Why can't they auto-block RMT posts?

[SamSmoot]

My guess is it would probably take resources SqE don't have or as a company gain very little (to them) to use. It's a machine learning problem where someone / a group will have to first sit and manually flag several millions/billions of examples of chat messages as RMT or not, to train and run a Bayesian network or support vector machine algorithms (or RNN, FCN, etc... i.e. different AI neural network archetypes) on.

The players would do the pattern recognition. After many RMT reports from different accounts, the only thing software would have to do is find matching strings in, say, 100 of the reports, then start blocking anything that contains that string. Such as: "[Ultrafast Deleviryin 10 Mins]". If they change a few characters, players will report the new ones, and they get blocked again, along with the old ones.

Example: Take this popular posting:
Ysteueqr Kejajuue: ----[SCUMBAGS.COM]----[Best Buy Gil Store]----[Cheapest Price]-----[SCUMBAGS.COM]---[Ultrafast Deliveryin 10 Mins]--[6% OFF Code;LOVE]---8qwet

After 100 reports, the blocking algorithm sees the highlighted strings in 50% of them, and starts blocking on just ONE of them.
The spammer then has to figure out which one is causing the block and change it.. Next 100 reports, and it picks a different string than it blocked on previously, and it continues. The spammers will have to keep finding and changing the blocking strings every 100 posts, and they may eventually run out ways to say what they want.

[manamoppet]

The players would do the pattern recognition. After many RMT reports from different accounts, the only thing would have to do is find matching strings in, say, 100 of the reports, then start blocking anything that contains that string. Such as: "[Ultrafast Deleviryin 10 Mins]". If they change a few characters, players will report the new ones, and they get blocked again, along with the old ones.

That's true, but can take a while (10 hotbars with different macros on them pr. bot), and after a while uses more storage than necessary when new string variations are added. Think of how many combinations words can be altered in a Twitter message, and that's only 140 characters. With machine pattern recognition you don't need exact matching strings, just be close enough when calculating e.g. cross-entropy for loss equations, and the number of parameters to learn is mostly unchanging (the values of each parameter will change). I meant that initially one would need to manually label that many chat messages, but of course after the initial training one would augment the training by adding the players new RMT reports (players doing the pattern recognition).