A simple experiment suggests a way to encourage truthfulness

Just to make this technique a bit clearer, since "subtract 50 and double" was a bit vague to me: you ask 100 people the question about illegal activity, and assume they're going to toss the coin and follow the protocol. That gives you 100 "yes" and "no" answers, of which ~50 "yes" answers are due to following the protocol and can just be ignored. The remaining yes/no answers are from the tails results and that's the data you use to get a final answer. As an illegal opiate producer in Afghanistan who answers "yes", you have plausible deniability by just saying "I got heads."

The fairness of the coin toss means the researcher asking the question doesn't need to know who got which coin result so it can be kept secret.

This description is a bit confusing, so I'll assume that the protocol was this:

100 people are asked a possibly self-incriminating question - "Are you indulging in bad activity X?" They are expected to toss a coin. If the coin said tails, they are expected to give a truthful answer. If the coin said heads, they are expected to say "Yes."

Now, no single person needs to worry about self-incrimination, since they can always claim that their "Yes" answer was the result of a "heads" on the coin toss. But still, you can get an estimate of the percentage of bad activity X going on, by subtracting 50 from the number of "Yes" answers and doubling.

(This is assuming of course that the subjects believe your claim that there is no way for individuals to self-incriminate - which is certainly not a given, IMO.)

Similar methods are widespread in interrogation tactics as well. With enough lies, you can determine the truth (eliminate all contradictory statements and what remains is either the truth or an amazingly executed piece of misinformation). Along the same lines, in SERE training, you're instructed to not answer any questions, and if you can't hold out from answering anymore, just tell the truth because lying helps as much as telling the truth.

Think it's called "Bayesian truth serum" usually: http://www.sciencemag.org/content/306/5695/462.abstract

In middle school, some of my friends and I discussed systems for transmitting information about crushes. Wouldn't it be nice if some trusted person knew everything, and could tell you whether your crush liked you? But then we faced a certain use case. One of my friends liked a girl; if she liked him back, obviously he would like to know that; but, he said, if she didn't like him back, then he didn't want to know that, and would prefer to keep hope alive.

So, we thought of this: Have some trusted person learn her preferences, and then flip a coin: if heads, then tell him if she likes him (and say nothing otherwise), and if tails, say nothing either way. This seemed to serve pretty well. [This was all hypothetical, of course.]

I think this is a good koan--something to provoke further thought. For example:

1. This is sort of how a scientific investigation may go. Suppose you have some theory, and some test that only yields results sometimes: 50% of the time, the test will yield result A, but 50% of the time, it will work properly, and in that case, if your theory is right, then the test will yield result A, and if your theory is wrong, then the test will yield result B. (Example: "There are no animals in this area", and imagine that they'll try to avoid being caught and that you can conduct an exhaustive search of only half of some area at once. Or: "Bob isn't holding a key in his hands", and he'll only let you look at one hand at a time and can switch from hand to hand between searches.) Repeated testing will let you approach certainty about your theory (but never really reach it).

2. One could talk about Bayesian probability and priors. Especially if one varies the coin's heads/tails probability.

3. I think this sort of thing is an actual trope in human interactions. Some people give "hints" of various sorts; to hint proposition P might be defined as "to decide to do some action X such that, if not-P, then there's a small probability that a generic human in your position would do X, but if P, then there's a high probability that a generic human in your position would do X". Other people (well, probably the same people) may perform tests, to put some other person in a position where he might do X or something else, and if P were true, he would do X with high probability, but if not-P, then he would do X with low probability. And people who resent such "tests" may deliberately randomize (or genericize) their behavior.

Consider this: Imagine that someone has asked me a question to which, if P, then my answer is rather embarrassing, and if not-P, then my answer is not embarrassing; suppose lying is not an option. Suppose I am a naive person. Then, if not-P, I will simply give the non-embarrassing answer, and if P, then I will refuse to answer or something. Now, if everyone knows I am naive, and P is true, then refusing to answer is futile; everyone can see that and deduce P anyway.

