Robots may take on a lot of jobs, but they can’t write… or can they? We spoke with a company that has software designed to do just that.
Suppose you want to generate a lot of product descriptions without the expense of paying a human writer to create them. Maybe you have fitness wearables or supplements or skin care products that come in a number of different sizes or configurations.
You know that using the same descriptions with only slight changes will keep Google from indexing your website. But human creativity would be wasted, right?
The software essentially takes a spreadsheet with features of the various products. Your human writer creates a few sentence patterns. Then the software uses the features and the patterns to create unique product descriptions.
A real-world example?
Gideon noticed a passage at a website and thought it might be an example of this autogenerated content.
“Some hikers were crestfallen,” it said, “when they felt discomfort in the ankle area.” The summary paragraph had some trouble with pronouns and carried on with one odd statement after another, culminating in the comforting news that “maybe with a little intervention, comfort is within one’s grasp.”
It sounds as though humans were involved at some point, but not as though it was actually written by a human.
We did a site search for “crestfallen,” and sure enough we discovered quite a few sentences using that in mix-and-match style:
- “A handful of day hikers were crestfallen when…”
- “Quite a number of testers were crestfallen after…”
- “A significant number of wearers were crestfallen as…”
Continued testing shows that the review analyses use this pattern:
- a phrase denoting numbers of users (“Droves of testers”)
- followed by a verb or emotional phrase ( “appreciated” or “were delighted”)
- and then a feature or problem (“good fit”)
We’re pretty confident that this is robot writing.
How’s it working out?
The site ranks well for “crestfallen hikers.” It doesn’t show up for our searches for “hiking boot reviews” but different people see different things in search. Google doesn’t like unnatural language, but they probably choose this website for some searchers, especially for those looking for less popular brands or models. This site has lots of reviews and posts summarizing reviews.
The more serious question is how human readers respond to obviously unnatural language. It’s possible that this website contains enough sheer data to overcome our reservations as well as Google’s.
Do you get those phone calls with a recorded voice telling you, “This is an important message. Do not hang up the phone!”
I hang up the phone. However, if I were receptive to what they’re selling, I might not let the fact that it’s an automated phone call bother me. Still, we know that 67% of people in a survey said they had hung up on an automated call service.
Companies that offer automated phone answer systems, chatbots and the like claim that listeners can’t even tell they’re conversing with a robot. The software makers we spoke with said that it takes just a minute for your human writers to check the autogenerated content and spiff it up for publication. In the example above, they might be doing a good enough job that most people don’t think it’s a robot writer, even if they think it sounds funny.
TechCrunch tried out a service that autogenerates blog posts or articles. They were crestfallen over the poor quality. Human beings would have no trouble identifying the autogenerated stuff as non-human. If you step outside the realm of content that can realistically be based on a spreadsheet of features, the robot’s lack of control over human language becomes obvious.
Should you try it?
We think autogenerated writing could work for product descriptions, weather reports, and maybe sports stats. We can set up chatbot interactions that could pass muster for a few minutes. It’s not going to work for your blog in the foreseeable future, but we’re certainly going to watch it with interest.