But they can figure out when a video has copyrighted music at 1% volume in the background.
The most horrid thought occurred to me while reading this, that somewhere within Google there may be a glossy (autogenerated or otherwise) report detailing exactly how many people are watching these videos and falling into this particular gradient of the recommendation model. Perhaps even read and passed up as too difficult to extract value from by their ads team. A company of this size (YouTube, not just Google) couldn't possibly be oblivious to this
Just advertise Spy Cameras, Thailand Vacation Packages, and Annual Passes to Disneyland, no ?
Detecting if there is a child in the video is more difficult, but do-able with current ML models. Now, however, determining why there is a child in the video - if it is a family "fail" video or pedo material, for example - via AI is about as impossible as trying to distinguish between satire, hate speech or propaganda via AI. It's not possible at all, as AI will for the near future totally lack context.
This distinction will require humans, and this is something not viable at all for fb, youtube, twitter & co, as it is a huge cost... the saving of which is offloaded to society though in form of e.g. undermined democracies or psychological trauma in sexual violence survivors.
And banning keywords won't help, that's just whack-a-mole'ing.
That is exactly the point, they have this ability and are not using it. Not even to disable comments.
People would be a lot more willing to accept blanket and risky experiments to solve these problems if there was a viable and reliable reconciliatory/redemptive path for when those systems get it wrong.
For better or worse, YouTube hasn't focused on scaling that part of their system, so there will be apprehension and scepticism with any approach they take.
It doesn't matter why the child is in the video. The pedophile only cares that there is a child. Videos uploaded with children in them will be exploited, regardless of whether they are legitimate.
I feel like you have missed the point. YouTube is actively helping people find this content and making money off of it.
The article claims that there is nudity (via timestamp comments). Separate models could be used to score the video, where a child plus any other potentially sexual content (genitalia, nipples) withholds the video. A withheld video can be appealed (for scenarios such as "breastfeeding tips" or what-have-you), much like copyright strikes, which have an existing process.
Obviously the model might have problems discerning between a 17 and 18 year old, but it would catch the most egregious content.
As long as child is not performing some sexually suggestive stuff on youtube video ("popsicle challenge", wtf) it is OK, probably.
The reason why they find licensed music in videos is because they have a financial (profit) motive for doing so. Find a licensed track, and get a few fractions of a penny every time someone watches that video. There is no financial incentive to blocking CP, so they don't bother. This is a real shame.
It can't detect context, however.
>The reason why they find licensed music in videos is because
It's technically easier to do so. ML is still new, and it still makes mistakes, this is why people complain about content ID because it takes down videos that are "fair use".
Explain to me, Mr Expert, how a ML algorithm can understand the nuance of the law, interpret it flawlessly like a human would (say, in a court of law).
Unless you're saying these cases no longer require humans, courts and judges, and ML is at a point where it's _flawless_ and accurate 100% of the time.
It's a shame you're so technically weak minded, I thought the Hackernews community would be educated in this regard.
Please go read about how ML works before making such comments in the future, they make you look really stupid.
You don't actually need to, at least initially. You only need to detect kids, and when you have some threshold of reports on comments on a video where the computer vision algorithm detected kids in the frame that rise to the level a human should take a look, the human can flag the video as containing elements of CP, which would kick in the following: - Deeper ML analysis of the dependency graph of commentors, related videos. - Blacklisting or law enforcement notification for commentors that are violating the law. - Training data that can be used to train more advanced algorithms.
Because there is no profit motive to do so, like there is with music licensing, they won't do it.
> It's technically easier to do so.
I'm not talking about determining context with ML. That is a hard problem, of course. But identifying "child human" with a computer vision algorithm is very simple, then relying on human moderators to initiate the dependency graph traversal.
> Please go read about how ML works before making such comments in the future, they make you look really stupid.
No need for the ad hominem. I'm not an ML expert, but I understand enough of the basics to know what is possible.
Even dental scans and x-rays can't give you a reliable chronological age for somebody [0].
In that context, I consider it extremely doubtful that some ML algorithm solved this whole problem by just using pictures/visual recording, that sounds a bit too much like magic/wishful thinking.
[0] https://www.newscientist.com/article/mg21428644-300-with-no-...