Considering the above, I might decide to be a defensive person. I will recognize questions like that, and at least some of the time, I will simply refuse to answer even though the answer is non-embarrassing. It seems people do exactly this: refusing to answer "personal questions", as they are called, seems to be relatively common behavior. Even on the scale of just you and someone else, you can establish yourself as a defensive person by refusing to answer a personal question when the answer is innocuous, and maybe revealing the answer through some other means or at another time (be a bit careful: e.g. if you refuse to answer, he pressures you to answer, and you say the innocuous answer, then it would still seem to make sense next time you are pressured and still refuse to answer to deduce the embarrassing answer); then, the next time you refuse to answer a personal question, you have demonstrated that this doesn't necessarily imply that the answer is embarrassing.

It seems to me polite to act as though someone has won his status as "defensive" without making him work for it, and rude to do the opposite. Or to make it irrelevant. "May I ask <personal question>" makes it easy for the person to say "No" (much more so than just asking the question), which establishes that "no answer" might easily happen even if the answer is innocuous. (I also take the parallel approach that it shouldn't be embarrassing to tell me anything. I think embarrassment is the anticipated feeling that someone might tease you/put you down for something, and so if telling me something won't lead to that happening, then you needn't be embarrassed to tell me. Then my task is merely to make that true.)

Consider also encryption. Suppose people are naive. Then they won't encrypt anything (and do other things to make their information inaccessible to others) unless they have "something to hide". Then goddamn authorities can say "This person is hiding his stuff, so he is probably guilty, and that's all the evidence we need to justify using forceful means to make him reveal his secrets", and there is little defense against this. (In some situations, you may have an obviously "legitimate" reason to hide things, like hiding your company's information from your competitors. But there are other situations, like if the government comes and demands that information (ignore the possibility that they'll leak anything--I imagine the government will claim that's ridiculous, and I imagine most people would buy that), in which those reasons don't apply.)

Meanwhile, if people are defensive about encrypting and hiding their stuff, then statements like "If you had nothing to hide, then you wouldn't be hiding anything" are obviously false. Imagine trying to make that accusation of Bruce Schneier or Colin Percival. Now if only people like that were extremely common, and considered the norm. (Oh, another example is the "Don't talk to the police" crowd.)

If nothing else, I consider it rude to demand potentially embarrassing or incriminating information from someone who isn't willing to give it up, and use their response to get that information anyway; and I'd rather rude people not win out relative to non-rude people.

Good point. Obviously this becomes implausible once the mean reaches 3.5 (as it apparently did in the n=74 single die followup, with 3.4).

And yes, the data should be available for download.

That is one scary prospect. I think there are a bunch of people that would fall for such an offering. Look at how alot of the perks are already set up to comp you a drink or other such thing based on how long you have been playing. If you were to say the longer you play continuously, the better your odds,I think you would take every dollar an addicted gambler could scrounge. Sad.

Thats a great connection you've made between seemingly disconnected & distant disciplines. If we figure out the actual biological and chemical processes involved in a human body while it is indulging in synthesizing an 'original sin', that would be a big win in terms expanding our species' knowledge-base.

I'm sure there's research out there that explains at least fews bits of it through experiments where they monitor bodily processes when a person is lying. There are some common methods like the lie-detector and GSR methods, but apart from those external monitoring methods, probably also some molecular/cellular-level observations have been made.

From the paper (http://home.medewerker.uva.nl/s.shalvi/bestanden/Shalvi%20et...), page 5-6: "Shalvi et al. (2011a) asked participants to roll a die under a paper cup with a small hole at the top allowing only them to see the outcome, and earn money according to what they reported rolling (1=$1, 2=$2, etc.). As participants’ rolls were truly private, the authors assessed lying by comparing the reported distribution to the distribution predicted by chance (Fischbacher & Heusi, 2008). Participants were asked to roll three times but to report only the outcome of the first roll. Although all three rolls were private, the distribution of reported outcomes resembled the distribution of choosing the highest of the three observed rolls. Modifying the task to allow participants to roll only once reduced lying. Participants clearly found value in being able to justify their lies to themselves. The authors concluded that observing desired counterfactuals, in the form of desired (higher) values appearing on the second or third (non-relevant for pay) rolls, modified participants’ ethical perceptions of what they considered to be lying. Observing desired counterfactual information enabled participants to enjoy both worlds: lie for money, but feel honest."

