Copying AI moves online

[gaius]

I fear internet go might be running into a new problem... Analyzing a KGS game from yesterday, I just found that my opponent copied AI moves at every stage. This guy is supposedly 2d, although I noticed he has so far won every ranked game he ever played, excepting one loss on time. I had not seen this before.

This is the win rate graph according to a fast Kata go analysis. There was one joseki in the beginning where my opponent really played a suboptimal move according to Kata. Might be because the human wanted to try out something or because he has a different network than me. Other than that, every single move he played was a computer suggestion and most of the time it is the #1 suggestion.


It's really frustrating because it is not so obvious when you are playing. AI just gets ahead and then starts playing very solid moves, almost looking slack sometimes. I was wondering why my opponent's moves looked so funny. I only realized what was going on this morning, when I analyzed the game in Lizzie.

Is this something that is known to occur more frequently? Is there a policy to deal with this? If there is, I couldn't find it. It really is a problem in my opinion, it has the potential to ruin internet go if too many people start doing this.

For those who are interested, the full game (I am black). Moves #42 and #82 were particularly blatant, and also both were exactly KataGo #1 suggestions.

[Kirby]

White played well... Moves 42 and 82 don't really look that unusual to me - especially 42. I like the consistency of 82, though, I'm not sure if I would have thought of it. I might have played on the right side... But that being said, 82 makes sense after you see it.

It's really hard to tell if someone cheated in any given particular game.

The sacrifice sequence leading up to 58 was not on my radar, either.

[Knotwilg]

I'm playing at the same level. There are clearly different types of games. Those which oscillate wildly when reviewed by LZ and those which go monotonously towards a result. The latter are suspicious and especially when the moves are bluelighted all the time.

What can you do? Just move on, I guess, until the behavior becomes ubiquitous and we're all 7d. At worst you're playing a bot while you intended playing a human.

[baduk]

Yes thats a problem,i dont think there should ever be prices on online games anymore,its absolutely impossible to tell if someone cheated if he did it the right way..

[Uberdude]

Yes, it's a growing problem in online go. There's been lots of discussion about it on this forum, both 2 years ago with a case in the Pandanet European Go Team Championship and some more recent threads; check the general go chat subforum.

I've not looked at your game yet, but if they aren't a stable rank a plausible non cheating explanation is that are a much stronger player still working their way through the rank system to a realistic rank. More suspicious is when they've been a stable lower rank for a long time and then suddenly win loads with lots of AI moves, ShaddyShenoute recently posted an example. Even with a relatively small rank difference it's possible to get such a one way slope of doom without cheating, e.g here is me as a 4d Vs a weak 2d: viewtopic.php?p=241268#p241268, at least I do have a few little steps up.

I was recently asked to investigate a suspected case of online cheating, in a not important game. Given the players were supposed to be around 3k and one played at least as strongly as me with lots of bot similarity I estimated the chance of cheating at 70% (but with caveats about not seeing control group games) with most of the remainder probability being they got a strong friend whose studied AI to play. But then I learnt the player was 30k less than 2 years ago, 10k about a year ago, ~4k 7 months ago in last real life tournament. So continuing the rapid pace of improvement in covid lockdown to mid dan is not impossible. The game was still pretty suspicious with some strange timings matching bots but to risk falsely saying a non-cheating rapid improver was cheating would require more effort than I was prepared to spend so I was glad the situation meant "inconclusive" was a valid conclusion.

[Gomoto]

Is this something that is known to occur more frequently? Is there a policy to deal with this? If there is, I couldn't find it. It really is a problem in my opinion, it has the potential to ruin internet go if too many people start doing this.

Yes, frequently.

Dont bother, just continue to play the next game.
(If I am sure during analyzing my opponent is a cheater, I just block him.)

It will not ruin internet go because it has not ruined internet chess yet. In chess automatic cheat detection is quite good already. I think with a neural network approach cheat detection will be quite good in go too in the near future.

[Bill Spight]

It will not ruin internet go because it has not ruined internet chess yet. In chess automatic cheat detection is quite good already. I think with a neural network approach cheat detection will be quite good in go too in the near future.

With the widespread availability of AI reviews, IMO, the average level of online amateur play will increase by around 2 stones within the next 5 years. That does not mean that ranks will increase that much, because one's opponents are getting better, as well. But some people will advance quite rapidly, and we may see some ratings instability because of that.

