Browse all

Artificial intelligence still has a lot to learn about writing

AI journalism amid plagiarism, gross errors and an inability to create interesting content

Artificial intelligence still has a lot to learn about writing AI journalism amid plagiarism, gross errors and an inability to create interesting content

Since OpenAI, the organization dedicated to research and development in the field of artificial intelligence, unveiled its new chatbot ChatGPT to the public at the end of November, many people have been amazed. The software actually offers unexpected qualities: for example, it can explain complex algorithms in the slang of a character from a 1940s gangster film or write new Bible verses on current topics. Recent studies show that people are often unable to distinguish between a text created by an artificial intelligence and one written by a flesh-and-blood human being. And it is precisely because of their amazing potential in writing that systems like ChatGPT are coming to the fore among laypeople. Since AI is unlikely to be just a bubble, as it was with cryptocurrencies, many are imagining the possibilities of how these systems could be used in the future, e.g. for copying during an exam (since one can write quite convincing essays with a simple hint). And more than a few have wondered why such technologies are not also used in journalism or, more generally, in publishing. However, the question in these cases is somewhat more complex than it first appears, precisely because systems like ChatGPT are not such perfect products.

Although software is capable of producing perfectly useful articles, the lesser-known case of CNET - a news site dedicated to technology - has also highlighted its limitations. Without clearly communicating it, CNET published at least 70 articles written by an artificial intelligence between November and today, until another website dedicated to technology took notice. Futurism had noticed a number of trivial errors in these articles, which CNET journalists had to correct by pointing them out (a typical practise in Anglo-Saxon journalism). In other cases there were warnings such as «We are currently checking the accuracy of this story». The problems were not just about a lack of journalistic reliability or gross errors, but also an inability to create interesting content for readers. CNET's auto-generated articles were simplistic, full of clichés and unappealing: one, for example, said: «Choosing between a bank and a credit union depends on the person. You need to weigh the pros and cons against your own goals to figure out which bank is best for you.» This is, of course, a sensible consideration, but one that is completely obvious to a user who wants to know more. Following a firestorm of controversy, CNET's editor-in-chief Connie Guglielmo announced the suspension of publication of articles written by artificial intelligence. She explained that this was an 'experiment' to test whether the technology could help journalistic staff "cover issues in a 360-degree way".

Another common problem is plagiarism: software scours the web for available sources, but without any critical consideration. Recently, the writer Alex Kantrowitz discovered that an automatic article generator had copied several sentences from one of his articles. The fact is that artificial intelligence tools are nothing new in journalism: in 2014, for example, the Associated Press news agency, one of the most important in the world, began to use such a tool for its compilation articles, i.e. articles that are always structured according to the same pattern. The Financial Times, in turn, uses a tool that automatically checks whether too many men are interviewed in its published articles to the detriment of women, while BuzzFeed has announced that it will use artificial intelligence systems to improve its online quizzes and other types of content.

The kind of articles in which artificial intelligence is increasingly being used, with all the problems this entails, are the so-called "SEO" (i.e. 'search engine optimization') articles, that try to appear in the first results of search engines in order to get more visitors. These articles do not have to be particularly elaborate or interesting, the only important thing is that they are there and that as many people as possible find and enter them so that the advertising revenue is higher. They therefore often follow a fixed pattern and heavily use the keywords of a certain topic to appear among the first search results. The risk of a massive use of artificial intelligence to create 'SEO' articles is that the search engines will only be filled with content that is very similar or that will be used in the future outside of marketing, for example to influence public opinion. «The fact that ChatGPT and similar tools can generate endless variations of persuasive text quickly, for free and potentially indefinitely is an opportunity for those who want to run propaganda operations, coordinated harassment campaigns, spam and other malicious activities» writes journalist Casey Newton in her newsletter Platformer«Platforms have historically struggled to determine with a high degree of accuracy which of their users are genuine and which are bots; if bots can use tools like this, the damage is potentially high.»