Amazon, Google follow Apple's lead on voice assistant review policies
Following Apple's decision to temporarily halt Siri grading as it evaluates the program's privacy safeguards, Amazon and Google this week followed suit and updated their respective policies on human reviews of recorded voice assistant audio.
Apple on Thursday suspended its Siri grading program, which seeks to make the virtual assistant more accurate by having workers review snippets of recorded audio, after a contractor raised privacy concerns about the quality control process.
Now, Apple's competitors in the space, namely Google and Amazon, are making similar moves to address criticism about their own audio review policies.
Shortly after Apple's announcement, Google in a statement to Ars Technica on Friday said it, too, halted a global initiative to review Google Assistant audio. Like Siri grading, Google's process runs audio clips by human operators to enhance system accuracy.
Unlike Apple's Siri situation, however, a contractor at one of Google's international review centers leaked 1,000 recordings to VRT NWS, a news organization in Belgium. In a subsequent report in July, the publication claimed it was able to identify people from the audio clips, adding that a number of snippets were of "conversations that should never have been recorded and during which the command 'OK Google' was clearly not given."
The VRT leak prompted German authorities to investigate Google's review program and level a three-month ban on voice recording transcripts.
"Shortly after we learned about the leaking of confidential Dutch audio data, we paused language reviews of the Assistant to investigate. This paused reviews globally," Google told Ars Technica.
Google did not divulge the halt to global reviews until Friday.
Amazon is also taking steps to temper negative press about its privacy practices and on Friday rolled out a new Alexa option that allows users to opt out of human reviews of audio recordings, Bloomberg reports. Enabling the feature in the Alexa app excludes recorded audio snippets from analysis.
"We take customer privacy seriously and continuously review our practices and procedures," an Amazon spokeswoman said. "We'll also be updating information we provide to customers to make our practices more clear."
Amazon came under fire in April after a report revealed the company records, transcribes and annotates audio clips recorded by Echo devices in an effort to train its Alexa assistant.
While it may come as a surprise to some, human analysis of voice assistant accuracy is common practice in the industry; it is up to tech companies to anonymize and protect that data to preserve customer privacy.
Apple's method is outlined in a security white paper (PDF link) that notes the company ingests voice recordings, strips them of identifiable information, assigns a random device identifier and saves the data for six months, over which time the system can tap into the information for learning purposes. Following the six-month period, the identifier is erased and the clip is saved "for use by Apple in improving and developing Siri for up to two years."
Apple does not explicitly mention the possibility of manual review by human contractors or employees, nor does it currently offer an option for Siri users to opt out of the program. The company will address the latter issue in a future software update.
Apple on Thursday suspended its Siri grading program, which seeks to make the virtual assistant more accurate by having workers review snippets of recorded audio, after a contractor raised privacy concerns about the quality control process.
Now, Apple's competitors in the space, namely Google and Amazon, are making similar moves to address criticism about their own audio review policies.
Shortly after Apple's announcement, Google in a statement to Ars Technica on Friday said it, too, halted a global initiative to review Google Assistant audio. Like Siri grading, Google's process runs audio clips by human operators to enhance system accuracy.
Unlike Apple's Siri situation, however, a contractor at one of Google's international review centers leaked 1,000 recordings to VRT NWS, a news organization in Belgium. In a subsequent report in July, the publication claimed it was able to identify people from the audio clips, adding that a number of snippets were of "conversations that should never have been recorded and during which the command 'OK Google' was clearly not given."
The VRT leak prompted German authorities to investigate Google's review program and level a three-month ban on voice recording transcripts.
"Shortly after we learned about the leaking of confidential Dutch audio data, we paused language reviews of the Assistant to investigate. This paused reviews globally," Google told Ars Technica.
Google did not divulge the halt to global reviews until Friday.
Amazon is also taking steps to temper negative press about its privacy practices and on Friday rolled out a new Alexa option that allows users to opt out of human reviews of audio recordings, Bloomberg reports. Enabling the feature in the Alexa app excludes recorded audio snippets from analysis.
"We take customer privacy seriously and continuously review our practices and procedures," an Amazon spokeswoman said. "We'll also be updating information we provide to customers to make our practices more clear."
Amazon came under fire in April after a report revealed the company records, transcribes and annotates audio clips recorded by Echo devices in an effort to train its Alexa assistant.
