Read on Twitter
So I have a (not really) shocking revelation.
@twitter @twittersafety you have failed the Turing test.
I can now prove that reporting a tweet is almost, but not entirely a waste of time.
1/n
@twitter @twittersafety you have failed the Turing test.
I can now prove that reporting a tweet is almost, but not entirely a waste of time.
1/n
Let's consider a vile tweet. Remember that lovely piece of trash I told you about this morning? Here's one of her followers responding:
"Sir, if you had any clue as to what it’s really like being a woman, then you’d understand this."
Cont'd.
2/n
"Sir, if you had any clue as to what it’s really like being a woman, then you’d understand this."
Cont'd.
2/n
"But you don’t. Because you’re a man living in a self-made world of delusion, trying to lecture women on how pregnancy after rape isn’t worse than being raped.
Seriously, FVCK OFF."
(I changed the spelling there to avoid twitter punishing me, see below).
3/n
Seriously, FVCK OFF."
(I changed the spelling there to avoid twitter punishing me, see below).
3/n
What is a Turing test? You may be familiar with "CAPTCHA" - the Completely Automated Public Turing test to tell Computers and Humans Apart.
The idea is simple. Computers are logical machines, and are inherently incapable of intuitively telling computer input from human input.
4
The idea is simple. Computers are logical machines, and are inherently incapable of intuitively telling computer input from human input.
4
There are algorithms, which can make probabilistic guesses based on certain data, but the top level problem is to a machine, data from a human and data from a computer are indistinguishable.
5/n
5/n
CAPTCHA tests typically rely on data interpretation. For example, poorly displayed text that we humans can figure out, but to a machine there is no discernible pattern. Thus, if you get it right, there's a high probability you're a human.
6/n
6/n
You've probably seen that variation that says "Click on all squares with traffic lights". The test here is not whether you get it right, but *how* you get it right. A single misclick doesn't count against you, but 50% success rate? Nah. That's a machine using random numbers.
7/n
7/n
Timing is a factor too. If you're clicking too fast, you're probably a computer algorithm trying to work it out. Humans don't move that fast.
The point here is, there is no 100% infallible way for a computer to figure out if you're a human, or a machine spouting garbage.
8/n
The point here is, there is no 100% infallible way for a computer to figure out if you're a human, or a machine spouting garbage.
8/n
And so we come to twitter failing the Turing test. Remember that vile tweet? Here's twitter's official response:
"We’re writing to let you know that after reviewing the available information, we didn’t find a violation of our rules in the content you reported."
9/n
"We’re writing to let you know that after reviewing the available information, we didn’t find a violation of our rules in the content you reported."
9/n
Now, anyone with half a brain knows this is a lie. Twitter's rules are very clear: no hateful conduct. This is clearly hateful conduct. Hell I've been suspended for 12 hours just for telling someone in caps to fvck off, and that was classified as hateful.
10/n
10/n
So I tried an experiment.
If you reply to their answer, then they "review it a second time". Or do they?
This was my response:
11/n
If you reply to their answer, then they "review it a second time". Or do they?
This was my response:
11/n
"Hello,
Hubcap space bar melon rabbit? Insecticide muffin cello veranda cease!"
As you can see, this is gibberish. It has proper sentence structure, punctuation, etc. Even has a greeting. But the words are random and make no sense.
12/n
Hubcap space bar melon rabbit? Insecticide muffin cello veranda cease!"
As you can see, this is gibberish. It has proper sentence structure, punctuation, etc. Even has a greeting. But the words are random and make no sense.
12/n
What you would expect here is something like: "I'm sorry? We didn't understand this." or perhaps ignoring the message completely.
I think you know what actually happened.
13/n
I think you know what actually happened.
13/n
"Hello,
We appreciate all your help with this report. We’ve reviewed again, but we couldn’t find a violation of our rules."
That's right, "twitter" responded as if I'd said something like "You can't be serious?" or "Can you check again? I think you got it wrong."
14/n
We appreciate all your help with this report. We’ve reviewed again, but we couldn’t find a violation of our rules."
That's right, "twitter" responded as if I'd said something like "You can't be serious?" or "Can you check again? I think you got it wrong."
14/n
Twitter failed the Turing test. Their email processor failed to distinguish between human data, and what would feasibly be randomized computer data.
So here's my conclusion: When you report a tweet, it goes into an algorithmic analyzer queue.
15/n
So here's my conclusion: When you report a tweet, it goes into an algorithmic analyzer queue.
15/n
That analyzer simply looks for specific words or word combinations. Something that looks, with a high probability, to be problematic gets weighted towards to the "bad" end.
If there's enough bad weight, the automated system takes action.
16/n
If there's enough bad weight, the automated system takes action.
16/n
If not, it sends out a stock email saying "Yeah, we looked but couldn't find a violation." A complete fabrication, as no human has looked at all.
When you respond, this goes back into the analyzer queue tagged with "try harder". One of two things now happens:
17/n
When you respond, this goes back into the analyzer queue tagged with "try harder". One of two things now happens:
17/n
1) It adds extra options to look for more obscure hateful conduct (unlikely).
2) It lets it sit in the queue for a bit before sending out another stock answer. (more likely).
18/n
2) It lets it sit in the queue for a bit before sending out another stock answer. (more likely).
18/n
In other words, unless the person you're reporting has said something with the right trigger words and phrases, your actions are almost, but not entirely pointless.
Why not entirely?
The mass-report algorithm.
19/n
Why not entirely?
The mass-report algorithm.
19/n
See, if one person reports something there are two possibilities: it's either a personal grievance (in which case twitter doesn't care), or a genuine problem that is detectable by the algorithm.
But twitter does acknowledge that it's not perfect.
20/n
But twitter does acknowledge that it's not perfect.
20/n
Borrowing from the Todd Akin school of genius, if it's a genuine problem user, more than one person would be reporting it.
I suspect the system works two ways:
More reports weights the analysis towards the bad end, increasing the probability of action being taken, or....
21/n
I suspect the system works two ways:
More reports weights the analysis towards the bad end, increasing the probability of action being taken, or....
21/n
Eventually, the tweets reach a trigger threshold and a disgruntled actual human is summoned to take definitive action.
These are the ones that have an effect.
If you want to test this out for yourself, it's easy. Next time you get that "didn't find anything" email....
22/n
These are the ones that have an effect.
If you want to test this out for yourself, it's easy. Next time you get that "didn't find anything" email....
22/n
Just resond with "Hi," then a string of random words in an otherwise perfectly formed sentence.
You'll get back an email saying "Yeah, we took another look, but no...."
23/23
You'll get back an email saying "Yeah, we took another look, but no...."
23/23
