Rava is an online news portal providing recent news, editorials, opinions and advice on day to day happenings in Pakistan.
If you enter the page of the giant of the electronic commerce Alibaba, a service of wholesale purchase that is a species of Chinese answer to eBay, you will see images and descriptions of anything that you want to buy, from sinks of kitchen until luxury yachts.
Each product has a short title, but for the most part they are little more than a list of keywords: search terms chosen to make sure that that cellphone USB charger or that pair of fireproof coveralls emerge first in a sea of thousands and thousands of similar products.
It seems easy, but it has a certain art.
Alibaba, however, recently revealed that it is training Artificial Intelligence (AI) to generate these descriptions automatically.
And they are not the only ones.
In recent decades, Artificial Intelligence has been taught to compose music, paint pictures or write (bad) poems. Now he also writes ad text, at a rate of 20,000 lines per second .
“Generative bots are the new chatbot,” explains Jun Wang of University College London. “Generating text is just one of its applications”.
The ideal text to a couple of clicks
Launched by the branch of m to rketing digital Alibaba (Alimama), the text writer for Artificial Intelligence applies deep learning ( deep learning ) and technology of natural language processing millions of product descriptions on Tmall and Taobao pages from Alibaba to generate new text.
“The tool eliminates the discomfort of having to spend hours looking for design inspiration by looking at competitors ‘listings and manufacturers’ pages,” says a spokesperson for Alibaba. “The user can create his ideal text with just a couple of clicks “.
Beyond their incursions into the art world, creating bland text such as the one used in an advertisement is where the generating systems will have a greater impact in the short term.
The software will produce millions of words and images that will see, and influence, millions of people each day.
And if they do the job well enough, we will never realize the difference.
The line between what is the agency of men and machines is already blurred on the internet.
Twitter bots disseminate seeds of misinformation, spam bots generate strangely poetic emails about Viagra, and automatic aggregators find and republish online news articles so quickly that it can be difficult to determine who posted what before and when.
“Welcome to the future”
Consider the news about the Alibaba editor. The English version of the press release was used by several news sites, mostly from the United Kingdom, the United States and India.
But among those first reports he appeared a video on an unknown Youtube channel called “Breaking News” ( “Latest News”).
There, a synthesized voice reads the news, with subtitles appearing on a series of archive images related to Alibaba and e-commerce.
And hidden in the description of the video there is a link to the source of the text: an article published more or less an hour before by the International Business Time, a web page based in India.
The speed and the strange bungler with which the original story was reused – the subtitles are copied over as if they were part of the main text – suggest that the video was generated automatically .
So does the fact that, in addition to Alibaba’s video, the channel seems not to post anything other than international football news reports, also republished from other sources.
It is possible that someone is choosing which stories to republish, but you do not see any human activity in the channel or in the Twitter account associated with it.
So we have news about an Artificial Intelligence produced by another . Welcome to the future: at the same time rare and mundane.
“It’s not science fiction,” says Wang.
The UCL researcher believes that advertising is an ideal environment for generating Artificial Intelligence because it has a clear objective. “What you want is to maximize the number of people who click and then buy,” he says. “We are not talking about generating art.”
According to Alibaba, using this tool is simple . You provide a link to the object for which you want a description and then click on a button. “This generates numerous text ideas and options,” says the Alibaba spokesperson.
“The user can then alter everything, from the duration to the tone, as it seems best”.
The tool is also prolific. Alibaba says it can produce 20,000 lines of text per second and that it is being used almost a million times a day by companies -even by American clothing brand Dickies- who want to create multiple versions of ads that keep getting our attention when they appear in spaces of different sizes on web pages.
And not only Alibaba does it. Its main rival, JD.com, says it also uses software, which it calls “Artificial Intelligence writer robot” to generate product descriptions.
According to the ZDNet technology website, the JD.com system can produce more than 1,000 “pieces of content” per day and has a weakness for flowery language, describing for example wedding rings as a symbol of “drops of holy matrimony of heaven. “
Mark Riedle of the Georgia Institute of Technology is skeptical that these tools are as good as all these favorable public relations suggest.
Even if we ignore the claim in Alibaba’s press release that his Artificial Intelligence editor can nail the Turing test , during which Artificial Intelligence must pass as a human, there are questions about this approach.
For starters, we do not know how good these systems are to achieve that clear goal of getting people to click and then buy, a process known as conversion.
Learning to produce text that describes an object is definitely the kind of thing in which the generating systems have become good, says Riedl. “You can take a picture or a few keywords and produce something that looks like a product description .”
Artificial Intelligence can recognize the image of a camera, for example, find what you know about this object and make a short description that seems written by a human.
But this is only half the work. “Creating text is really about long tail (in Spanish, long queue, or aggregation of niches),” says Riedl.
To convert clicks into sales, especially when the competition for online attention is so fierce, you need to address the specific concerns and interests of a particular audience , possibly niche.
“You do not want to just say ‘this is the camera and these are the features’, you mean why someone should buy it or why this camera solves problems that others do not solve,” says Riedl.
