Have you ever watched a bizarre Super Bowl ad—Mountain Dew’s Puppy Monkey Baby, perhaps, or Burger King’s 45 seconds of Andy Warhol eating a Whopper—and wondered if an algorithm could produce similar results?
Or perhaps you’ve looked at how famously formulaic most Super Bowl ads are, with their celebrity cameos, animal hijinks and inspirational voice-overs, and questioned whether a machine could generate an ad idea that would fit right in.
Enter Adweek’s Super Bowl Bot, a text-generating AI trained on nearly 3,000 descriptions of Super Bowl ads—134,000 words in total, so far—sourced from around the web, including Adweek’s Super Bowl Ad Trackers from recent years. We created the AI bot to test the creative limits, possibilities and problematic weirdness of today’s most advanced neural nets available to the public.
A few months ago, we began tweeting some of the AI’s ad concept pitches from the handle @SuperBowlBot as its training process accelerated.
We’ll continue to update the Twitter account daily with new pitches in the remaining weeks leading up to the Big Game, hopefully adding features like predictions or halftime show observations as online chatter around the game picks up steam.
The bot can also take prompts, so if you’re active on Twitter, feel free to @-mention it with specific brands or stars you’d like to see in a commercial, but allow a little time, as we have to manually feed the prompt back to the AI. (We’re hand-curating the bot’s output to avoid any sort of “Tay situation.”)
We might even have the bot flesh out full scripts for a few of the best concepts. Let us know if you see a tweet worthy of the full storyboard treatment.
The blurbs the bot produces often straddle the line between thought-provokingly absurd and inexplicably disconcerting, but it’s already had some hits as well:
Throughout the training process, we’ve gotten to know the bot, from its quirks—including a slightly disturbing fixation on apocalyptic violence—to its various creative kicks, such as passing fascinations with actor John Krasinski or chimpanzees.
Sometimes the explanations for such idiosyncrasies reveal unexpectedly insightful observations about the source material. The bot’s preoccupation with war, for instance, is a distortion of the abundant mentions of military veterans and shows of troop support in real Super Bowl ads, sometimes blurred with the dystopian settings that have been hallmarks of the game since Apple’s “1984.”
The bot’s takes on the American Petroleum Institute, an occasional subject of its pitches thanks to a Super Bowl spot the trade group ran in 2017, tend to be savage:
It can recognize subjects in text prompts and inadvertently skewer them, too:
In technical terms, the bot is derived from a version of a model called GPT-2, a cutting-edge language generation AI developed by the research group OpenAI.
GPT-2 is trained on more than 8 million websites to understand the basic mechanics of language. From there, it can be fine-tuned through training with a relatively smaller dataset—such as a list of ad descriptions—which it will mimic in its output.
While our bot isn’t about to replace human copywriters or helm a full Super Bowl campaign anytime soon, it’s already proven it can at least string together an entertaining story concept. Stay tuned as it continues to evolve, and please join us in hoping it never becomes self-aware.
You can follow Adweek’s Super Bowl Bot at @SuperBowlBot on Twitter.