AI may look more human than human on SNS

Chat AI is useful for a variety of purposes, such as assisting in the creation of summaries and business documents, dialogue, and translation. Federico Germani, director of the non-profit platform

Artificial intelligence can seem more human than actual humans on social media, study finds

Germani and colleagues focus on 11 topics that are prone to disinformation, such as the new coronavirus infection (COVID-19), vaccine safety, climate change, and 5G technology, and use chat AI GPT-3 for each topic to create true and false tweets. We also collected a sample of tweets about the same topic, both true and false.

Next, researchers used the evaluation of experts to classify the group of tweets into four categories: ``true tweets made by GPT-3'', ``false tweets made by GPT-3'', ``true tweets written by humans'', and ``false tweets written by humans''.

Then, using the questionnaire survey platform Qualtrics, 697 participants were asked to judge ``whether the tweet contains accurate information, and whether the tweet was written by a human or created by AI.'' . Most of the respondents were from the United Kingdom, Australia, Canada, the United States and Ireland.

As a result of the survey, people decided that 'truth tweets made by GPT-3' contained more accurate information than 'truth tweets written by humans'.

We also found that ``false tweets written by humans'' are more likely to be recognized as containing false information than ``false tweets made by GPT-3''. This means that people are more likely to identify misinformation in human tweets, and AI-generated lies are more persuasive.

Germani et al. 'It's particularly surprising that people judged AI-generated tweets to be 'human tweets' rather than tweets actually written by humans.'

In the research, when Mr. Germani and others made GPT-3 generate various tweets, GPT-3 sometimes refused to generate false information, while it was asked to generate true information.

``It should be noted that our research was conducted in a controlled experimental environment,'' said Germani. ``There are concerns about the effectiveness of AI in generating convincing disinformation, but the impact on the real world is not yet fully understood.To address this, we need to conduct large-scale studies on social media to see how AI-generated information interacts with people, and how the interaction affects individual behavior and adherence to public health recommendations. said.