Academic research is increasingly anthropomorphising technology – a trend that could mislead the public about how powerful artificial intelligence and other cutting-edge developments really are.
and her colleagues at Stanford University, California, analysed the content of more than 655,000 academic publications released between May 2007 and September 2023, along with the headlines of approximately 14,000 news articles citing some of those papers. They rated the extent to which each text used human pronouns such as “he” and “she” rather than “it”, as well as how often they used verbs that imply human emotion.
The researchers found that the way technologies are discussed in academic papers has changed over time, with a roughly 50 per cent increase in the level of anthropomorphism – where we attribute human-like qualities to non-human objects, tools or concepts – according to their rating system. News coverage was more likely to anthropomorphise technology than academic literature, though this hasn’t increased by a statistically meaningful amount over time.
Advertisement
In particular, academic papers that focused on large language models (LLMs) such as ChatGPT were more likely to rate higher for anthropomorphism than those dealing with other forms of technology.
Cheng says she wanted to look at the issue after noticing anthropomorphic language in academic papers she had been asked to peer-review. “I don’t think you should use language this way, because it has implications,” she says.
Those implications include overstating the capabilities of the technologies, potentially reframing conversations about how to handle and regulate them. This is particularly pertinent for artificial intelligence, which, as Cheng points out, is named in a way that is already inherently anthropomorphic.
“Anthropomorphism is baked into the way that we are building and using language models,” she says. “It’s a double-bind that the field is caught in, where the users and creators of language models have to use anthropomorphism, but at the same time, using anthropomorphism leads to more and more misleading ideas about what these models can do.”
“This is something that’s plagued AI from the beginning,” says at the Santa Fe Institute in New Mexico. “It very much influences the way we think about these systems.” Mitchell says that the problem has got worse as chatbots have become more fluent and human-like in their text outputs – resulting in users projecting human qualities onto them. “People are very prone to project human qualities on systems then engender assumptions about how these systems are likely to act and how trustworthy they are,” Mitchell says.
Team member , also at Stanford University, says that imbuing technology with human characteristics is common. “It really is natural to use a standard language,” she says. “Anthropomorphism is really a metaphor or a mental shortcut that we take to explain things.” But she believes that – particularly within the scientific community – it is important to use the right language and for researchers to be more mindful of how technology is presented.
The public can also play a role in this by being more sceptical of technology and its capabilities – something that starts with how we talk about it. “[Anthropomorphism] has ended up making us trust the system a little bit too much,” says Mitchell. “And then you’re unpleasantly surprised when they fail, or you see something very unhuman-like.”
arxiv
