Hacker News Reader

Comment by simion314 | Hacker News Reader

Humans are to blame for using horrible AI/software. At least if you know your AI is that broken then have some human supervision.

View on HN

The real problem is when the task is no longer doable by Humans. Then the AIs have taken over and we can't turn them off. Imagine what happens when the AIs run our cars and food production, security, etc. We will be unable to turn them off and they will be unaccountable except to the people who set their policy which will be mired no doubt in layers of deep obscurity.

Or the problem where the developers will not know why the AI did something, say an AI decides to do some bad thing, the judge will ask the developer why and the dev will say: -I don't know, we feed them a ton of input and it worked fine on my machine and without test data so we shipped it

Google prefers to automate everything. Having human supervision would affect the bottomline so is out of the question.

Well, let's say it takes 30 minutes to fully check for copyright compliance 10 minutes of video. This means that the speed of one person checking videos is 1/3 unchecked-minutes per minute, or 1/3upm. 400 hours of content are uploaded every minute to youtube, which makes for 24000upm. You will need to employ 96000 people 24/7 to check all videos. Of course, they can't work 24/7, so a more realistic figure is 150000 people. This isn't menial work, it requires legal knowledge and familiarity with copyright law, so you're looking at hiring entry-level law clerks, which is about 16$/hr or about 33k$/yr. It will cost Google about 5bln$/yr to check all uploaded content manually. Which seems like a lot but is like 10-20% of Alphabet Inc.'s gross profit, so not actually impossible.

Are the only 2 choices (1) no human in the loop and (2) only humans in the loop?

Why not simply have humans review the videos that are automatically flagged? That would be a significant improvement over the current state of affairs, and would command a significantly more reasonable price.

Exactly, have the code extract the exact section of video that it considers problematic on left side and on the right side put the copyrighted work, the human will lose a few seconds to have a look.

Also I do not understand why Google does not use a reputation system, if a channel has a good reputation then take this in consideration in the automatic detection.

At the very least do it for major accounts (4k hours a year). It's pretty ridiculous that pulling in hundreds of dollars won't get you even a 30 second review.

The question is whether you are more worried about false positives or false negatives.

If you are only worried about false positives, then the humans only need to review the content flagged by the AI. This has 2 benefits: 1) the "starting number" would be much much lower than 400 hours of content per minute. 2) If a 10 minute video is obviously pirated, you don't need to watch the entire video to decide that.

There are some strong corrections that need to be made to your math.

1. Only a subset of uploaded videos need to be checked by humans, those that were flagged by the automated system.

2. Only a subset of the runtime of each flagged video needs to be checked. In other words, if a human decides within 12 seconds of a video that it is infringing, there is no reason to check the remaining portion of the video.

today in order to do a Google search I had to prove to Google that I wasn’t a robot. And even though I did it right the second time, they still made me do it a third time. I just remember sitting there going, Google, I actually am a human and your AI sucks so if I have to tap on storefronts and cars one more time to do a freaking Google search then I’m just using Bing.

It's partly checking that you're a human.

It's partly using you for free labour to train their next generation of SkyNet.

We've reached a point with CAPTCHAs (originally conceived as a way for humans to exert their will over machines) where the machines are now setting tasks for the humans, and the humans are carrying out the tasks without putting up too much of a fight.

That. CAPTCHAs are used as free labour. The only reason you don't see them everywhere is because people would stop using Google.

We've done it. We've cracked the code.

This thread is complaining about the fully automated process used by Google to check for copyright violations, and how infeasible it is to hire someone to check each video.

You're talking about how CAPTCHAs are now being used to get free labour.

> Prove you're human: Does the following video contain Beat It by Michael Jackson?

I am pretty sure they have considered it. Probably there is a legal hiccup laying here -- you either have reasonable suspicion that the material you show is copyrighted, or you acknowledge your algorithms and AI suck and you require crowd-sourcing. Kind of a lose-lose legally, and they would possibly introduce liability.

Recall the goal of Google is to avoid liability via the statement "we are pretty good and have really low false negatives", and have as few false positives as not to make the platform unusable.

Google does not care about content creators or propagation of knowledge (instructive videos) and ideas. Simple statement here. Google engineers might (at least the friends I have), but not as a whole organization.

But yeah crowd-sourcing is a good way to do a second sifting through results. I know it is being used, by Google and others, already to classify video sources.

Why do site owners use Google CAPTCHA? Even Cloudflare uses it. I resent being forced to help Google when not even on a Google site. I won’t put any of my stuff on Cloudflare specifically because they force people to help train Google’s image recognition.

Why do site owners use Google CAPTCHA?

Because it's free, reasonably hard to beat by bots and reasonably low-effort for the user if they let Google track them (enabling the single-checkbox captcha).

A few days back, to allow me to download the software I had just bought, the site kept putting up those damned pictures for me to click on. Street signs were the first three, I apparently got all of those wrong (even though I got them right), shop fronts was the next lot (that's a little harder to work through, I can't tell Chinese shop fronts from Chinese warehouses by text alone) and I had two of those, then more street signs, and finally cars.

It took longer to confirm I was human than it did to download the software. My partner came through to see if I was alright, after I was shouting abuse at the computer for an extended period. She sat through one of them with me, confirmed that I had it right, then swore and walked away when it decided I was wrong.

If I'd known that it was going to screw me over that much, I would have recorded it and put the video on YouTube.

Would that have infringed on Google's copyright?

Reply on news.ycombinator.com