How We Upgraded Adweek's Ad-Pitching Super Bowl Bot With a Smarter, More Eerily Human AI

When we first launched Adweek's Super Bowl Bot last year, the goal was to see whether the most sophisticated text generation AI at the time could write a plausible pitch for a Super Bowl ad on its own. The answer then was that it could—but only very occasionally. More often, it generated ideas that, while lucidly worded, were basically akin to fever dreams. That was entertaining enough but rarely mistakable for the output of any but the most deranged Big Game advertiser. However, that all might have changed, thanks to upgrades we've made to the Super Bowl Bot. In the short time since we first launched the bot, long strides have been made in the machine-learning tech that powered the bot. This year's version of the bot is far more sophisticated and takes advantage of the most up-to-date version of content-creating AI—one so powerful, access to it has been carefully limited. While our Super Bowl Bot still produces its fair share of ridiculousness, the bot is now notably better at staying on topic, registering the nature of a given product or brand and creating ad concepts that feel like they have a point. Sometimes, its ad ideas are so run-of-the-mill that they are downright... boring? https://twitter.com/SuperBowlBot/status/1355199832307224580?s=20 No matter how articulate the bot might seem, however, at the end of the day, it's still performing essentially the same function as your iPhone keyboard or Gmail window—guessing the next words in a sequence—but on a much more massive scale. It's still far from our sci-fi-colored conceptions of hyper-intelligent AI, but it's getting more lifelike by the day. Over the course of the next week, we'll be using the Super Bowl Bot to examine the opportunities and limitations of this kind of natural language processing tech for creative industries and having fun with the wacky world of neural net-produced content. You can follow along and request brands and/or celebrities you want to see in the ad concepts at @SuperBowlBot on Twitter. How we built this year's bot The original Super Bowl bot was built on a state-of-the-art, open-source machine learning model called GPT-2 . Last summer, the research group behind that program, OpenAI, released GPT-3, which is around 100 times more powerful than its predecessor and trained on nearly 1 trillion words from text across the internet. Unfortunately, we at Adweek—like most of the world—don't have direct access to GPT-3. OpenAI has made the system available through an API that it only allows a limited number of developers to use, and despite repeated applications to join the list, we haven't gotten a response. So we decided to hack it. OK, we didn't literally hack it, but we did come up with a pretty interesting workaround. For this year's bot output, we've been using a version of GPT-3 integrated into a text-based role-playing game called AI Dungeon. The game offers an option to create a custom premise for a game, and if you prompt it with enough text in a certain style, it will drop any pretense of fantasy adventures and instead adopt the format of the input material. The new Super Bowl Bot is clearly better at recognizing an actual product or brand, as seen here where each version is fed the same prompt. It is, admittedly, not a perfect representation of GPT-3. The creators of AI Dungeon have said that they put certain limitations on the AI in order to create more coherent story lines for the purposes of the games, and it has features that improve context recognition and eliminate fragments and repetition. But the base technology is the same, and the loophole is known among social media and online forum posts. While GPT-2 requires a large volume of training text to hone it to a specific style—around 5,000 ad descriptions in our case—GPT-3 is so powerful that it can pick up on a theme from only a few example passages. It only took a handful of Super Bowl ad summaries to get the AI writing its own, although it has gotten much better at it as we've continued to add more of the training data. It's also created a running log in which the system can continue learning on both the training data we feed it, its own text output and any other contextual information we might want to feed it. https://twitter.com/SuperBowlBot/status/1354874981432389637?s=20 What it means With the advent of GPT-3 and other massive models of the same kind, text generation AI has gained more traction this year among companies looking for commercial applications. OpenAI launched its first paid product around GPT-3 with brands like Reddit and Quizlet signed on to use it. And Google already taps a language model similar to GPT-3 to give its search results more context. But the tech is still in its infancy in many ways, and it remains too unpredictable at scale for a lot of business applications beyond role-playing games like AI Dungeon and novelty gags. Still, the Super Bowl Bot shows how it might eventually aid in the creative process for advertisers, giving creatives a way to generate outside-the-box ideas or play around with a concept in collaboration with the technology. https://twitter.com/SuperBowlBot/status/1354500467838119936?s=20 For all the latest Super Bowl advertising news—who’s in, who’s out, teasers, full ads and more—check out Adweek’s Super Bowl 2021 Ad Tracker and the rest of our stories about the Big Game. And join us on the evening of Feb. 7 for the best in-game coverage of the commercials anywhere.