On handling online cheating with AI
-
jlt
- Gosei
- Posts: 1786
- Joined: Wed Dec 14, 2016 3:59 am
- GD Posts: 0
- Has thanked: 185 times
- Been thanked: 495 times
Re: On handling online cheating with AI
Just a note on Bill's problems: for almost every board position, I preferred the human move. Perhaps Bill's point is that in some situations, it may happen that moves A and B have similar winrates, but A is much more bot-like than B. This might be detectable by training a neural network on human games, but that's hard work.
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: On handling online cheating with AI
Example 12
Who played the atari on four stones, who made the ko?
The winrate difference is 1.9%. Elf's play got 65k rollouts, and the human's play got 99.
Who played the atari on four stones, who made the ko?
The winrate difference is 1.9%. Elf's play got 65k rollouts, and the human's play got 99.
gains ⅔ pt. locally, but also takes away the potential eye there. If Black has to capture the 
stones, because there is no eye here, then this move gains more than ⅔ pt., and 
may have been OK, or even superior to the 1 pt. play on the top side.
takes his 1 pt. sente on the right side and then 
plays the hanging connection. So Black will not have to capture the 
increases White's territory and makes The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: On handling online cheating with AI
Well, that's your opinion, and you are welcome to it. But, as I indicated, if matching the bot's moves is an indicator of cheating, then not matching them is relevant to the discussion. IMO, of course.Adin wrote:I'm sorry if I've been a bit blunt about it. But this belongs to another thread. Posting 12 messages with diagrams and then maybe getting responses to them really disrupts the main discussion.
And the discussion of online cheating has a history, both here and in other forums. Several people seem to believe that matching some number of some bot's suggestions, in itself, is enough to decide the matter. But this is something that we actually know very little about. And people without a scientific background are unaware that confirmatory evidence, which is what matching is, is very weak. So instead of simply repeating once again that matching is very weak evidence, I decided to gather some scientific data. OC, it is not enough to settle anything, but it is a start.
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: On handling online cheating with AI
What about the data so far? It is next to nothing, OC, only 12 examples from a single game. (BTW, the game is GoGoD 1979-11-08a, Hashimoto Utaro, 9 dan, (W) vs. Ohira Shuzo, 9 dan. Ohira won by ½ pt.) I came up with criteria that allowed me to collect the data in an unbiased way: fewer than 500 rollouts for the human moves, with a winrate difference of 2.0% or less.
It is fairly easy to learn the preferences of bots in the opening. Most pros now play nearly perfect openings, according to the bots. And Elf did seem to play a more accurate endgame than Hashimoto. Other than that, my impression is that the bots like to play sente. In one example, both the peep and the narabi are sente, but Elf got in both of them in its mainline variation, while the human only got in one of them in the actual game. A number of people have already noticed that bots like to play sente, perhaps to a fault.
Elf calculates differently from other bots, so a winrate difference of 1% might be better with them. Still, I think that comparing the human play to a bot's top choice when the human play gets few rollouts but still has a small winrate difference is a good research method.
It is fairly easy to learn the preferences of bots in the opening. Most pros now play nearly perfect openings, according to the bots. And Elf did seem to play a more accurate endgame than Hashimoto. Other than that, my impression is that the bots like to play sente. In one example, both the peep and the narabi are sente, but Elf got in both of them in its mainline variation, while the human only got in one of them in the actual game. A number of people have already noticed that bots like to play sente, perhaps to a fault.
Elf calculates differently from other bots, so a winrate difference of 1% might be better with them. Still, I think that comparing the human play to a bot's top choice when the human play gets few rollouts but still has a small winrate difference is a good research method.