[Uberdude]

I thought I would give my policy network top 3 (or fewer if only move) choices for the game in question for the next move as I clicked through, usually within 10 seconds. Then we can compare how similar the suspected cheating KGS 1d, gaius 1d, and me as a 4d are to whatever bot. Where the game move matches one of my choices I put a Wm or Bm to mean white/black match. If I match similarly with white and black then that suggests moves have similar human obviousness, if I match lots more with white that means I can play like a suspected bot without using a bot. If I match less with white that suggests white's play is more botty and less like a 4d who studies with bots so more likely to cheat. Where my suggestions don't match, does bot say the game move is better or worse? (I wrote these hypotheses before going through the game.)

Code: Select all

1. q16 r16 r17   Bm
2. d16 d4 d17    Wm
3. q3 d4 q4      Bm
4. d4 c4 d3      Wm
5. o17 o16 p17   Bm
6. r5 q5 r6      Wm
7. r5 q8 c17     Bm
8. c16 d17 r3    Wm
9. c16           Bm
10. c14 c15 d15  Wm
11. c15 b14 r5   Bm
12. b13 c13 d15  Wm
13. c15 b15 d15  Bm
14. d14 d15      Wm
15. d15          Bm
16. e15          Wm
17. e14 e16 d18  Bm
18. f15 d18 e16  Wm
all matches so far, not surprising during joseki, but thought might get some misses in fuseki where >3 good choices, but didn't
19. d18 e13 b12  Bm
20. e18 e13      Wm
21. c18          Bm
22. f17 e12 g17  Wm
23. e13 d12 e16  Bm
24. e17 e12 f16  Wm
25. e13 d12 b12  Bm
26. c12 e13 b15  Wm
27. a13          Bm
28. c13 b11 c11  Wm
29. B11 r5 b15   Bm
30. c11 b15 e13  Wm
31. b10 c10      Bm
32. c10 f13 d9   Wm
33. b9 c9        Bm
34. c9 c8 r3
first non match! if bot DIDN'T want to play this on previous moves but DID this one that's suspicious, but it wanted to play all moves then not.
35. a14 d8 b16   Bm
36. a12 c8 r3    Wm
37. r2 r7 d8
first black non-match
38. d3 c4 q2     Wm
39. c4           Bm
40. d5 c5 c6     Wm
41. c5 b6 c6     Bm
42. d6 c6 d7     Wm
so 42 not wacky genius bot move to me but my 3rd instinct
43. c6 d6 r2     Bm
44. e6 c6        Wm
45. c6           Bm
46. e7 f7 q2     
another sente move timing Q
47. a11 e7       Bm
48. a15 q2 e7    
49. c8 b8 k17    
a tenuki
50. f5 c2 g4     
am I following black too much with my policy? Seeing as I thought black shouldn't tenuki, it's natural for white to punish with good sente profit.
51. a12 c8 e5    Bm  
52. b8 f5        Wm
53. b16 a9!      Bm
54. f5 c2 c7
not a match but an obvious candidate if I don't blitz
55. d2 f5 r2
56. f5 e4 c2
Game e2 was 4th choice, this is a case where need reading not policy
57. f2 f5 c2     Bm
58. f2 (f4 rejected on reading ladder) f5 Wm
59. e4           Bm
60. g3 h3        Wm
61. f2 e3        Bm
62. g2 f4 q2     Wm
63. d2 f4 f1     Bm
64. f4 q2 m17    Wm
65. e3           Bm
66. h4 q2 m17    Wm
67. m15 o18 r2   
68. q16 q2 o3
again I am following black too much, given black m15 was good to me before white there now is good too.
69. n17 r2 g4    Bm
70. r4 q16 n4    Wm
71. p2 n14 s2 
I guess black move is bad
72. h4 j4 l4     Wm
73. k4 k3 h2?    
Gosh, need to read, my policy assumed this cut didn't work
74. h3 k3 g5     Wm
75. h4           Bm
76. h5           Wm
77. h3           Bm
78. k3 k4 q16    Wm
79. k3 l2 o4     Bm
80. l4 o3 s2     Wm
81. n2 n3 m4     
82. n5 n10 r9
If this exact spot is blue that's indeed suspicious, though I did pick one floaty middle move further away at n10. Locally speaking the large knight to grow centre is a moderately common haengma. I was also consider moves like q16 o14 as >3rd choices.
83. q8 k14 l10
84. o4 r8 q16    Wm
85. q17 q15 r17  Bm
86. p17 r17 q15  Wm
87. p16 q15      Bm
88. q15          Wm
89. r17 q18 r15  Bm
90. o17 p18      Wm
is p18 better?
91. n16 s14 p14  Bm
92. r15 n16 q14  Wm
93. n16 s15 r13  
94. s16 s15 q18
game 94 was random sente move, does bot like it now, or human timesuji?
95. a16          Bm
96. s16 s15 q18  
is white move better or worse than mine?
97. s15 r11 n16
98. s13 s15 q13  Wm
99. s14 q13 r14  Bm
100. r14 q13 s15 Wm