While it may come as a surprise to some, human analysis of voice assistant accuracy is common practice in the industry; it is up to tech companies to anonymize and protect that data to preserve customer privacy.
Apple's method is outlined in a security white paper (PDF link) that notes the company ingests voice recordings, strips them of identifiable information, assigns a random device identifier and saves the data for six months, over which time the system can tap into the information for learning purposes. Following the six-month period, the identifier is erased and the clip is saved "for use by Apple in improving and developing Siri for up to two years."
Apple does not explicitly mention the possibility of manual review by human contractors or employees, nor does it currently offer an option for Siri users to opt out of the program. The company will address the latter issue in a future software update.
Comments
"Google did not divulge the halt to global reviews until Friday" is incorrect.
The Germans on Thursday announced Google had previously paused their program and additionally would now be restricted from using it for the next three months by the German Data Protection Commissioner. In the same statement the German agency recommended that Apple and Amazon follow Google's lead.
"In a statement released Thursday, (and due to time zone differences I believe Wednesday here in the US) Germany’s data protection commissioner said the country was investigating that contractors listen to audio captured by Google’s AI-powered Assistant to improve speech recognition. In the process, according to the reports, contractors found themselves listening to conversations accidentally recorded by products like the Google Home. In the statement, the German regulator writes that other speech assistant providers, including Apple and Amazon, are “invited” to “swiftly review” their policies.
Nothing may come of it but common business practices may not be acceptable going forward, and certainly more transparency required.
Of some note Ireland had already been looking at Apple and Siri due to a citizen complaint but Apple's response to them at the time was no disclosure was necessary since the recordings were anonymised. That defense tactic has probably now changed.
Interesting that you're so quick to defend Google that you'd make a careless error like that.
Also unchanged: Apple among the three companies was the only one that was always anonymizing its voice clips before all this controversy even started, as per their white paper. Anonymizing is not 100 percent foolproof against identifying someone (for example, they identify themselves in the recording, or its obviously a famous person with a distinctive voice etc), but it was and is better than what was previously the policy at Google and Amazon, which left identifying information intact.
Good to see the others doing the same. But also a bit unnerving to see that what Apple was doing in this regard was qualitatively no different from what Google and Facebook were.
'grading' sounds like 'culling' unbelievable that this term is used, wow.
For example The Verge reported Thursday that Google had voluntarily stopped the reviews at some earlier date which could have been weeks or days earlier, and not specifying whether in just the EU (it was never just Germany) or worldwide. Common sense would say worldwide since it was a contractor problem that needed to be addressed not a regional one.
Ars didn't report the story until Friday and added the word "globally", which still doesn't show "Google and Amazon following Apple's lead". Factually it was the German authorities telling Apple and Amazon they should seriously consider following Google's lead. I assume you read what was announced. Saying Google didn't divulge the program pause until Friday is a half-truth at best IMO.
No that is not defending Google, it's pointing out that it had nada to do with anything Apple chose to do with their program. Rather than Apple being proactive and everyone else following their lead Apple for their part was forced into it to avoid a formal inquiry by yet another country's data privacy commissioner IMO, tho it's reasonable to assume Apple might have done so of their own volition at some point.
Also Google was not supplying voice snippets for review that were not anonymized. They did not "leave identifying information intact" for contractors. What we don't know is whether Apple was more similar to Amazon and including things like location, gender and device ID for instance, not that it would make those recordings as presented personally identifiable anyway. The leak implies some form of identifying data was attached but Apple has volunteered very little information, nor has Google for that matter. They are both pretty vague and not exactly forthcoming, avoiding comment unless pushed into it.
Even from the article: "Shortly after we learned about the leaking of confidential Dutch audio data, we paused language reviews of the Assistant to investigate. This paused reviews globally," Google told Ars Technica. AI changed the context of that quote when they changed the attribution.
The quote from the Ars article actually reads:
"Shortly after we learned about the leaking of confidential Dutch audio data, we paused language reviews of the Assistant to investigate. This paused reviews globally," Google told Ars today. They had already paused the reviews when they spoke to Ars on Friday.
You're wrong about the claim that Apple was the only one anonymizing the voice clips. Google was anonymizing clips that were listened to as well. Google also made the storing of audio clips opt-in and even if you do you can opt-out at any time. If you've opted-in, you can set your account to auto-delete every 3 or 18 months. You can also manually delete them at any time. From the earlier Ars article on 11 Jul https://arstechnica.com/information-technology/2019/07/google-defends-listening-to-ok-google-queries-after-voice-recordings-leak/
In case you misunderstand my intent, I'm not defending Google. I am countering your misinformation.
