People are becoming more trusting of and warm towards AI models, according to a year-long survey of those living in the US.
at Stanford University in California and her colleagues gathered this information on the crowdsourcing platform Prolific. Between May 2023 and August 2024, roughly 1000 participants a month completed the researchers鈥� questionnaire, although 鈥� due to technical issues with the platform 鈥� only 12 months of data was collected over the 16-month period surveyed.
The participants, who were nationally representative of the US population, first had to answer an open-ended question: 鈥淲hat is the best metaphor for how AI works?鈥�
Advertisement
鈥淭here鈥檚 such a mixed bag of different perceptions of AI, and we decided to look at metaphors, because metaphors are a good way to get at people鈥檚 implicit thoughts,鈥� says Cheng.
The participants were then asked what AI tools they had heard of or used themselves, along with questions that probed their trust in AI.
Analysis showed that the metaphors evoked by study participants clustered into 20 groups of dominant images. Around 10 per cent of participants referred to AI as a tool like a Swiss Army Knife or calculator, another 10 per cent as a brain capable of reasoning and logic and the same percentage compared it to a powerful search engine. Another 4 per cent of respondents compared it to a genie. And around 1 in 200 people 鈥� which, expanded out to the general population, would still represent millions of Americans 鈥� equated AI to a thief.
鈥淭he genie metaphor showed up disproportionately for different populations,鈥� says Cheng. 鈥淭he prevalence of that metaphor, and also the metaphor of AI being a thief, was surprising to me.鈥�
Answers like this indicated that participants thought of the AI as having non-machine qualities 鈥� they were anthropomorphising it. Women, older people and people of colour were more likely to anthropomorphise AI than others, Cheng and her colleagues found.
Because the researchers tracked perceptions over a prolonged period of time, they also noticed an uptick in this anthropomorphisation of AI systems. The proportion of people who anthropomorphised AI increased by a third over the 12 months studied. 鈥淭hroughout our data, we see a lot of things that are influenced by cultural discourse,鈥� says Cheng. 鈥淭he way that AI is marketed and talked about in general definitely has an impact.鈥�
Whether someone used AI also shifted their perceptions of its capabilities: frequency of use was the largest predictor for trust in, and willingness to use, AI going forwards.
鈥淭he findings聽highlight聽the strong instinct we have to engage socially and positively with systems that show signs of human-like behaviour,鈥� says at King鈥檚 College London. 鈥淚t shows the benefits gained by tech companies by setting up their聽models to respond in the first person, as if they are our engaging and helpful friends.鈥�
However, Devlin points out, the paper establishes that this can sometimes lead to misplaced聽trust. 鈥淚s it responsible聽or ethical for companies to聽encourage this illusion, without making clear the limitations and risks?鈥� she asks.
The study also highlights 鈥渢he importance of sociotechnical, interdisciplinary research around AI, looking not just at the systems people use but the way that they interact with them, talk about them and think about them鈥�, she says.
arXiv
