> content predicted to evoke reader sentiments like self-confidence or adventurousness
I'm not convinced that is different from Facebook's model - it's just less powerful.
Advertisers have always tried to place their ads alongside content that favors their brand identity, sure, but that's about the style of the content. That looks like Burton sponsoring top-tier snowboarders or Chanel advertising in Vogue. Watching Olympic snowboarders might make some people feel adventurous, but Burton's placement is also aspirational and cultural, a way of simply forming an association between the brand and high performance.
The NYT model didn't just put content alongside stories about adventure, it put content alongside stories expected to evoke adventurous sentiments. If a story about a daring Arctic expedition makes you feel relieved to be comfy at home, it could still associate a brand with adventure, but it's outside the sentiment target. That is inferring the emotional state of users, rather than of content. The main difference is that the NYT was making session-level judgements, rather than long-term ones. I find that much less objectionable (even if it's only out of a lack of data), but it's still in the category of mind-state targeting rather than content alignment.
Advertising based on the type of content is nearly as old as advertising. Some TV commercials air during dramas, others during comedies, others during sporting events. If you want to get more specific, you can elect to have your ad air during a certain show.
The only difference between this and what the NYT is doing is the former requires slightly more research on the part of the advertiser, to learn what the show or article is about.
Again, though, my point is that the NYT didn't just sell advertising based on content type. That wouldn't be new. "Project Feels" was an ML initiative to study how readers felt after reacting to stories, and create an ad-targeting tool based on that. It's very specifically about offering advertisers the chance to choose stories based on predicted reader demographics, behaviors, and emotional response, instead of simply targeting stories by category or topic.
To decide whether to air your commercial alongside a drama or a comedy, all you need to do is watch the show (and perhaps collect viewer demographics). To decide which section of the print NYT to advertise in, all you need to do is read it. But Project Feels was only possible by studying the behaviors and emotional responses of readers. It didn't try to alter those emotions in specific ways, so it's not equivalent to Facebook's project, but it's also not the same as content-based targeting.
The NYT is replacing a human reader determining a story's emotional value with an ML program determining a story's emotional value.
You're implying that the NYT's ML software is capable of finding some kind of secret, subliminal emotional traits that wouldn't be detectable to a human reader. I don't find this believable at all.
I'm not. If the NYT had allowed advertisers to handpick stories to appear alongside, I wouldn't consider this substantially different.
What I'm interested in is the switch from choosing to appear alongside a category or keywords (content targeting) to appearing alongside a type of story, with its more specific impact. The impact of ML isn't better-than-human parsing, it's almost certainly worse-than-human. It's just a question of adding story-level targeting which wasn't previously available, along with access to user studies that go beyond demographics and engagement to self-reported emotions.
I'm not convinced that is different from Facebook's model - it's just less powerful.
Advertisers have always tried to place their ads alongside content that favors their brand identity, sure, but that's about the style of the content. That looks like Burton sponsoring top-tier snowboarders or Chanel advertising in Vogue. Watching Olympic snowboarders might make some people feel adventurous, but Burton's placement is also aspirational and cultural, a way of simply forming an association between the brand and high performance.
The NYT model didn't just put content alongside stories about adventure, it put content alongside stories expected to evoke adventurous sentiments. If a story about a daring Arctic expedition makes you feel relieved to be comfy at home, it could still associate a brand with adventure, but it's outside the sentiment target. That is inferring the emotional state of users, rather than of content. The main difference is that the NYT was making session-level judgements, rather than long-term ones. I find that much less objectionable (even if it's only out of a lack of data), but it's still in the category of mind-state targeting rather than content alignment.