artificial intelligence in chess

Imagine someone secretly developed and ran a superior chess engine on a supercomputer that could baffle its contemporary engines (say it is 500 ELO higher than them) and decided to tell nobody about it. If analysts, cheat detection programs, and observers received only the moves of its games, would it look like a human was playing? Even with a large sample size of games, I think it would, as the analyses would fail to understand the depth of some of its moves. But still I wonder, would it agree with current engines enough to be caught?

It would get caught. It would NOT make human type moves. It would just be better computer moves.

I think it would not get caught because of its moves but it might get caught by its results and some other factors - however, if it was trying especially to avoid getting caught by playing in situations where it wouldn't attract a lot of attention, and then easing up when it knew it was winning, then it could be undetectable.  If it were truly intelligent it probably would do that, having realized that pride is in vain. 

Maybe like Marvin the elevator from the Hitchhiker's Guide.

It might get caught if it would cause a blackout across California when facing a new situation.  It might get caught if whoever controlled it registered it anonymously or suspiciously for online competitions.

Whether it would get flagged by chess.com for cheating would depend on whether chess.com is willing to accept false positives once in a while.   In a machine learning class, I am just a beginner, but I can see some shape of some systems that might be used, and what's behind detection seems to be statisitcs.  humans and experiments decide what kind of data the detection system will use, and humans decide what examples a machine learning system will use to learn from  Without insider information proprietary to chess.com I can't answer specifically at all whether such an engine would be detected.  If you want to know more, courses on Coursera are free unless you opt to pay for a certificate, and I'm sure there are other online courses for it too, or you can just read up on Machine Learning from published books.  The light of knowledge is better than the confusion of darkness!

If the best human anybody knows about can play chess at a level of 2800 or whatever it is, and the best known program plays at 3400 or whatever, then an unknown anonymous thing comes along that is really significantly better than that, any reasonable statistical detection system would flag it as an anomaly.  That's what the topic is called within the machine learning subject - anomaly detection algorithms.  Fraud prevention is an example - the same types of statistical machine learning techniques that are used to detect fraud (maybe with credit cards or insurance claims) could be used to detect weirdness in chess playing.