AI conference.

The supplier of video game-creation software is reaching consumers through 3 billion devices every day, she said, providing rich opportunities for advertisers — and lots of data for learning about user behavior. “We’re taking a step before [the ad to understand] ‘What is the optimal outcome for the player or user?'” said Shumaker.

Advertisers might have very specific goals, like selling a $17 game install to a 22-year-old player, she said. They might not think about a 65-year-old woman. But machine learning may reveal that this woman is likely to spend about $3.99 over the course of three days. And if the cost of acquisition is 75 cents, it yields as good an ROI as higher-dollar goals for more typical ad targets.

“You’re able to try the crazy stuff,” said John Koetsier, VP of insights at marketing data platform Singular. One crazy example: a client that ran an advertisement for a game without showing any actual gameplay. It generated a huge amount of conversation about the game among a certain group of players. “You can try many, many things, because you can let the machine [learning] then figure out in real time what’s generating impact,” Koetsier said. “You can do stupid stuff, and sometimes stupid stuff is smart stuff.”

Guessing also happens on the creative side, said Rishi Shiva, CMO of Bidalgo, which provides ad automation technology to in-app marketers. The company recently rolled out a machine learning service called Creative AI that analyzes images to identify approaches that are likely to succeed. “Before you go investing hundreds of thousands of dollars developing video assets, you can actually run your historical images and videos through our system, and it will actually give you insights,” Shiva said. “What actually had a positive impact on the audience? What is it that people liked?” It can get as granular as the poses people take in images, he said. The software then develops a creative brief for content teams “that talks their language.”

Using AI to develop new types of insights is particularly important in view of new privacy requirements, like Europe’s General Data Protection Regulation (GDPR) rules and the California Consumer Privacy Act of 2018. “In light of GDPR and California’s GDPR-lite and laws coming up in different states in the U.S., the idea of using behavior data in advertising is under peril for a number of reasons,” said Ben Plomion, CMO of GumGum.

The company intuits the context of a particular online session by analyzing the images on the page. It then places relevant advertisements over a portion of the image. (Users can click to close the ad.) For its client Jeep, for instance, GumGum was able to place ads for the Cherokee over images of competing models, such as the Toyota RAV4. So instead of building a big behavioral model of the consumer, the company uses machine learning to understand the particular context in that moment.

A particularly clever example was a campaign for Vodafone in the U.K. last year. The telecom company wanted to advertise that it would carry the iPhone X, but Apple’s restrictive guidelines made it difficult to mention the product. So GumGum’s machine learning-based technology found images of the iPhone X and placed Vodaphone ads on top of them. “So by that association, even though we never mentioned the words ‘Apple iPhone X’, consumers knew that they were selling service for the iPhone X,” said Plomion.