SEATTLE — Amazon’s new Alexa Answers program, which encourages unpaid volunteers to fill in gaps in its databases to better respond to user queries, raises a host of questions about information quality and transparency in the fast-evolving voice computing era.
Many people make purchases, control home automation systems, play music and ask questions using the Alexa system — software that runs in the cloud and on microphone-equipped devices from Amazon and other hardware makers.
Alexa-enabled devices number more than 100 million and are being pushed into homes, workplaces hospitals and hotel rooms as Amazon competes with Google, Apple and other technology companies in voice-controlled computing. Amazon will unveil new services and devices with Alexa later this month.
Alexa Answers allows anyone who sets up an account to answer questions posed repeatedly by other Alexa users. The answers are limited to 300 characters and are typed into a website, not spoken into an Alexa device.
The program, rolled out broadly last week, began as a limited test last year in a bid by Amazon to provide answers that weren’t present in its hundreds of Alexa data sources, which include Wikipedia, Stats.com and Yelp.
It’s an area where Alexa, the market leader in voice computing, suffers in comparison with second-place Google Assistant, which has a deeper well of answers to draw from thanks to its dominant Internet search engine. Google spokesperson Ashley Thompson said the company does not utilize users to answer questions posed to Google Assistant.
The tech industry abounds with cautionary tales of crowdsourced question-and-answer programs providing dubious or manipulated results, intentional misinformation campaigns on social media and user-trained chatbots that quickly went off the rails.
Alexa executive Bill Barton told Fast Company that Amazon uses a mix of machine learning software and human editors to ensure the quality and appropriateness of responses provided by users through Alexa Answers.
An Amazon spokesperson said certain topics are off-limits to Alexa Answers, including politics, but declined to elaborate on other topics that would be filtered out.
Barton also said the company is “leaning into the positive energy and good faith of the contributors,” who will be awarded digital points and badges for their work in answering Amazon customer questions.
The points have no apparent value and the contributors will not be compensated. (They’re not routinely compensated on question-and-answer sites, though Quora experimented with financial rewards sponsored by companies in 2016.) Amazon operates a separate online service called Mechanical Turk in which people can earn money for performing small digital tasks for third parties, such as answering questions.
Amazon itself presaged the Alexa Answers concept in a 2018 Super Bowl commercial starring Jeff Bezos and a handful of celebrities who fill in for Alexa to respond to user queries. In the spot, a boy asks, “Alexa, how far is Mars?” Singer Cardi B answers, “How far is Mars? Well how am I supposed to know? I never been there.”
That answer, presumably, would not be presented through the new program, which highlights responses that have been up-voted by other users, though the company won’t specify how many up-votes a response needs to get before it’s presented to Alexa users.
Users who want to assess the credibility of responses provided by Alexa Answers can go to the program’s website and see how other users have rated the answers. But there is no process for providing citations, and it’s up to the individuals giving the answers to decide if they want to provide their identity.
In the commercial, Cardi B answered in her own voice, but the answers from other users are presented in Alexa’s default, inalterably female voice, albeit prefaced with the phrase, “according to an Alexa customer.” In short, all wisdom received from Alexa — whether it be from the crowd or another source — sounds the same.
The company spokesperson said there’s currently no way to turn off responses provided by Alexa Answers.
