Hacker News Reader

Comment by noodlesUK | Hacker News Reader

I’m genuinely concerned about data privacy, but I’m kinda surprised about the outcry about all of this. I guess it’s relatively common knowledge for hacker news readers that all of these systems need to have human intervention to improve, but I feel like all this could have been avoided by just explaining that this was how the service was able to improve... I’m sure more steps could be taken to anonymise data from the people doing the transcripts, but I feel the best part of this is to improve the wake word detection to reduce the false positive rate.

View on HN

I think the outcry is mostly:

1.) Apple literally plastered "What happens on your iPhone stays on your iPhone" ads around CES to make fun of competitors. - https://www.macworld.com/article/3331597/apple-privacy-billb...

2.) Apple always slams competitors over their lack of privacy on every keynote.

I think the outcry is more than warranted when the company is blatantly lying to your face in their marketing.

Only that Siri doesn't happen on the iPhone, and it's easy to prove: try using Siri with aeroplane mode on.

The audio to text translation does happen on the iPhone. Siri is a service, yes a network is required to access the service.

Nope, it does not. Only the trigger phrase (“Hey Siri“) is recognised locally on the device (according to Apple), everything that comes after that is sent as audio recording snippet to Apple servers where the actual speech recognition happens.

There are rumours [0] that Apple might enable local speech-to-text at some point, but as of now, that’s not how it works.

[0] https://appleinsider.com/articles/18/11/15/apple-considering...

https://support.apple.com/en-us/HT208343

> You can use dictation instead of your keyboard to enter text with many apps and features that use the iOS keyboard.

> On iPhone 6s or later, and iPad, you can use dictation without being connected to the Internet. Earlier models of iPhone and iPad require an Internet connection.

Sadly, offline dictation is a separate feature, it is not the speech-recognition used for Siri.

Quote directly from Apple:

"The recording of the user’s spoken words is sent to Apple’s voice recognition server. If the task involves dictation only, the recognized text is sent back to the device. "

Source:

Page 69 in https://www.apple.com/business/site/docs/iOS_Security_Guide....

I just put my iPhone in airplane mode and used speech to text to enter a full paragraph in Notes.

Siri is the service. Speech to text, while it can be improved with network access, works fine without it on current iOS.

It is not at all a given that machine learning systems need covert review of user data. The training data can come from paid users or explicitly opted in users. Do not normalize this behavior, it is not normal, OK, or required.

Alternatively, the users could choose to annotate their own data from time to time I suppose. Maybe there’s some way to incentivise this.

> I’m kinda surprised about the outcry about all of this. I guess it’s relatively common knowledge for hacker news readers that all of these systems need to have human intervention to improve

There are plausible ways of doing this ethically and maybe many people thought it was being done that way.

(Eg prompting a subset of users for opt-in to a paid or volunteered time limited study )

I mostly agree, however, there are some issues I'm foreseeing. I've done some NLP work myself and have had access to very private information because of it. The team we were working with were all highly educated and well paid. Because of this, we were aware of the implications of the data we were working with, so at one point, in one of our randomly selected data sets, we found data that was from a public figure and removed it to avoid any possibility of a conflict of interest.

Where my concern is, is that a lot of annotating work is now being done by low wage, low education workers and because of the increasing demand of annotated data sets this group is increasing. Also, because this work is increasingly getting outsourced, there is less direct control of who is doing the work and therefore the chance that a bad apple slips through is bigger. That's what's scaring me going forward.

i’d wager that in most cases “low wage, low education” workers have much better morals than your “highly educated well paid” team.

Why do you think that?

power corrupts.

I'd probably guess the "gap" of morals isn't that large between the middle class the GP is talking about compared to the low class.

The problem isn't a single-value metric of morals, whatever that may be, the problem is that one group is bored and not invested in the project while the other has far more exciting things to do than wondering who they might be listening to.

Thanks, that is exactly what concerns me.

Maybe, I'm just making an observation. My biggest issue is that lack of control over outsourced workers over internal workers.

In my situation it was very clear that breaching the agreements would not just lead to direct end of contract, but also possible legal steps and a permanent record, leading to a much smaller chance for a similar job, for which I've specifically had an education. For uneducated workers, this if often less of an issue, especially for off-shore workers.

I also don't think low wage/education workers have less morals, but as the workforce gets bigger the chance of bad apples also gets bigger. I wrote exactly that in my previous comment.

Assuming wage is a proxy for power, I would bet they just haven’t had a chance to excercise their morals yet. There’s a saying about power corrupting.

I'd wager "low wage, low education" workers are more likely to be simply acting immorally in different ways.

The "high education well paid" team is more likely to do something with far reaching harmful impact whereas the "low wage, low education" worker doesn't have the power to do anything at that scale.

Do low wage, low education workers have poorer morals? They are fighting to make ends meet and could care less about your 1st world problems.

The real moral gap is between the very rich and everyone else:

https://phys.org/news/2019-03-rich-parents-unethical.html

Educated or not you can't 'unhear' something. What if the conversation is political or financial and has major implications on someone you know? Will you still be able to remain professional and not act on the information?

I feel like devices like this should at least give the user the opportunity to play back whatever is sent for manual evaluation. e.g. it could send you an email at the end of the month listing the recordings it would like to use.

The outcry is over the leaks from the Dutch contractor that worked for google to the media.

He was acting as a whistleblower because he was alarmed at the amount of data he was asked to review that was just people talking privately in their homes (not intentionally interacting with their smart devices)

Reply on news.ycombinator.com