THE DIPLOMATICON

In the previous post in this series I explained that, in July 2015, a paper was published that had looked at whether there were any clues in language that might indicate a friendship was about to end. This research was based on a study of online Diplomacy games.

So, what did they find out?

Well, first, how did they look for clues? “To ensure that we are studying conversational patterns that occur only when the two individuals in the dyad [2-player pairing] are ostensibly being friends, we only extract features from the messages exchanged before the last act of friendship. … we can only hope for subtle linguistic cues: if there were salient linguistic signals, then the victim would notice and preempt the betrayal.”

So they’re looking at the messages sent before the act of betrayal, the stab. They compare these with other conversations between 2 players at a similar stage in the game, from a similar length of alliance, but where the alliance continued.

By comparing the patterns in language from these two types of game, they came up with the following evidence.

Sincerely Insincere

Positivity Imbalance

It isn’t surprising that alliances are built on positive interactions between players. You’re not going to willingly ally with someone who tells you how bad things have gone, are going or will go. Alliances mimic friendships; it’s hard to like a moper.

What they found was that a player who was about to betray another tended to use a lot of positive sentiment. In fact (and this is surprising) they would use more positive sentiment than players who were in an alliance that lasted. Why should this be? As the authors suggest, possibly over-compensation. If a player is about to betray you, they could be using excessive positivity to mask the impending stab.

Planning and Persuasion

They looked at the occurrence of language of persuasion & discussion, and at language indicating planning activity. What they found was that a victim was likely to have more planning related language in messages than the betrayer. This may be because the betrayer doesn’t see any future in the relationship and has subconsciously given up.

On the other hand, what they found was that a betrayer would use more persuasive language than a victim. The betrayer may have stopped thinking in terms of planning, but they’re certainly trying to persuade the victim of something… how strong the alliance is, perhaps?

Politeness

It’s not difficult to see politeness in messages. Every message should be polite. However, again, they found something which seems less predictable: that before an alliance broke down the victim used fewer politeness markers than the betrayer. Maybe that’s because the betrayer, knowing the power they have, is more polite as a way of distracting from the coming stab or possibly is softening the victim up. Of course, it could be that the betrayer is preparing to stab because they’ve got fed up of the other players impoliteness!

Patter

They looked at the number of messages sent, the number of sentences per message, and the number of words per sentence. What they found was that there was an imbalance between the talkativeness of the two players when a betrayal is about to take place. What they saw was that a betrayer is likely to use more sentences per message than a victim.

Warning Signs

The researchers built a model, using their findings, to see if these findings would allow someone to predict that a stab was coming. They presented this to players and ran a classification model.

With players, the clues were missed. People were unable to see the betrayal coming. In fact, the results suggest that you may as well toss a coin to decide whether someone is likely to betray you or not. We all know this: it’s where the paranoia in Diplomacy comes from.

However, the classifier was able to predict betrayal at a level significantly higher than chance. This tells us that, given a raised level of awareness, we should be able to better predict when someone is going to betray us.

What are the clues? “Overall, the selected linguistic features capture a consistent signal that characterises people’s language when they are about to betray: they tend to plan less than their victims, use less structure in their communication, and are overly positive.” We can also see that they will be increasingly persuasive and write longer messages.

Of course, some of this may just be the way people communicate. Perhaps it’s even more to do with the profile of a player who is more likely to stab and one who is less likely to stab.

So we need more information and the researchers looked back over previous turns. What they found was that betrayers begin to show a change up to two turns before the act of betrayal. This isn’t surprising: the best stabs are planned ahead of time.

Positivity imbalances occur just preceding the stab. If someone is still expressing opinions through ‘claims’ (that is, stating what you think through perceived evidence), they’re not likely to stab you immediately. Betrayers become increasingly more talkative than their victim the closer the stab looms.

They also found that, if one player tends to talk more about plans, they are more likely to be stabbed, possibly because of the pressure of planning having a negative effect on the betrayer. Interestingly, if a player suddenly becomes more polite, then they’re likely to be betrayed!

So what can we learn from all this? Well, first, we’re bad at predicting when a stab will happen. This isn’t news. Even leaving aside our inability to read the clues the team looked at, some of us are bad at reading the board and predicting when a stab is likely to happen. Perhaps we prefer to snuggle beneath the comfort blanket of a long-term, on-going alliance than to peek out head out to see if we have to throw the blanket off.

---

POSTS IN THIS SERIES:

• Harbingers of Betrayal: Introduction
• Harbingers of Betrayal: Findings
• Harbingers of Betrayal: Lessons