Many people enjoy the experience of using AIs like ChatGPT or voice assistants like Alexa to find out information, but it turns out doing so makes it less likely you will spot inaccurate information.
The conversational tone of an AI chatbot or voice-based assistant seems like a good way to learn about and understand new concepts, but they may actually make us more willing to believe inaccuracies, compared with information presented like a static Wikipedia article.
 |
| Voice assistants provide information in a casual way Edwin Tan/Getty Images |
To investigate how the way we receive information can change how we perceive it, Sonja Utz at the University of Tübingen, Germany, and her colleagues asked about 1200 participants to engage with one of three formats.
The first involved text appearing letter by letter in the style of a large language model, similar to OpenAI’s ChatGPT; the second provided information through a voice-based device like Amazon’s Alexa; and the third was a static, text-based copy of Wikipedia. In all cases, participants weren’t able to interact with the systems as they might with the real versions, in order to keep the experience consistent across the study, but they were presented with the real brand names.
In some cases, participants were given entirely accurate information, while others saw factual errors of the kind that can be produced by large language models. Some of these errors were plain wrong, such as naming Sofia as the capital of Slovenia, but others were merely internally inconsistent, meaning a statement was contradicted elsewhere in the same response.
The downside of chat
When asked to rate the accuracy of information on a 7 point scale, people were most able to detect errors when reading a Wikipedia-style article
 |
| SOURCE: ANDERL, C., KLEIN, S.H., SARIGÜL, B. ET AL. SCIENTIFIC REPORTS DOI: 10.1038/S41598-024-67829-6 |
The participants were asked to rate the accuracy, trustworthiness and believability of the information they encountered on a seven-point scale, where 7 is entirely accurate. In the answers where errors were introduced, information presented in a static text-based format was rated as 4.24 on average. The ChatGPT-like set-up was rated as 4.76, while the voice assistant was rated 5.29.
Participants rated all of the systems higher for answers where no errors were introduced, but the differences were smaller for the voice assistant and chatbot, suggesting people found it more difficult to identify errors in these formats. “It has this dialogic presentation, which is why you tend to believe it,” says Utz.
“It's long been understood that people are easily convinced that objects that display or mimic lifelike qualities are in fact alive,” says Mike Katell at the Alan Turing Institute, UK. “It’s not surprising that people find conversational chat agents more convincing and credible than static text because humans seem to be hardwired to see life where it doesn't exist and to assign credibility to things that seem lifelike.”
The way the chatbots and voice agents are programmed to speak in a friendly and helpful conversational tone also confuses people further, says Katell. “The makers of chatbots aren't necessarily trying to deceive us, but they are trying to put users at ease. The goal of a large language model is ‘plausibility’ and that is achieved through believable content and inoffensive – if not soothing – delivery.”
Utz says it is important to recognise that the way information is presented to us can make a difference to how much we believe it. “People now have learned that these models hallucinate sometimes,” she says. “But what people do not know yet is that just this interaction mode makes us more prone to believe whatever is written there. That should be in media literacy training.”
OpenAI declined to comment on the paper’s findings, while Amazon didn't respond to a request for comment.
Journal reference:
Scientific Reports DOI: 10.1038/s41598-024-67829-6
0 Comments