The Guardian recently published an essay written entirely by a text robot. GPT-3 is a headline-grabbing, powerful Artificial Intelligence (AI) text engine. In this essay, it tries to convince readers that “robots come in peace.” The text was only slightly edited and of an impressive quality.
The Guardian’s example reminded me of a digital project we did with a robot writing texts for Germany’s leading business and financial daily, Handelsblatt. Our robot definitely came in peace — but went to pieces.
I’d like to tell you the story about the fate of Handelsblatt’s text robot. You know, there are many successful examples on how automated texts provide weather reports to Web sites, write summaries of soccer matches, or update real estate transactions. We decided a few years ago to use a robot to write about stock price moves.
The strategy behind it was simple: Handelsblatt has a decent CPM. Advertisers pay a lot to show their ads alongside financial news on our Web page. There also is a lot of search traffic around stock price movements in Germany. Additionally, our SEO ranking is pretty decent. So, adding some SEO optimised and automated content for those looking for a quick update on a certain stock was deemed worth trying.
We picked a regional AI company (Textomatic) to help us with the project. We taught its algorithm a lot about how media reports on stock markets. We also decided to integrate some human-produced news snippets about reports published by stock market analysts, which were provided by a German-language wire service (dpa-AFX). Thus, we created a continuous stream of reports on more than 100 German stocks, republished every 15 minutes with current stock market prices mixing text robot input and human written text.
Our approach was innovative and the strategy paid off. We attracted plenty of traffic and quickly calculated that the money we spent was refinancing our efforts. All lights on green!
So why did the robot go to pieces anyway? In a way, we became a victim of our own success.
For instance, our automated stories were of such quality that Bloomberg added them to its news system for financial market experts. Because of that, a lot of attention was put on the service. On days of dividend payouts for the stocks, our robot reported that a certain stock would lose a lot of its worth. But even if the stock price dip was mainly triggered by a dividend payout, headlines might read: “ABC stock crashes by X%.”
Well, the media department of company ABC didn’t like that — and neither did some Bloomberg terminal users. Our data set simply didn’t have any information on dividend payouts and thus couldn’t run a better headline like “ABC pays dividend — stock drops by X%.”
We put the robot on hold and went back to teaching it. It was quite difficult to find an up-to-date and affordable data set on dividend payouts. After we found the data and integrated it into our services, our text robot went back online, now in perfect shape. Robot-human texts with dividend knowledge.
But by the time, many German Web sites were publishing automated stock market texts. They competed for the same search engine traffic. And even though they didn’t care about dividend payouts nor added human written text, they got traffic anyway — although readers didn’t expect as much accuracy from those brands as from Handelsblatt.
With others competing, our click rates weren’t as good anymore. We also considered the reputational risk to be too big compared to the return on investment. Plus, we were putting more emphasis on our subscription strategy and didn’t want to put more management attention and development capacity on the text robot.
And, so, we pulled the plug.