Shalvi et. al. (2011a) is: Shalvi, S., Dana, J., Handgraaf, M. J. J., & De Dreu, C. K. W. (2011a). Justified ethicality: Observing desired counterfactuals modifies ethical perceptions and behavior. Organizational Behavior and Human Decision Processes, 115, 181-190.

Yes, that is interesting and true. The Economist article is a bit lame.

76 was enough to convince me that the effect isn't all publication bias (P<.01, and it's intuitively plausible). I would rather see a more diverse population (than college freshmen) than larger numbers.

The call for larger sample sizes isn't always appropriate. It can often lead to spurious inferences.

As Jacob Cohen (famous statistician) has said, "all null hypotheses, at least in the two tailed forms, are false".

That is, with nearly any hypothesis about differences between groups, given a large enough sample size, you're likely to find a significant difference.

Most likely. As another poster pointed out, what this shows is that our reason--given time to work--can override more instinctual desires to cheat. Part of rationalizing is considering situations, real and unreal, that may influence the outcome. E.g. I may be being observed and if they think I am cheating they may discard me from the pay group.

Without that time you wouldn't consider that "chosen at random" may be a lie. So, to the automatic mind your odds of getting paid are the same whether you pick 3 or 6 but the reward for lying is significantly higher. A simple heuristic the automatic mind may use would be (roll)x(payout): 6 x 10 = 60 vs 3 x 10 = 30

I'd be interested if the money incentive might have influence lying in the other direction, however. E.g. I did roll a 6, but if I say the maximum number they may not pay me; if I say a lower number they may be more likely pay me. This could lead to a new rational heuristic including a chance of being paid: 6 x 10 x 0.10 = 6 vs 3 x 10 x 0.50 = 15

This is what I'm thinking as well. The more they think about it the more they realize that the possiblity of cheating may be part of the experiment and decide not to do it because they don't want to look like a crook.

It's an interesting hypothesis but it doesn't seem to change the conclusion that "cheating is instinctive".

You mean that the most likely explanation for us hearing about this study is publication bias, right?

But we're talking p<.01 here, so most likely you're wrong. You'd have to believe that 100 similar studies have been performed and we only heard about the one winner.

They did two experiments that conceptually replicate research in the area. I agree that results like these need to be replicated by other researchers, but chalking this up to measurement or design error is not the most likely explanation.

This is a valid point, and I can only assume you are being downvoted due to "stereotyping." The sample group IS an important factor. For example, Milgram's famous "electric shock" experiment was repeated in many different countries. Germans for example, were very susceptible to authority figures, Australians were lesser so. Results are going to change based on the group being tested. It still doesn't change the fundamental finding of the study, however.

Probably more relevant to say this experiment was done on college freshmen, but yes, it would be interesting to see differences by country/etc. provided you can provide uniform experimental conditions (at a minimum, keep participants in the dark as to the composition of the experimental group, the fact that it's being done worldwide and comparisons will be made of their nation vs. others, etc).

You've clearly never been to Southern California.

In the paper in question (http://home.medewerker.uva.nl/s.shalvi/bestanden/Shalvi%20et...), it does. Chi-square test shows p<0.01 and p=0.02 for significance for the high and low time pressure groups, respectively.

After reading it I was going to say the same thing. What I don't understand is why they compared it to the statistical average rather than just recording what they rolled. It would have been fairly trivial to eliminate that particular statistical variability from the experiment.

n=76 also seems low to me for this type of experiment. There was a great New Yorker article titled "THE TRUTH WEARS OFF", which if I remember correctly, discussed the difficulties of reproducing scientific experiments because we underestimate the occurrences of statistical anomalies. Good time to read that article again.

http://www.newyorker.com/reporting/2010/12/13/101213fa_fact_...

76 is plenty for this. Read the paper before complaining.