Your reluctance to train the AI has been dully noted.

Is the note taker bored?

I'm kind of surprised that recaptchas are 'training AI' at google.

I went through a phase of trying to use tor just for the principle of it. So I was hit with a lot of recaptchas. And I really couldn't see how they were testing AI directly.

Usually, its the 'pick pictures of road signs' and 'pick pictures of cars' and so on. Almost all images seem to be taken from streetview cars, or possibly waymo cars.

So, the human doing the recatcha has to parse an image to extract the implicit depth information to pick out cars etc.

Whereas google know the exact location of the image so know where the roads are etc, and probably have lidar scan so they know the depth information etc too.

Picking a road sign out of a picture is kinda hard for a 2D image. Picking it out from a 4D scene created from all the different sensors their cars have, on the other hand, is a piece of cake.

Eh, I don’t think they’re training Waymo AI. They’re training their general image classification AI which doesn’t benefit from extra sensory data.

So if google want to train general image classification to pick out things, e.g. road signs:

google can use their 4D datasets to select a gazillion real still images that contain what they _know_ are road-signs or not road-signs or whatever.

And then they can feed these to their image classification stuff.

How does humans clicking on tiles help train their image classification AI?

I always deliberately try to mis-train it. But they can undoubtedly filter it out.

I can't prove it. But I have been hit by unsolvable captcha's when I purposely give it wrong answers.

It's hard to know. I generally try "well enough" and provide a couple of incorrect positives and negatives. I was forced to do about 8 yesterday, and I messed up all 8 slightly. It might depend how overtly you try to game it. I've tried to be subtle, but I don't know whether or not it's effective.

Back in the days of book digitising reCaptcha it was fairly easy to distinguish which word was from the book vs computer generated and only answer half the captcha and pass.

I always inputted the optional word with slight errors. A totally wrong answer would be easy to filter out, but one or two letters being wrong while it's still a proper word would hopefully be harder to get out of the training data.

I don't get your argument.

I get the original post's argument of not willing to be tracked by Google, but what hurts you so much in training Google's AI?

Well, mainly because Google is a 1 Trillion $ company and why should you help them in exchange for nothing? It's not like they're open sourcing all their data for others to use. They use and control whatever they can in their own interest.

You aren't the customer. The website is. They're providing a valuable service to the website. The website is providing a valuable service to you. You get to choose whether or not doing one unit of Google AI training work is worth using the services of the website. Generally speaking, I find that trade to be quite fair.

> You get to choose whether or not doing N units of Google AI training work where N measures how much Google hates you is worth using the services of the website.

FTFY

Well they do release tons of datasets,and their open source contributions have a lot of importance as well sinxe the past 10-15 years. Sure other companies do OS stuff too,i know. Even, the massive improvements in computational photography in the Pixel phones are explained by them in great detailed blog posts. They even built a special search engine especially for Datasets.

The AI you help train will be used to assist authoritarian governments (e.g. Dragonfly) imprison and murder millions of people so that Google can make a few extra bucks.

How do you see the car and traffic sign recognition AI being used to imprison and murder people?

Because the AI they're perfecting can and will be used for other purposes as well?

How do you see that data improving the AI for other purposes?

I think it's safe to assume that the code used to recognize cars and traffic signs can also be used to recognize other things, such as muslims, people using apps the government can't spy on, people wearing/holding objects that the government want to eradicate (perhaps a banned book). I think it's also safe to assume Google benefit immensely from having millions of people ready to test their changes in order to optimize their code and algorithms. I'm also sure that they do more than just check if you picked the "right" squares, such as storing every square you've ever chosen along with other stats such as completion time and correctness. I wouldn't be surprised if the Chinese government (as well as others) had interest in using such data to detect "rebels".

I would also be surprised if the work wouldn't be beneficial to other AI related work that can be used to build detailed profiles among other things.

Time and effort wasted. Work imposed on me by a hostile company with which I have no direct dealing and that financially doesn’t need this.

Agreed, AI is technological advancement, and Google makes most of its datasets available to everyone and goes great lengths to explain the technology it creates for those who are willing to read it. I don't understand what's the issue with semi-passively contributing to a potential technological advancemet.

This is irrelevant in case when I want to access a service of a third-party whom I'm paying to provide the service, or even worse - a service of public authorities.

> what's the issue with semi-passively contributing to a potential technological advancemet.

You're wastly overestimating your contribution to the technological advancement though this channel.

You're doing work for them for free. It's unpaid Mechanical Turking.

Devils advocate: you are getting access to a service that uses google's assistance for spam or abuse protection, not entirely unpaid.

No, the company that integrates it makes me pay (by doing work for google)

Again, worthy trade off, not much of a work to solve one recaptcha, and the service that you use remains available more reliably by filtering out spam bots that could potentially render it out of service otherwise.

Either way, you train ai and watch ads in exchange for content.