Ask HN: Why are we expected to train Google's NN for free through captchas?
I keep seeing the captchas more and more whose obvious purpose is to train their self-driving NN, and they're getting out of hand. Sometimes I have to work through 5+ images. How is that legal that they can just interrupt me on a whim and ask for me to do work for them? Is there an alternative solution available?
60 comments
[ 3.1 ms ] story [ 118 ms ] threadI think it's awful we've allowed Google to trick us into placing trackers on every site in existence that also force us to do free work for them. There has to be a better solution.
That said, I'd expect a large chunk of users to use their browser's password manager, if nothing else.
Probably something the website opted into. No clue.
[1]: https://nearcyan.com/you-probably-dont-need-recaptcha/
I'm sure they didn't care, or just thought I was tilting at windmills, but if I'm one of many that cancel for that reason, the service might understand it's costing them money.
BTW I'm happy people aren't spending their time on meaningless tasks, at least it's useful for something.
Or did you mean the 30 seconds-1 minute of time to solve the captcha (assuming it doesn't infinitely keep going)?
No bigger middle finger than an indefinite unsolvable captcha!
Blame the pages you go to.
As a consumer, it’s your choice to not support them if you feel it has become too burdensome to fill out the captcha.
The legality of this shouldn’t even be a question. No one is forcing you to use these sites.
Now you're opening up a bigger question of why are these massive companies that have hit on a basically infinite source of money allowed to falsely subsidise non-related industries and killing off any potential of innovation there by offering things for below-cost/free?
Also, come to think of it, I'd like some assurance that I'm not being bothered more if I'm using privacy enhancing tech and basically being corralled into behaving a certain way.
This is why we have consumer protection laws
What's worse is that ReCaptcha isn't an honest CAPTCHA. Instead, ReCaptcha relies in large part on building a tracking profile. If you take privacy enhancing measures, such as clearing cookies and using a VPN, then you can't pass ReCaptcha. It will present you with a endless loop of "please try again" no matter your answer.
maybe there is something in your particular setup that is breaking it.
This doesn't harm consumers. They are very slightly inconvenienced, likely less-so than if websites were forced to use other methods to achieve the same thing.
That argument would work if government sites didn't use reCAPTCHA in their online services.
Yes, you could probably complete the process by phone but thats akin to telling people to boycott Facebook when its the place to find out about events and participate in community discussions.
reCAPTCHA is EVERYWHERE. There are no viable alternatives. Asking people to boycott reCAPTCHA is asking them to boycott a huge chunk of the net.
Besides, this 'think of all the poor website operators' argument is bunk. Bots are going to happen whether the operator likes it or not. Better to lean in than out.
https://www.hcaptcha.com/
In a way, if a website is doing this instead of ads, it's better, still, getting to choose how I repay a website for content, I would prefer more control over my form of payment. Less acceptable cases exist as well, I ended up cancelling auto-payment of a service which recently started putting up an hCaptcha for every login attempt.
Here's the relevant section from their website:
hCaptcha is designed to solve the most labor intensive problem in machine learning: labeling massive amounts of data in a timely, affordable, and reliable way.
More data generally produces better results in training machine learning models. The recent success of deep models has led to increasingly large datasets, almost always with some human review. However, creating large human-reviewed datasets via Mechanical Turk, Figure Eight, etc. is both slow and expensive.
hCaptcha allows websites to make money serving this demand while blocking bots and other forms of abuse.
Powered by HUMAN Protocol
The hCaptcha marketplace is powered by the HUMAN Protocol, an open decentralized protocol for human review that runs on the Ethereum blockchain. Websites earn Human Tokens (HMT) whenever users use the hCaptcha widget on their site, and machine learning companies pay Human Tokens to get their data labeled.
I have not seen a captcha in months if not longer, I don't remember doing one anytime recently.
I have browser and network level ad and malware blockers in place and sometimes come upon sites I cannot access at all, but never unusable ones due to no captcha.
But, I also have a Google account linked to chrome that stays logged in, and I wonder if that allows me to avoid them.
Asking the web developers on here, how difficult would it be to make something like that? Or does that already exist?
The image recognition task has not changed for a very long time. How many captchas do you think have been solved by users in the last 5 years, all the same kind of recognition problems on the same kind of data? Trillions? It would be way past any kind of diminishing returns for this task to get more labels. If there was value there, they would have at least switched to a different kind of task.
In addition to that they're trying to get sites to move to Recaptcha v3, which does not have the user solve any kind of. Why would that be if the answers had any kind of value?
The fundamental problem with using captchas for training is that it devalues their value as a tool for security, since it sets up conflicting incentives. And the latter is what Google is selling. (Yes, selling. Recaptcha Enterprise costs money to use at scale, and the list price seems to be $1 / 1000 calls). The people paying for that service are using it to block attackers and let good users through, not to challenge the good users just to get some more labeling data from them.
sure we have models that can do it running on server on static images, but the problem they are trying to solve is not street view, the problem is training ML models capable of doing this in real time in an autonomous car.
taking a step back, identifying things in images is a problem very complex for a computer, best way in this day and age is to use ML but you need to train it to do it, for that you use a bank of know images and that is what they use reCaptcha for, to built those image banks that they will later use to train and evaluate ML models.
Second, they can use reCaptcha for that and still provide all those things that you mentioned.
another thing, they are not continuously using the same set trillions of times, they are increasing the image set over time, and this is what they use reCaptcha for to classify those new images, they dump new images in the set and every time you see a reCaptcha it will show you a few know images and a new unknown one that need to be classified.
Your input alone say nothing about that image, but after hundreds of peoples saying that, for example, there is a traffic light in a image while getting all the other know traffic light images right indicate there is a big likelihood of that picture containing a traffic light, once they have a big confidence on the image classification after hundreds or even thousands of captcha they move it to the know images set.