Did ChatGPT do this homework? It's difficult to prove, research shows


Welcomed by students, feared by teachers, AI chatbots are set to change education. Many educators have pinned their hopes on new AI-detecting software tools in the fight against cheating with ChatGPT. But one plagiarism researcher has bad news for them. — Photo: Philipp von Ditfurth/dpa

BERLIN: Over the past year, researchers have found clear evidence of the use of AI text generators in several scientific papers.

And yet when it comes to the illicit scholarly use of AI chatbots, we may have only seen the tip of the iceberg.

Authors have been caught publishing works with legitimate-looking scientific references that were entirely invented by AI. Others were shown to have left snippets of text from ChatGPT's user interface pasted in their work.

One maths study was even retracted after it was shown to have inadvertently featured the AI interface phrase "regenerate response", as reported in a Nature article from September 2023.

The use of AI text generation is at its most problematic when it is not being made transparent. In other words, when services are possibly faked.

The question of how to deal with the shift towards AI text generation has been debated at universities ever since ChatGPT was made public at the end of 2022.

Despite the global hype surrounding its ability to summarise information and write texts at the linguistic level of a human, experts say tools like Microsoft-affiliated ChatGPT and Google's Bard are still a long way from human intelligence.

Study: AI detectors often fail

Many teachers have had their hopes pinned on software that claims it can detect AI texts. Does this mean the end of cheating?

Not at all, says one expert.

"The hope of a simple software solution for unmasking AI texts will not become reality," says plagiarism researcher Debora Weber-Wulff from Berlin's HTW University of Applied Sciences.

"There are lots of self-proclaimed detector programs, but they don't do what they're supposed to." Some manufacturers have even admitted shortcomings and limitations themselves, she says.

Weber-Wulff was involved in a study for which 14 apparent AI detectors were tested. According to the study, these tools did not deliver reliable results when it came to the question of whether a human or a machine had written a text.

The research team published their findings at the end of December in the International Journal for Educational Integrity.

"In uncertain cases, the systems tend to assume human authors," Weber-Wulff explains. "Because, of course, it is not desirable for people to be wrongly accused. That would also be disastrous in the education sector."

Plagiarism easier to prove than AI use

However, the study identifies the core problem as being that around one in five texts created with AI was not recognised as such. According to the study, the rate of AI usage not recognised by the programs increases further if the AI text has been revised by humans.

The results of the detectors are also not easy to interpret for the average user: Some provided a percentage indication of the probability that the text had been produced by an AI tool.

There was a lack of concrete evidence, meaning that universities may find it very difficult to prove misconduct on this basis. "Unlike with plagiarism, it is not possible to compare the text with the original," says Weber-Wulff.

Weber-Wulff says she is aware of cases in which lecturers have raised suspicions and students have admitted to using AI. She also assumes that AI chatbots are being widely used – often without the students being aware of what they are doing wrong.

In the study on detectors, the experts state that higher education institutions were not prepared for how quickly and radically freely accessible AI tools have improved. And using an AI for help is also not always unethical.

"We need to think very carefully about how we measure performance," Weber-Wulff says. This may mean that future tasks should be set in a completely different way than they have been in the past.

Getting the student to find mistakes in the responses of AI tools needs to be part of the job, she says.

After all, AI chatbots are little more than parrots Weber-Wulff says, since they only parrot what they have heard. That's why it's important to teach students the standards of academic writing, such as the purpose of footnotes.

If AI systems are used, they need to be handled transparently. "And you have to take full responsibility for all the rubbish produced by the system. No excuses." – dpa

Follow us on our official WhatsApp channel for breaking news alerts and key updates!
   

Next In Tech News

Atos creditors reach deal to rescue debt-laden group, La Tribune says
In an online world, a new generation of protesters chooses anonymity
After two winsome Ori games, a pivot into dark fantasy
Teenager in China dies of heart attack after teacher forces her to exercise, insists illness is ‘fake’, delays first aid, enrages mainland social media
NoSpace is Gen Z’s answer to MySpace
What if customers were rewarded for tipping their meal delivery drivers?
Reddit CEO beneficially owns 61.5% of class A shares, regulatory filing shows
Exclusive-Stanford AI leader Fei-Fei Li building 'spatial intelligence' startup
Tech platforms make pitch for ad deals as TikTok is roiled by politics
Intesa targets new digital-only clients after antitrust blow

Others Also Read