Hacker News Reader

Comment by DangerousPie | Hacker News Reader

Important to note that this is not 10% of all messages being falsely flagged (= 10% false positive rate), but 10% of flagged messages being false positives (= 90% precision). As someone who works with these types of classification problems in a different context, 90% precision is actually quite good - especially assuming there is some sort of manual review process to take care of the 10%.

Whether that makes this whole plan a good idea or not is obviously a very different question, but I think it's important to be clear about what this number actually means.

View on HN

Everybody here seems to be saying "the EU commision doesn't have any idea what they're doing", but it seems barely anyone understands what the 10% is referring to.

However, I hardly doubt that they will actually adapt their recall to actually ensure a 90% precision based on court evaluation. I think this is all handwaving and throwing around numbers which as described won't be part of any legal act.

It's easy to achieve high precision when you just define a hit very broadly. See also the political flap about teachers factually describing the existence of alternative sexuality as 'grooming'.

If you want to see the lie behind any of these child protection surveillance initiatives when they talk about things like 90% precision or millions of hits per year ask them how many of those detection of vile child predators resulted in an arrest warrant -- not even an actual arrest, or an actual conviction, but just an attempt.

The answers is extraordinarily few and that tells you everything that you need to know.

> It's easy to achieve high precision when you just define a hit very broadly.

No, it's easy to achieve high recall when you define a hit very broadly. Precision will come down starkly.

Code is law the algorithm is always right 90% of the time.

I don't mean in the algorithm itself, I mean in your evaluation of the algorithm, where you also don't do so equivalently for recall (or don't report on recall).

Evaluate your algorithm thusly: If it made a hit, it's a grooming true positive unless its extraordinary undeniably a false positive. Absent any ground truth data you just don't evaluate recall, of if you have any test data it's only a false-negative if it's undeniably abuse. Benefit of doubt always goes to the algorithm. All hail the algorithm. All hail.

For a moment, consider that you're running a company that supports some form of user communication.

How much time and money are you willing to risk in a gray area considering all the legalese? Google doesn't seem to be willing whatsoever: https://archive.ph/W41mf

> 90% precision is actually quite good

It is good from the perspective of comparing this to other ML models. It is not good from a real world perspective.

Is it really good? What's the rate of false negatives? I could build a search engines that searches only one specific text classified as grooming. I wouldn't find much of the other grooming in chats but if I find a positive it's with a pretty high precision.

https://archive.ph/W41mf

At present the manual review process consists of "fuck you"

What that means is having some random person going over the nudes your kid sent to her boyfriend.

I wonder what kind of person would take a job like that…

Reply on news.ycombinator.com