There almost appears to be a concerted spin effort since in general each of the articles these past days here at AI concerning Apple's review program have contained bits of misinformation, inaccuracies or misleading statements, too many to be coincidental IMO but perhaps they are just simple mistakes. I know it's not typical of AI writers to make such obvious ones though.
As many have said, there are very legitimate reasons for human listening. There just has to be improvement. First thing, they need to get rid of the 3rd party outsourcing and use in-house employees only. Yeah, it's going to cost more for them but none of these companies are hurting for a dollar. Second, implement controls in the work environment that make stealing data a difficult process. Third, in plain language inform customers there's a chance their interactions can be recorded and used for blah, blah, blah. Third B - a pop up that makes participation OPT-IN not opt out.
Not high priority, but I think they should all use raw number when reporting to the public. Vague "less than 1%" or "approximately 0.2%" doesn't really paint an illuminating picture for customers using these services. What raw number represents less than 1% of interactions per day x 365 days per year? 1000 per day? 100,000? A million?
Baaaaa!!!! 🐑
You didn't actually read the article did you?
Google stopped human reviews sometime back, shortly after the initial stories that a contractor had stolen several hundred recordings and given them to a news organization. If anything be surprised that Apple wasn't more forthcoming that they were doing the same thing, but would stop doing so and review the program. That would have been proactive.
It's not as tho Amazon and Google had already been "discovered" and received negative press, while Apple remained silent and carried on.
There almost appears to be a concerted spin effort on your parts. You certainly never asked why they waited so long to "confirm"....
Perhaps more accurately the AI author didn't know until Friday, that's possible. Many others already knew. The article as written is misleading IMHO.
In truth it was Apple and Amazon following Google's lead, with a little prodding from the Germans.
As for not wondering why Google took so long to discuss the review pause it might be the same reason why Apple chose not to say a word about doing the same kind of human reviews and continue doing so after the negative press surrounding Google and Amazon being discovered. The less said the better. Say nothing and admit to nothing unless pressured to do so.
I have to agree, I think the headline is Gaitorbait, mainly because it says something different to what is written in the article.
But
This appeared on the 1st August, which I think was the Thursday.
Still, there seems to be a few details that have been lost in the media rush to attract clicks.
Google suffered a data breach.
Apple did not.
In Google's case, "thousands of recordings" were leaked to the press.
https://arstechnica.com/information-technology/2019/07/google-defends-listening-to-ok-google-queries-after-voice-recordings-leak/
In Apple's case, a contractor claimed to have heard people having sex and people making drug deals overheard over the sound of a car engine. But as far as I've read, he did not provide any recordings to back up his claims.
https://arstechnica.com/gadgets/2019/07/siri-records-fights-doctors-appointments-and-sex-and-contractors-hear-it/
Loath as I am to accept any claim without evidence, I'm going to assume he's an upstanding citizen and that he's talking to the press out of innate sense of altruism, though that still doesn't really answer the question as to why he didn't provide any evidence.
He also claims that accidental activations on the Apple watch are alarmingly regular. This doesn't surprise me since Siri on the Apple watch can be activated by raising your wrist. Not sure that counts as an accidental activation if you've deliberately set the watch to behave that way.
Others seem to think that the problem is hiring outside contractors. Again, I'm not sure that going completely in-house would fix the problem. I have seen little evidence that Apple employees are less likely to leak information than contractors.
https://www.businessinsider.com/apple-fires-iphone-x-engineer-viral-video-2017-10?r=US&IR=T
Past behaviour is no guarantee of future conduct.
So no recordings was leaked. Is Apple hiring a better class of contractor, or does Apple have a system in place that ties the contractor to a recording, making capturing and releasing them to the press a much more risky proposition?
I'm also guessing that they will continue to store these snippets (with the user's explicit consent this time) in the future, but right now they're looking at ways to make it less likely these recordings are leaked.
Not allowing customers to opt out? C'mon Apple! Even Google allowed this! (though Google has been known to ignore what their users have requested, and run roughshod over their privacy regardless).
If Apple holds itself to a higher privacy standard than Google, then I think Apple's more knowledgeable customers should do the same.
…
…
Yeah, okay, I think I see what happened there …