So of those first 100 moves, black played 40 moves matching my 3 instinct choices, and white also played 40. So black and white are both playing the same level of "obvious moves for a dan player". The question then comes down to those 10 moves each that didn't match mine, are white's moves generally better than mine, whilst black's generally worse? 100 moves are essentially reduced to 20 interesting data points for study. Some of those, from both white and black, were on my radar just >3rd choice. The most suspicious of bot cheating moves from white were the timing of the sente moves on left edge (does bot like only them or all the time?) and l12 jump. Move 56 was a good move for a low dan, but not a mid dan.

[Bill Spight]

OT, hidden for courtesy.

WIP =? West In Peace, Elmer Fudd

[Gomoto]

Just to be clear. I do not promote comparing player moves to network moves for cheating detection. I promote a network approach for cheating detection.

[goame]

Cheating detection works only against noobs, which are winning all games.

Other things does not work because we have 2000000 nets and x some 1000 gpus + cpus and x time is also a factor and even I see sometimes better moves than the strongest KataGo net running on two rtx 2080 ti.

[Shenoute]

Cheating detection works only against noobs, which are winning all games.

And even when cheating is obvious*, it seems some servers have no interest in taking action.

*By 'obvious', I do not mean a player winning one or two games in a very convincing way with a couple of very good moves. I mean a player going from a stable 4d level to 7d in a matter of days, winning 80% or even 90% of his games at every level, constantly over dozens of games, suspiciously resigning a couple of won games here and there, etc.

If even this sort of cases cannot be dealt with swiftly, I have no hope go servers will develop a system to curb more subtle cheating cases anytime soon.

[Uberdude]

Ah sorry Shenoute, I get you and Shaddy, the 2 mid-dans starting with 'S' mixed up.

[John Fairbairn]

And even when cheating is obvious*, it seems some servers have no interest in taking action.

I was watching a tv programme recently which mentioned a very successful experiment in a town with major problems in the pubs - drunkenness and fighting. Rather than usual remedies such as closing the pubs, calling in the police or overtaxing booze, they employed an Arts & Crafts architect to rebuild the pubs. With tiled floors instead of sawdust and spit, even men started enjoying the ambience, and these drinkers themselves started regulating their boozier friends. This and the rest of the elegant décor began to appeal to other, quieter, drinkers and to women, and so they ended up in a benign circle where the boozy men even started to regulate themselves. This was some time ago and the pubs are now listed building, but pub chains have since followed this sort of model.

We see a similar approach in other problem areas, such as graffiti - providing an empty wall with no risk of prosecution for daubing on it seems to turn offenders away from acting like spraying tom cats. Similarly head-shaking and tut-tutting is always more effective than road rage.

I don't claim to understand the psychology behind this, but I suspect it's nothing like reacting to saintly turn-the-other-cheekers. I reckon it's just that the offenders feel they have been sussed out as acting like naughty children and they are not gaining the sort of attention or smugness they crave.

In that light, I wonder whether AI cheating can be alleviated simply by servers announcing clearly that using bots is perfectly acceptable. The cheaters then know that everyone else knows they exist, thinks of them as unfortunates who have yet to grow up, and will react merely with a head shake.

[Uberdude]

Trouble is lots of people, myself included, don't want to play people using bots online. It's a rare occurrence (like once a month) I get 1 hour+ of uninterrupted time where I'm allowed to play a game online so wasting that precious time losing to a bot, something I can do other times if I want (I don't), is annoying. Servers could add a flag to user profile "I use a bot" and allow people to self declare. I would do so for the LeelaZero7 account I used on Fox a while back to openly and without deceit play as a bot. Fox actually has a flag for AI users and that got automatically applied to my account. But how many of the people who use a bot now without admitting it would tick that flag, allowing people like me to avoid them? I'm guessing not many. And a server with a "using bots is allowed and you are indistinguishable from non-bot humans" policy wouldn't attract me.