AI-Generated Text Is Getting More Common—Here’s Why That’s a Bad Thing

It's getting harder and harder to tell the difference between sentences generated by artificial intelligence (AI) and people. 

The good news is that a new paper shows that people can learn to spot the difference between AI-generated and human-written text. The ability to determine if computers craft online information is becoming critical with the rise of large-scale language models such as ChatGPT, which many predict will touch virtually every corner of our lives.

"If you receive text messages suggesting your grandmother is in urgent need of funds for a medical procedure, you'd like to know if it was a mass AI scam or a real family member," Kentaro Toyama, a professor of community information at the University of Michigan, and author of "Geek Heresy: Rescuing Social Change from the Cult of Technology," told Lifewire in an email interview. "And, at moments when real human connection matters, we'd like to hear from real people, not machines."

Telling the Difference Between Real and ChatGPT

Researchers from the University of Pennsylvania wrote their paper by examining data collected using "Real or Fake Text?", a web-based training game. Participants are asked to indicate whether a machine has produced a given text in a yes-or-no fashion. This task involves classifying a text as real or fake and scoring responses as correct or incorrect. The study found that participants scored significantly better than random chance. 

"Our method not only gamifies the task, making it more engaging, it also provides a more realistic context for training," Liam Duga, one of the study's authors, said in a news release. "Generated texts, like those produced by ChatGPT, begin with human-provided prompts."

Toyama said there are many reasons why learning to distinguish between computer-generated text and human writing is important. For example, teachers would like to know whether students are submitting essays they wrote themselves or text written partly or entirely by computer. He noted that recently, Vanderbilt University sent an email written by ChatGPT to its community linking the tragedy of the Michigan State shooting to campus efforts toward inclusion. There was a swift backlash. 

"There are plenty of other such situations," he added. "In fact, I believe that one essential form of regulation of AI is that the law should require any text, image, audio, video, or other creative output generated by a computer to be clearly marked as such." 

Spotting AI-Generated Text

Toyama is pessimistic that people can consistently spot the difference between AI-generated and human-generated text, saying, "in the long term, it will become virtually impossible because the AI will become better and better." He pointed to informal experiments that suggest that even experienced teachers have difficulty distinguishing student writing from ChatGPT. 

"I have a colleague who claims that he spotted two instances of ChatGPT submissions because the students' writing had suddenly dramatically improved—but, ChatGPT can be directed to write at different levels of ability or to introduce errors," he added.  

Computers might be used to identify if a particular text is human-made. Parag Arora, the CEO of Kwegg, an AI content system, told Lifewire in an email interview that research is underway to develop programs that spot machine-generated writing. 

"AI classifiers being launched by research organizations, including OpenAI itself, look promising," he added. "However, GPT advances seem to be always a step ahead of them, and by nature of research, classifiers are always at least one generation behind generators. Some work has been done on leaving cryptic signatures as a policy for every GPT generator, which can be a sure-shot step toward solving this problem."

There are some ways for users to spot machine-generated text, Robert Brandl, the CEO of Tooltester, said in an email interview. He suggests trying out AI tools like ChatGPT to see what kind of content they generate to get more familiar with them. Among the giveaways are formulaic structures where sentences start with "Firstly" and "Secondly," etc., and a lack of sarcasm or humor in the text. 

"As AI develops, however, the tools that are designed to spot AI text will learn too," he added. "Indeed, it may get to a point where people need to rely on these tools to understand when AI is being used, that is, if publishers don't openly disclose AI usage anyway."