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
-
Knotwilg
- Oza
- Posts: 2432
- Joined: Fri Jan 14, 2011 6:53 am
- Rank: KGS 2d OGS 1d Fox 4d
- GD Posts: 0
- KGS: Artevelde
- OGS: Knotwilg
- Online playing schedule: UTC 18:00 - 22:00
- Location: Ghent, Belgium
- Has thanked: 360 times
- Been thanked: 1021 times
- Contact:
Re: On handling online cheating with AI
I got 10/12. In the last 4 examples I found it much easier to determine the better move and so I would only think the bot played the other move for some statistical win probability maximizing reason. But it didn't.
Now I admit, what's the point here? Me calling out the bot move 10/12 doesn't say a lot I guess. Suppose more people can do it, what does that prove? That this forum has good insight in what bots play? That this forum is strong on average? What kind of benchmark is this establishing?
Now I admit, what's the point here? Me calling out the bot move 10/12 doesn't say a lot I guess. Suppose more people can do it, what does that prove? That this forum has good insight in what bots play? That this forum is strong on average? What kind of benchmark is this establishing?
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: On handling online cheating with AI
If I have a point, it's that I don't know.jlt wrote:Perhaps Bill's point is that in some situations, it may happen that moves A and B have similar winrates, but A is much more bot-like than B.
But I also know that I don't know.
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: On handling online cheating with AI
The ability that you and jlt show in being able to distinguish between human and bot moves is hopeful, because that suggests that a neural network could learn that distinction, as well. Whatever the basis of that distinction is, it does not have to do with playing well, since I controlled for the quality of the plays.Knotwilg wrote:I got 10/12. In the last 4 examples I found it much easier to determine the better move and so I would only think the bot played the other move for some statistical win probability maximizing reason. But it didn't.
Now I admit, what's the point here? Me calling out the bot move 10/12 doesn't say a lot I guess. Suppose more people can do it, what does that prove? That this forum has good insight in what bots play? That this forum is strong on average? What kind of benchmark is this establishing?
And if a cheater learns how to make human style good plays, then he doesn't have to cheat, does he? (OC, he might fall back on Leela 11, but there you go.
)The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
-
Adin
- Dies in gote
- Posts: 28
- Joined: Thu Jun 16, 2016 1:25 pm
- Rank: 1 kyu
- GD Posts: 0
- Been thanked: 2 times
Re: On handling online cheating with AI
It proves how one can get to 10.000 posts while knowing he does not know while others not knowing he knows that.Knotwilg wrote:Suppose more people can do it, what does that prove? That this forum has good insight in what bots play?
-
Knotwilg
- Oza
- Posts: 2432
- Joined: Fri Jan 14, 2011 6:53 am
- Rank: KGS 2d OGS 1d Fox 4d
- GD Posts: 0
- KGS: Artevelde
- OGS: Knotwilg
- Online playing schedule: UTC 18:00 - 22:00
- Location: Ghent, Belgium
- Has thanked: 360 times
- Been thanked: 1021 times
- Contact:
Re: On handling online cheating with AI
If you're not careful, you're going to have more posts about Bill's posts than about what you actually want to discuss. Put away your anger for a moment will you? I wouldn't be confident in the refereeing of a person so easily aroused.Adin wrote:It proves how one can get to 10.000 posts while knowing he does not know while others not knowing he knows that.Knotwilg wrote:Suppose more people can do it, what does that prove? That this forum has good insight in what bots play?
-
gennan
- Lives in gote
- Posts: 497
- Joined: Fri Sep 22, 2017 2:08 am
- Rank: EGF 3d
- GD Posts: 0
- Universal go server handle: gennan
- Location: Netherlands
- Has thanked: 273 times
- Been thanked: 147 times
Re: On handling online cheating with AI
If the ratio is less than 5 (false negatives) to 1 (false positive), it would start to feel like a police state policy. The cure would be worse than the disease.Uberdude wrote:Adin, what is your view of the Carlo Metta cases from the PGETC?
And of Blackstone's principle thatWhat do you think the correct ratio is for AI cheating is in Go. I think it is reasonable to differ from that for say the crime of murder and the punishment of lenghty imprisonment or execution. But we shouldn't neglect the reputational damage of declaring someone an AI cheater, indeed I would be happy with a lower standard of evidence against an anonymous online account than one linked to a known human individual.It is better that ten guilty persons escape than that one innocent suffer.
