[bsder]

1) You can look at the "strength" of individual moves. Someone who plays at 2000-level normally but magically coughs up 2600-level moves when in trouble is probably cheating (watch some of the live chess streamers--you'll regularly see this in real time). Computers are quite good at estimating the strength of a move after the fact.

2) Quite often there are certain "play lines" that computers will play that humans simply can't find over the board.

For example, a computer can take a defensive "play line" that is littered with traps with only a single non-losing path for 30+ moves and work it out really quickly (there is only one non-losing path to take so it prunes the search space mega fast) and play it perfectly. A human playing such a line is almost always cheating--humans simply can't run those kinds of lines in real time.

If you look at computers analyzing even the highest end games, you see the humans making quite a few mistakes that the computers will spot and take advantage of immediately. Someone who walks down these kinds of paths regularly is a statistical anomaly.

That having been said, given the current crop of computer-trained chess kids, it IS possible that we'll grow a prodigy that can run those kinds of lines. However, it doesn't seem like that person exists, yet.

[Nomentatus]

I'm haunted by the possibility that humans might (at least half) catch up, too. When I look at how AI beats humans, I can't help thinking that AI shows us that human narcissism holds humans back. We don't want to look stupid, or mediocre - we don't want to make moves that are hard to explain the value of clearly.

In Go, we can't make ourselves spread our moves around the board as much as we should, we tend not to choose a maybe good move elsewhere over a clearly powerful move where the board is developed, for example.

Maybe there's a pattern to the moves AI chooses that is also a pattern humans can see without running every line; we're just reluctant to choose moves that we can't clearly justify in the shorter run.

[umanwizard]

> Computers are quite good at estimating the strength of a move after the fact.

Are you sure? I’ve never heard of such a program.

[Nomentatus]

He's just referring to the fact that after game is over you can let your program stew on any given move for a weekend or more before reporting back on how strong it was or wasn't.

[umanwizard]

But standard programs like Stockfish won’t tell you how strong a move was. They’ll just tell you how much it changes the evaluation.

E.g. if you initiate a queen trade in a straightforward position, on the next move I have to take back my queen; any other move will show a gigantic evaluation drop by the engine. But that doesn’t mean it’s a particular strong move — even an absolute beginner will play it. Thus it’s of no value for determining whether the person who played it is cheating.

It’s entirely possible that chess.com has access to more sophisticated software that can estimate the strength of players (they sort of allude to this with their “strength score” metric) but AFAIK it’s not publicly available and not clear how it works, or whether it can evaluate individual moves as opposed to the game as a whole.

[Nomentatus]

I think it's more a question of whether, given more time to calculate, the software changes it's choice of move to something else.