By simulating fake people while interacting with popular smart voice assistants, such as Amazon鈥檚 Alexa, Google Assistant and Apple鈥檚 Siri, researchers have uncovered the different approaches each system takes to learning users鈥� personal preferences and habits.
鈥淚 don鈥檛 know that consumers have the intuition that when you鈥檙e talking out loud, that also could potentially be used to profile you and then target you with ads,鈥� says at Northeastern University in Massachusetts.
All of these voice assistants rely on artificial intelligence models to interact with users, including more than 142 million people in the US, but they differ in how they gather information about their customers. Choffnes and his colleagues didn鈥檛 have access to proprietary details about how these assistants build a user profile 鈥� so they took another approach to find out.
Advertisement
The researchers spent 20 months feeding voice assistants from Amazon, Google and Apple more than 24,000 voice queries and text-based web searches. These interactions, recorded in the voice of one of the researchers, were designed to simulate a variety of people, such as a 鈥渇ashion persona鈥� or 鈥渂ooks and magazines persona鈥� for Amazon Alexa users. The researchers then used data disclosure requests to find out how the voice assistants had labelled these profiles and to examine their accuracy.
Amazon鈥檚 Alexa builds consumer profiles based on shopping and purchasing commands, whereas general questions don鈥檛 result in profiling, says co-author at the University of Southern California in Los Angeles. The voice assistant鈥檚 close integration with Amazon鈥檚 e-commerce platform made Alexa鈥檚 profiling 100 per cent accurate at labelling each persona the researchers had designed with that character鈥檚 interests.
鈥淪imilar to other Amazon services, we may use customers鈥� interactions with Alexa to provide more relevant ads,鈥� says an Amazon representative. They also described how customers can choose whether to receive interest-based ads by changing a single setting.
By comparison, Google Assistant delivered imperfect profiling. For example, it applied the 鈥渕arried鈥� label with 70 per cent accuracy and it identified users as having an advanced degree with 50 per cent accuracy. It also assigned initial profile labels such as 鈥渉omeowners鈥� or 鈥渘ot parents鈥� for new user accounts before any voice interactions occurred, says co-author , also at Northeastern University. A Google spokesperson says users can select the categories used to show ads or turn off personalised ads.
Only Apple鈥檚 Siri didn鈥檛 appear to profile users based on voice interactions 鈥� but Apple鈥檚 privacy policy reserves the right for the company to do so. 鈥淲e clearly state on our website and privacy pages that 鈥楽iri Data is not used to build a marketing profile, and is never sold to anyone鈥�,鈥� an Apple representative says.
The findings remain 鈥渕ostly speculative鈥� because such black-box experiments lack back-end access to the voice assistants, says at Clemson University in South Carolina. But it was helpful that the researchers confirmed that both Amazon and Google respect requests to opt out of profiling, says at the Electronic Frontier Foundation, a digital privacy non-profit organisation based in San Francisco, California.
Understanding how voice assistants work remains a challenge because tech companies keep changing and updating them, says , also at the University of Southern California. For example, in voice experiments with Google Assistant, the system鈥檚 profiling accuracy dropped by an average of 20 per cent between the spring of 2023 and summer of 2024 鈥� but the researchers don鈥檛 know why.
One major change is that some companies are upgrading their voice assistants with generative AI to improve their understanding of human language. 鈥淚f voice assistants improve, I鈥檇 guess more people will use them more often, possibly for more personal questions,鈥� says Klosowski. 鈥淎nd as those lines between our devices blur even more, it鈥檚 ripe for user confusion about what data is kept private and what isn鈥檛.鈥�
arXiv
Article amended on 30 September 2024
We added comment from a Google representative
