Most people won’t be surprised to hear that AI has been used to create music, art, and even fashion. Now it’s coming for your ads. Artificial intelligence can now write adverts that have a better chance of success than the ones you might think up yourself, according to recent research from IBM. That could make an impact on everything from the kind of content that appears in e-mail blasts to the layout of web pages.
AI is already being used by some firms to create headlines and ad copy for marketing campaigns, but it can also generate text on a large scale.
For example, IBM’s technology has been trained to generate product descriptions based on images. The software uses deep neural networks (a form of AI) to teach itself what words go together and how language is structured.
Why artificial intelligence is being used to write adverts?
At first glance, this might seem like a shortcut but the results are mixed. In one example, researchers fed a computer thousands of fashion images from Pinterest to create text descriptions of each picture. The resulting text read as follows: “She wore a yellow crocheted sweater and black leather shoes with laces. The red bag was made of wool.”
“The AI basically comes up with sentences that are semantically, structurally sound, but it tends to be quirky,” says Kunal Anand at IBM Research in Ireland. His team helped create an alternative system that uses AI to generate product descriptions based on pictures, but in a more natural and readable way.
To do this, the researchers used a different AI technique known as semantic parsing. It’s good at analyzing the reasoning behind sentences without delving into too much detail about individual words. “It is inspired by how we humans understand our world,” says Anand.
The system starts with basic facts about the product, such as its price and color. Then it creates sentences based on what people usually say about similar products (for example: “A new pair of designer heels from Jimmy Choo”). The resulting copy might not be Pulitzer prize-winning material but is still more readable than the sentence about the crocheted sweater.
Similarly, IBM’s researchers have developed AI software that can write product descriptions in various languages, based on images of what it has been trained to recognize. This means customers in Japan could read product descriptions originally written for French shoppers.
News outlets are also using AI to create stories and headlines
The Associated Press (AP) is a big user of automation, which it says helps it to produce more stories in less time. The AP’s software has produced more than 3,000 news stories so far, with some going out under the byline of one of the wire service’s human journalists. “I’ve seen great examples that make me think, ‘Wow, I didn’t know that was possible,'” says Lou Ferrara Jr, the AP’s deputy managing editor.
Facebook has also started to use AI for its news feed algorithms (it plans to eventually turn off the human curators). If a user writes anything related to technology, such as “I updated my iPhone”, their friends are more likely to see a story about the new iPhone before they see stories on politics. The same applies to any reference to social media, including words like “poking” or “tweeting”.
But AI is not just used by advertisers and news outlets. IBM has trained its software to write straightforward ad copy too. For example, it has written slogans for the World Cancer Day campaign, and product descriptions for Amazon.com.
“This is early days but it’s really cool,” says Ian Goodfellow at OpenAI, a non-profit research company that focuses on AI safety issues. He thinks we will see more automated content creation in the future, especially as systems become proficient at writing long, flowing narrative pieces. “You can imagine taking a short piece of text and then having an AI iteratively make small edits to it and improve it,” he says.
At one point or another, we will have to think about how much of our news is written by machines. But for now, Ferrara Jr has some advice. “I would take any computer-generated news story with a grain of salt,” he says. “You still want to read it and understand what’s going on, but you’re probably not getting the whole story.”