As artificial intelligence becomes increasingly advanced, how many years out are we from machines being able to write mainstream commercials?
It might seem like a fantastical notion or an academic exercise to some, but for copywriters and creative agencies, it’s also an existential question that some in the industry are taking seriously. While text-generating AI in its current form is probably ill-suited for creative projects beyond novelty gags, like Adweek’s recently launched Super Bowl Bot and its surreal ad concept pitches, even these projects show glimmers of impressive complexity or genuine cleverness that hint at the growing, inventive power of AI.
“We are elbows deep in this stuff, trying to figure out what it means to our clients, what it means to the kind of work we’re doing, what it means to our processes and what it means to where the industry is heading,” said Michael Morowitz, executive technology director at R/GA Chicago.
Morowitz said he recently had a conversation with the agency’s global chief technology officer about the potential of GPT-2—in particular, the cutting-edge language generation model that serves as the foundation for our Super Bowl Bot—after seeing it used in a Dungeons and Dragons-style text adventure game in which the AI generates new scenarios as the player types his or her actions.
Research group OpenAI unveiled GPT-2 last spring, at first declining to release it to the public in full for fear that its eerily realistic-sounding copy might be put toward the mass production of fake news or spam (it gradually relented with a staggered release throughout the year). The program, and other breakthrough models of its ilk, are notable for the massive size of their training data. Text from around 8 million broadly representative websites was used in GPT-2’s case.
That training forms an extensive knowledge base for AI’s understanding of the mechanics of language and context in a general sense from which point it can be fine-tuned on narrower datasets of specific styles and concepts, like Super Bowl ad descriptions. Some experts think this capability could herald a new boom in language-generation AI in the coming years, as an analogous milestone did for image and facial recognition around 2012.
But BuzzFeed data scientist Max Woolf, who wrote the popular guide we used to fine-tune GPT-2, said he hasn’t yet seen any examples of agencies or brands deploying his guide in creative work, despite a proliferation in GPT-2-powered Twitter bots and other creative endeavors it helped facilitate.
One concern might be brand safety. Randomness-generating bots like GPT-2 are entirely unpredictable by nature, and its sensibility is informed by the wilds of the web on which it was originally trained. The Super Bowl Bot, for instance, has a definite violent, apocalyptic streak, and some of its output was too violent or disturbing even for our experimental project.
“It still needs a lot of human help,” said Janelle Shane, author of the AI Weirdness blog, which spotlights the foibles of neural network-generated content. “Without some form of curation, its results can be very hit or miss, and occasionally offensive.”
Agency executives say that honing these rough edges into something useful will take exhaustive, highly tailored training on specific brand narratives, and in some cases, the available data doesn’t allow that yet.
“AI could generate something fun, it could generate something viral, it could generate something wacky,” said Trace Cohen, co-founder and strategy lead for Venables Bell & Partners’ new AI division, Braive. “But in terms of something poignant as it relates to the purpose of the brand, [that’s] highly unlikely until it is effectively trained, and that training takes reams of data, it takes knowing the purpose, the ontology and vocabulary that match to that purpose, and it takes knowing the people and the general conversations around the brand.”
Still, even with such weaknesses, Morowitz sees AI playing a modest role in the creative process sooner rather than later, working side by side rather than in place of real human creatives, perhaps helping to stir creative juices or power predictive text tools.
“Just as everyone on a creative team has a different style of working, when you think about text generation models or image generation tools or other kinds of machine learning, if you think of them as having a seat at the table, as being a member of the team, their value starts to look a lot different,” Morowitz said.
Shane said creators are only beginning to unlock the full potential of these advanced text generators.
“GPT-2 is a model that keeps on giving,” Shane said. “There are so many interesting ways to fine-tune it to do specialized things, and I think so far we’re only scratching the surface of things we can build with it.”