“This requires a lot more context, you ‘ll want to have a lot more information about your product and about your audience,” he explains.
The problem with an automatic learning approach like the one used by Alibaba and JD.com is that the generating system will tend to learn the average.
“Artificial Intelligence is very good with generic formats , but the more you want to specialize or customize, it becomes a much, much more difficult problem,” says Riedl. “I do not think we’re still there.”
Maybe not, but that’s where we’re headed.
To understand why, first you have to immerse yourself in the ways in which advertising is already personalized and adapted to your habits.
Wang, for example, is a co-founder of a company called MediaGamma, which uses learning by reinforcement – a type of machine learning that is also behind DeepMind’s AlphaGo software – to help advertisers buy ad space.
When you visit a web page with ads, the ones you’ll see will probably be different from what another person will see. This is because the ads have been selected specifically for you .
As soon as you upload a web page, the page allows Internet advertising agents to know who is visiting, and then a high-speed bidding war begins that usually involves about 100 advertisers.
The winners can show you their ads. And the whole process ends in 100 milliseconds , faster than the blink of an eye.
In this context, the MediaGamma AI helps advertisers make smarter offers in this automatic auction. How much is this person’s attention worth?
“We do not necessarily know who you are, but we know your online activity,” Wang explains.
Google crawlers operate on about 75% of the most popular websites. And the three largest ad networks, Adsense, Admob and DoubleClick, are managed by Google.
And there are few places on the Internet where Google can not track you . And if Google can not see you, Facebook, which has trackers on 25% of those sites, probably does.
These trackers record what we are looking for: what websites we visit and how much time we spend on them .
Let’s say that someone is interested in shoes and it is known that they bought a particular type of footwear in a particular store. The MediaGamma AI learns to categorize Internet users based on all the information they have available.
If your online activity is similar to that of the person who bought shoes before, it will tell a footwear advertiser to make an offer.
“It is very likely that you will convert,” says Wang. “We also estimate, for this type of user in this type of market, how much there is to bid to win.”
But this is just the beginning . Last month, MediaGamma won a grant from the ‘innovation agency’ of the UK government to develop more advanced AI that can generate text and images for specific ads.
This would involve in practice putting together something like the Alibaba editor with the current MediaGamma user analysis technology.
So, instead of using your online activity to decide which existing ad you should see, that information could be used soon to generate a personalized ad on the fly.
“We could have an ad specifically tailored to a person’s tastes,” Wang explains.
Or in the hands of Alibaba, such a system could generate a customized item description , designed to please your individual preferences, based on what you know about your buying habits.
This would be a direct sale to the long tail that Riedl imagines.
Next step: narrative intelligence
These AI systems are getting smarter, but are they becoming more creative?
Here’s a famous six-word story by Ernest Hemingway: ” It sells : baby shoes, never worn . “
It is an advertisement, similar to the descriptions produced by the Alibaba AI. But the emotional resonance of Hemingway’s words comes from his deep understanding of human life , something that machines do not have.
Even if they produced those words, we would not react in the same way. At least not yet.
Riedl, however, is interested in giving AI a form of narrative intelligence, the ability to build and understand stories more as people do.
One of their experimental systems, called Shezarade, generates short narratives based on information from open sources about common human activities, such as an exit to the cinema. Here is an excerpt from one of those stories:
With sweaty palms and a racing heart, John went to Sally’s house for his first date. Sally, her pretty white dress floating in the wind, carefully entered John’s car. John and Sally drove to the movies. John and Sally parked the car in the parking lot. Eager to feel prepared, John had already bought tickets to the film in advance. A pale-faced usher was standing in front of the door; John showed the tickets and the couple entered. Sally was thirsty, so John hurried to buy drinks before the movie started. John and Sally found two good seats near the back. John sat down and raised his arm so that he and Sally could snuggle. John paid more attention to Sally during the movie and drank his drink nervously. Finally, gathering courage, John extended his arm to hug Sally. He was relieved and euphoric to feel that she was approaching him in response. Sally got up to use the bathroom during the movie, smiling shyly at John before leaving.
It is not entirely Hemingway, but the narrative generation is a growing area for AI. And, for Riedl, the narrative intelligence will help the AI to understand the worldin a way more similar to how we do it; We often tell stories to make sense of things.
Such an understanding could make the AIs with whom we interact, like Siri, seem less strange.
On the other hand, in addition to telling stories and becoming better sellers, you can also use more creative AIs to generate personalized campaign emails or social media posts for political candidates.
We are also seeing how the first IA editors generate short news bulletins: China’s Xinhua news agency recently announced that it would start using software to write some of its news reports .
This has generated concern, since many consider that Xinhua is a propaganda machine for the Chinese government. And if you have hundreds of bots that propagate a particular version of a story, this could be difficult to counteract and could have big implications in relation to news bias.
However, this is where we are headed : we see that more and more companies, political campaigns and consulting firms are beginning to use AI to help with their communications.
We should try to detect this when possible. But that is easier said than done. We may discover that we know these machine-generated messages as much as the online ad auctions that run each time a website is opened.
“Nobody notices , “ says Wang.