Robert seems to think that ratio is infinity to 1. The USjusticeprison system seems to think it's flipped at 1 to 10. My first feeling is maybe a bit lower than Blackstone's, say 5 to 1, is an appropriate target to design for. What do others think is a suitable one for Go cheating?
Re: On handling online cheating with AI
IMO this is nearly impossible if a bot is only used for blunder checking. Not making bigger than a certain size of blunders needs dozens or hundreds of games to even remotely amount to meaningful evidence.Adin wrote:We really need automated detection but I heard of nobody working on that.
-
Uberdude
- Judan
- Posts: 6727
- Joined: Thu Nov 24, 2011 11:35 am
- Rank: UK 4 dan
- GD Posts: 0
- KGS: Uberdude 4d
- OGS: Uberdude 7d
- Location: Cambridge, UK
- Has thanked: 436 times
- Been thanked: 3718 times
Re: On handling online cheating with AI
I just checked the KGS TOS at http://gokgs.com/tos.jsp. Although it says "Dishonest play during rated games." is not allowed, it says nothing specifically against using bots, just prohibits "Having other people play rated games for you". It could be argued that using a bot to help you is not dishonest so long as you don't claim to not use a bot, so whilst I certainly agree it is bad etiquette not to volunteer that information through your obvious username (like leelazero7 I used on Fox), in your profile, or in chat before the game, under the current rules of KGS it does not appear to be cheating there either. Perhaps you should liase with the other admins and update the TOS, though changes to TOS should be clearly communicated to current users.Adin wrote:A few months ago I played a game on another server than KGS. I got crushed. Looking at opponent user info his latest about one hundred games were all wins. He had skyrocketed from a very stable 2k to 4d and showed no sign of stopping. I contacted the server administration and they won't do anything since there is no policy regarding AI cheating. And meanwhile instead of complaining or taking action people get into the most abstract philosophical discussions involving 0.5% win rate. Oh well.
About automatic bot detection, I have heard Yike server has some, though no idea of its details or efficacy.
-
Adin
- Dies in gote
- Posts: 28
- Joined: Thu Jun 16, 2016 1:25 pm
- Rank: 1 kyu
- GD Posts: 0
- Been thanked: 2 times
Re: On handling online cheating with AI
Using an AI as a human player in rated games does qualify as dishonest play. But it is certainly a good suggestion to add it specifically, I will bring that up.I just checked the KGS TOS at http://gokgs.com/tos.jsp. Although it says "Dishonest play during rated games." is not allowed, it says nothing specifically against using bots, just prohibits "Having other people play rated games for you".
If you care about something then you are passionate about it. I look at it as saying that some robbers are taking away our goods and we should do something about it. And then someone else talks for hours about whether a robber usually wears a hoodie or just a simple shirt, and do they use Android or IPhones etc etc. These things are totally irrelevant, a robber is that guy with YOUR phone in his hand.Knotwilg wrote:If you're not careful, you're going to have more posts about Bill's posts than about what you actually want to discuss. Put away your anger for a moment will you? I wouldn't be confident in the refereeing of a person so easily aroused.
-
Javaness2
- Gosei
- Posts: 1545
- Joined: Tue Jul 19, 2011 10:48 am
- GD Posts: 0
- Has thanked: 111 times
- Been thanked: 322 times
- Contact:
Re: On handling online cheating with AI
There are many openly AI enhanced players on KGS now, I haven't noticed anyone removing their rank. So I am not sure that you are right about that.Adin wrote:Using an AI as a human player in rated games does qualify as dishonest play. But it is certainly a good suggestion to add it specifically, I will bring that up.I just checked the KGS TOS at http://gokgs.com/tos.jsp. Although it says "Dishonest play during rated games." is not allowed, it says nothing specifically against using bots, just prohibits "Having other people play rated games for you".