I first read this some time ago and wrote about it then. It’s such a good report that I thought I’d write about it again here.
This is another example of Diplomacy being used as a tool to look into something that happens a lot in real life: betrayal. Dip is a great tool for this – it’s part of the game, Carebears notwithstanding.
I came back to this report because of the previous post. Here’s the question the research is aimed at answering:
Such scenarios suggest an important research challenge: is the forthcoming betrayal signaled [sic] by linguistic cues appearing in the (ostensibly friendly) conversation between the betrayer and the eventual victim? A positive answer would suggest not only that the betrayer unknowingly reveals their future treachery, but also that the eventual victim fails to notice these signals. Capturing these signals computationally would therefore mean outperforming the human players.Linguistic Harbingers of Betrayal: A Case Study on an Online Strategy Game; 1. Introduction
The report investigates ‘dyad relationships’. These are relationships between two participants; in Diplomacy – and for the basis of the research – these involve two people in an alliance. The report focuses on friendly relationships, which it defines as those evidenced by multiple acts of support by two players; and hostile relationships, which it defines as those evidenced by one player invading another’s space or supports a player already attacking a third player.
A betrayal is evidenced by a hostile act by one player against another when the two have been in a friendly relationship. The authors found that, in the game, a friendly relationship breaking down into a hostile relationship is about 5 times more likely than the reverse. Not, perhaps, surprising in the game, and in real life: building bridges is much harder than pulling them down.
There are two other points that worth mentioning about the research. It focuses on stable relationships. A stable friendly relationship is when at least two repeated and reciprocal acts of friendship take place over a period of at least three seasons, and are no more than five seasons apart. A betrayal occurs when an established friendly relationship ends with at least two hostile acts. The person who initiates this is the betrayer and the person being betrayed is the victim.
They also compare instances that are similar. So, for each instance of betrayal, they compare it with a similar example of a friendly relationship that didn’t break down, based on the length of the relationship and the number of seasons from the start of the game.
What they did was look at the correspondence between the two players. As a successful stab in Diplomacy tends to be a surprise to the victim, they were looking at subtle clues. After all, while I’ve played one game in which a player regularly signalled their intent to attack them, this is a huge rarity, for pretty obvious reasons.
Are there clues in correspondence?
Hmm. Well, if there weren’t, then it wouldn’t be worth writing the post, would it? So the answer is a definite YES. But what are these cluese?
Sentiment: How many positive, neutral and negative statements are made in messages? It was found that, when an imbalance occurs between two players in a friendly relationship, this is a clue to a betrayal. What the saw was that it is the betrayer was significantly more positive about the relationship than the victim.
This seems counter-intuitive at first sight. Positive statements seem to indicate that a player is likely to remain friendly. The report’s authors speculate that this is over-compensation for the eventual betrayal, which is probably one reason for it. However, I’d suggest that this could also be that the betrayer could grow tired of their ally being less positive, or even that they believe the lack of positivity from their ally.
Argument and discourse: How much planning goes on? They look at the content of messages. Are there indicators of persuasion, discussion, planning, etc goes on. Again, an imbalance in these markers is indicative of a betrayal.
I find it interesting, given my correspondence style, that more planning markers were common in correspondence from the victim. This is perhaps because the betrayer is less invested in the alliance and a lack of planning discourse signifies this. It could also be, more cynically perhaps, that the betrayer gets a tired of the long-term planning of the victim, or that the betrayer takes a clue from the amount of planning the victim displays shows that the victim is likely to be an easy target.
Politeness: These indicators relate to whether players offer suggestions, questions, validation and what I’d call ‘soft requests’ such as “I wonder if you should..?” Once more, there tends to be a ‘slight’ imbalance between politeness that indicates a betrayal.
They found that it is the victim that is less polite than the betrayer. This could be that the betrayer grows weary of demands or assumptions. Again, it could be that it’s a result of over-compensation. It isn’t surprising that this is indicated by a slight imbalance: most players recognise that rudeness isn’t the way to build or maintain a relationship.
Talkativeness: They compared the amount of conversation between the two allies. They measured the number of messages, the average number of sentences in a message, and the average number of words in a sentence. Again, where an imbalance exists, a betrayal is signalled.
What predicts a betrayal?
All of the above, is the answer. Interestingly, the authors looked at subjectivity and found that there were no clues linked to this. In other words, it doesn’t matter whether statements are opinion, accusation, suspicion and speculation didn’t have an impact on a betrayal. This is probably because players are aware that paranoia has a big impact on the game and therefore expect that others will express doubt in an alliance.
Imbalance is the key. When there are imbalances in sentiment, argument and discourse, politeness, and talkativeness, a betrayal is very likely to happen. Perhaps a good question is: When? How soon can I expect this?
There are other clues that suggest when the alliance is likely to break down. A betrayer is likely to become more positive just before a betrayal (hence the over-compensation suggestion).
Opinions, on the other, seem to indicate that a betrayal is not imminent – so a player who is still opining on what is happening in the game is still behind the alliance.
Discourse features – those that indicate that an effort is being put in – are less evident before a betrayal occurs. The victim becomes less eloquent prior to being betrayed… which, I guess, affects the confidence of the betrayer in their choice of ally.
Again, when a victim introduces more planning into a conversation, the betrayer becomes more likely to stab. Is this because they have less invested in the alliance, or because the stress of planning for the future has an impact on any relationship?
Politeness is very interesting. A betrayer will, initially, be more polite than their victim, however, as the point of betrayal approaches, the betrayer becomes less polite, whereas the victim will actually become more polite.
Similarly, as betrayal approaches, the victim will make more requests prior to being betrayed, perhaps signalling an increase in desperation in their position. The more pressure put on a relationship means that it is more likely to end.
If a player increases the complexity of their messages, then the change indicates that they are thinking of betraying the victim. However, this is not an imminent threat: the betrayer could be preparing the way by attempting to increase their victim’s confidence in the alliance.
What can we learn from this?
Well, first, we need to recognise that there is no infallible recipe to predict a betrayal. These are indicators and they suggest a betrayal is likely to happen. Interestingly, the programme the authors built to analyse correspondence did predict betrayal much more successfully than experienced players who had access to the same info. Even the best players don’t pick up on the clues within correspondence!
This should give us all confidence! We do have chances against good players… in this aspect, anyway!
The one big indicator is imbalance. When one player’s correspondence is different in form and content from another’s, an imbalance is created and this suggests the relationship won’t last.
If, then, you want to maintain an alliance you need to match your ally’s style and content. This isn’t surprising if you have some understanding of psychology. We tend to match with people who are similar to us. Difference makes us uncomfortable. And an imbalance indicates that one person is more invested in an alliance than the other.
Consistency is the key to not signalling a change in an alliance. If you want to give the impression that you’re wanting to maintain an alliance, when you’re actually preparing to stab your ally, then put effort into not changing your style.
It’s possible to predict that an ally is likely to betray you, too. If a player becomes less polite, then they are likely to be very close to stabbing you. If they become less communicative, it is also likely that a stab is in the offing.
We do need to take a holistic approach, however. Paranoia, which can lead us to question an ally’s commitment, can also lead us to read more into slight changes than is necessary. Read the board, too: if these changes happen when the position on the board doesn’t suggest a stab is probable, then take them a healthy does of scepticism.
And we need to remember that the game of Diplomacy is not a science. There’re plenty of players who don’t play in what seems a logical manner. There is no infallible set of clues that will predict a stab, or the imminence of it.
However, if you can use these clues, it’s possible you can prevent stabs, or prevent stabs being perceived.