On handling online cheating with AI

[Adin]

What were the recent stats? How many wins and losses in even games against KGS 1d and against KGS 1k?

I have to protect the privacy of the person so I can't get into such details.

But let me give you an example. EGD statistics show that over three years (from start of 2017 to start of 2020) there were exactly 295 games played between 6k EGD and 2k EGD. The 6k won 14.2% of such games. The chance of a 6k winning three consecutive such games is 0.27%. Now when you do not look at three games but months of many games what are the odds?

[jlt]

I can't calculate the odds, you didn't give an approximate number of games and the winrate.

[Adin]

It was a rhetorical question of course. The point is that they are extremely low. And if we get stuck on the mentality that anything is possible then we might as well give our credit card details to that Nigerian prince who wants to leave us his fortune.

[ez4u]

As a first step, you could post the player's rank graph with the name trimmed off. Like this example. That would at least indicate the gains in online ratings that you have mentioned.

KGS graph example.jpg (50.99 KiB) Viewed 19080 times

[Adin]

As a first step, you could post the player's rank graph with the name trimmed off.

Obviously I could not do that, it would be a clear way to identify the person, even with the name covered. By the way I don't know if you posted that graph just as an example or you think it's very suspicious. Because it actually looks fairly normal, that short huge spike is usually due to low number of games.

[ez4u]

I just picked it at random off the list of players. I find it difficult to imagine that it is easy to find a particular graph if you trim off the kyu/dan rankings as well.

[Bill Spight]

The ontroduction of automatic alleged cheating detection and more so of unproved sentences against others will drive away some other players.

In this era of big data, we are already seeing algorithms that estimate important things, such as the quality of classroom teaching or the propensity to commit crimes, which are used to make consequential real world decisions, but which are untested. In addition, they are proprietary, so that it is, as a practical matter, impossible for them to be independently verified. Yet they are used, faute de mieux, and may simply enshrine current flawed judgements, at best.

I think that it is not only possible, but easy to come up with an algorithm that would distinguish between the plays of go bots and the plays of human pros, in human games played before 2015. Not no more. We humans have learned from the bots, and we are still learning from them.

After the sprint cheating scandal on IGS in the 1990s, I would not have held an online tournament without monitoring for prize money. At all. And now, everybody can consult their own expert.

[RobertJasiek]

KGS dan ranks differ from real world EGF ranks by up to 6 ranks without any cheating. That is how unreliable KGS ratings are. For kyus, expect larger differences. Therefore, before you start assessing cheating, first consider the usual variation.

Furthermore, take into account that players can have very different strengths in offline and online play (2 or 3 real world ranks difference easily).

So before cheating, you must tolerate ca. 10 ranks difference between KGS and EGF.

Before AI cheating, also consider more variation due to new throw-away accounts (win or throw away and create a next account).

If a long 10k KGS is 8d a week later, consider both a rating system anomaly or AI cheating.

Hence, before even dreaming of detecting AI cheating, correct the rating system, limit each person to one account and verify him by his identity document!

[Bill Spight]

Leela analysis shows a fairly high matching rate.

You gotta be much more specific than that. Matching is a form of confirmatory evidence, which is of extremely low value. In addition, people will count all sorts of plays as matches, as we saw in the Metta controversy.

To put it in terms of police testimony in American courtrooms these days, a high matching rate with Leela is consistent with improvement in playing ability. (It is also consistent with cheating, which is perhaps more like police testimony. )

Edit: Since nobody has posted another note, let me clarify. At this point in time, matches with Leela or any top bot, my indicate skill at go, but cannot be taken to indicate playing like a bot instead of a human. Especially if the match is loose, and not only of the bot's top choice. But show me that the player plays Leela Zero's top choice 95% of the time, that's good evidence of superhuman play. It is unlikely that even KataGo or Elf can do that!

[jlt]

So before cheating, you must tolerate ca. 10 ranks difference between KGS and EGF.

10 ranks difference for people who play frequently sounds impossible.

In the SDK range, a difference of 4 ranks (e.g. 6k EGF and 2k KGS) is quite frequent, 5 ranks can happen but 10 ranks? Do you know actual examples of people who play frequently and are, say, 2d KGS and 9k EGF?

[RobertJasiek]

Right, for frequent players, the variation is smaller. However, even for them intermediate variation of their KGS graphs (4 ranks up or down) can occur for a day or some months. It does not affect all players.

[Javaness2]

How can we stop referees cheating? Like skipping the match, and then afterwards just making up the rules as they go along.

[Bill Spight]

How can we stop referees cheating? Like skipping the match, and then afterwards just making up the rules as they go along.

Give them training and pay them some money.

Rule for referees #1. Always make your rulings from the book.

Rule for organizers #1. Write the book.

Edit: And, I suppose, this:

Rule for referees #0: Show up.

[Polama]

The flip side to this conversation is that while people can mimic an AI and jump 12 stones overnight, they can also cheat more subtly. Being a stone and a half stronger in important games by consulting the AI at a critical juncture or two and choosing a human looking suggestion is not going to be detectable from statistical analysis. So while there's value in weeding out cheating, I think it has to be assumed to be happening in any un-monitored tournament.

I'd also note that if you have a reliable algorithm for detecting AI moves, you can use adversarial learning to train a bot that doesn't look like an AI anymore. Maybe the limitation forces it to be a couple stones weaker, but that's still plenty of room to be stronger than us mere humans.

[daal]

I can tell you as an admin that it takes a lot of effort to analyze and even more courage to take action against a cheater.

I agree that it needs to be done, but it seems like an excessive burden on admins. Their conclusions will always be challenged, and they have to live with the nagging doubts that they might be in the wrong. The only reasonable solution imo is what SoDesuNe suggested - making a server-side cheating detector bot. We have bots. We have programmers. We have models from the chess community. It seems that the problem is big enough for KGS (the AGA) to put some money into this instead of foisting the responsibility onto volunteers.