Same quote here: a sophisticated AI prevails.
"... Instead these systems are based more on probability. Big websites have millions of games to analyse - including games with confirmed GMs to learn from. A fraud detection algorithm can consider unlimited questions like:
If black has a 3 pawn lead but lots of unguarded pieces and down a knight, how likely are they to aggressively recapture instead of defend?
If black has been playing aggressively, how likely are they to play a very difficult to calculate defensive move?
When black has 15 available moves, 10 of which are pretty good, how long will it take for black to play a move?
Black has played 20 straight GM level moves and has a strong material lead. What are the odds that black will now play an amateur blunder? A "look I'm not cheating" move?
Is black suddenly playing like a GM whenever there's one really important move?
Does black suddenly play very well mid game if they are slightly behind? How likely is a human to do that?
Is black playing hard to calculate moves just as fast as easy to calculate recaptures?
Note how in all those cases, whenever I said "mid game" it could have been 10 turns in, 15, 17, whatever, each with different probability calculations. When I said "engine" it could have been any one of twenty engines. When I said 3 pawns and 1 knight, it could have been X pawns and Y knights.
If you analyse millions of games and programmatically construct probabilistic answers to thousands, even millions of questions like this, you can make a "threshold" system that is far more accurate than anything a human could come up with by comparing naively to an engine. Some questions turn out to be poor predictors of cheats. Some turn out to be great predictors - a human does not decide which are the best questions to ask, or with what parameters. I'm basically describing machine learning - which is definitely what the better systems must be doing. ..."
"... Instead these systems are based more on probability. Big websites have millions of games to analyse - including games with confirmed GMs to learn from. A fraud detection algorithm can consider unlimited questions like:
If black has a 3 pawn lead but lots of unguarded pieces and down a knight, how likely are they to aggressively recapture instead of defend?
If black has been playing aggressively, how likely are they to play a very difficult to calculate defensive move?
When black has 15 available moves, 10 of which are pretty good, how long will it take for black to play a move?
Black has played 20 straight GM level moves and has a strong material lead. What are the odds that black will now play an amateur blunder? A "look I'm not cheating" move?
Is black suddenly playing like a GM whenever there's one really important move?
Does black suddenly play very well mid game if they are slightly behind? How likely is a human to do that?
Is black playing hard to calculate moves just as fast as easy to calculate recaptures?
Note how in all those cases, whenever I said "mid game" it could have been 10 turns in, 15, 17, whatever, each with different probability calculations. When I said "engine" it could have been any one of twenty engines. When I said 3 pawns and 1 knight, it could have been X pawns and Y knights.
If you analyse millions of games and programmatically construct probabilistic answers to thousands, even millions of questions like this, you can make a "threshold" system that is far more accurate than anything a human could come up with by comparing naively to an engine. Some questions turn out to be poor predictors of cheats. Some turn out to be great predictors - a human does not decide which are the best questions to ask, or with what parameters. I'm basically describing machine learning - which is definitely what the better systems must be doing. ..."
