Just how is Alexa listening?
“They are not actively listening to their user,” said Umar Iqbal, assistant professor of computer science and engineering in the McKelvey School of Engineering at Washington University in St. Louis. “They only activate when users give very specific commands, such as ‘Hey Alexa’, in case of Alexa.”
Iqbal has previously studied how Amazon uses smart speaker interaction data to infer user interests and then use those interests to target personalized ads to the user.
“I focus on AI security and privacy, and general privacy concerns on the internet – how people are tracked, how the data is used and what people can do to protect against it,” he explained.
Iqbal examined the process of how the Amazon smart speaker Echo records user voice commands, saving them to the company’s cloud, to train its AI assistant Alexa.
In March, Amazon ended a privacy feature that lets some users of its Echo smart speaker prevent their voice commands from going to the company’s cloud computing centers for processing. Amazon said it made the decision to “no longer support this feature” as it expands Alexa’s capabilities with generative artificial intelligence features that rely on being processed in the cloud. While the change may sound alarming to some privacy-minded users, it was not widely used or available for all devices anyway.
“I think the general fundamental issue is the lack of transparency,” Iqbal said. “Users don’t know how the data is used.”
Iqbal said his assessment of Amazon being less open about the fate of those recorded Alexa requests, only adds to fears of the technology.
“If they (users) had that clarity, a lot of their concerns might be eliminated,” he said. “I would not say that the users are used to not having privacy, it’s just that when they are reminded, they are shocked.”
Iqbal described Alexa as similar to chatbots such as ChatGPT. He said Alexa is using large-language models to process information, like a voice-based chatbot. And Iqbal said there are other ways to use large-language models that give users more autonomy over the fate of their data.
“There are options that are more user- and privacy-friendly,” he said.
As AI assistants begin taking on more tasks for consumers, like planning vacations and scheduling work events, Iqbal said the threat to data security increases as the various algorithms swap user information and potentially open themselves to data manipulation. Iqbal and colleagues at McKelvey Engineering have developed tools to counter these threats. One such method, known as “IsolateGPT”, keeps external tools isolated from one another while still running in the system, allowing the AI assistants to do their job while also keeping user data secure.
