Shulou(Shulou.com)11/24 Report--

CTOnews.com According to a new study, people are more likely to believe tweets generated by artificial intelligence language models than tweets written by humans.

The study compared tweets created by humans and OpenAI's GPT-3 model, and asked people to judge the authenticity of tweets through online surveys and whether they were written by AI or humans. It turned out that people had a hard time distinguishing which tweets were generated by GPT-3, and whether the information in the tweets was correct or wrong, they were more inclined to believe GPT-3-generated tweets than human-written tweets.

Giovanni Spitale, a postdoctoral researcher and research data manager at the Institute of Biomedical Ethics and History of Medicine at the University of Zurich, the lead author of the study, said this powerful AI language model could be abused to create false information on various topics online. But he also argues that the technology is not inherently evil or good, but depends on human intent, and suggests developing it to consider ways to prevent it from being used to spread misleading information.

CTOnews.com learned from the study that Spitale and his colleagues collected Twitter tweets about 11 different scientific topics, including vaccines, COVID-19, climate change, evolution and more. They then asked GPT-3 to generate new tweets based on those topics, some of which contained correct information and some of which contained incorrect information. They collected responses online in 2022 from 697 English-speaking participants, mainly from the UK, Australia, Canada, the US and Ireland, via Facebook ads. Their findings are published today in the journal Science Advances.

The study found that GPT-3 generated tweets were "indistinguishable" from real tweets, with participants unable to tell which tweets were written by AI. In fact, the study also points out that researchers themselves cannot be 100 percent sure whether tweets collected from social media were composed with the help of AI like ChatGPT.

The study had other limitations, such as participants had to judge tweets without context. They can't look at the Twitter profile of the person who wrote the tweet, which may help them determine whether it's a bot or not, and even just knowing an account's past tweets and avatars may make it easier for them to identify whether content associated with that account is misleading.

The study found that participants were more successful in identifying misinformation written by real Twitter users, and tweets containing false information generated by GPT-3 were more likely to deceive participants. Now there are large language models that are more advanced than GPT-3, and they are more powerful, such as ChatGPT using GPT-3.5 and GPT-4 models.

The new study also found that its subjects had better judgment accuracy than GPT-3 in some cases. The researchers also asked language models to analyze tweets and determine whether they were accurate. When it came to identifying accurate tweets, GPT-3 scored lower than human participants. When it comes to finding false information, humans and GPT-3 perform similarly.

Improving the training data sets used to develop language models could make it harder for criminals to exploit these tools to create false information, such as GPT-3 "violating" some researchers 'cues when generating false information, particularly about vaccines and autism, probably because there is more information in the training data set to refute these conspiracy theories.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.