Материал: искусственный интеллект

32 A. Lambert-Mogiliansky et al.

growth and expansion. Tiger Forever was founded 2006 with the goal of reversing global tiger decline. It is active in 17 sites with Non-Governmental Organizations (NGOs) and government partners. The sites host about 2260 tigers or 70% of the total world’s tiger population”.

It is worth mentioning that the descriptions were formulated so as to slightly suggest that the elephants’ NGO (ECF) could be perceived as being more trustworthy (because of its link with of WWF, a well-known NGO). The text about tigers in contrast suggests a higher level of urgency (a mention was made of the absolute number of remaining tigers, significantly lower than the number of remaining elephants). Thereafter all respondents were asked:

“When considering donating money in support of a specific project to protect endangered species, di erent aspects may be relevant to your choice. Let us know what counts most to you”. It followed:

“The urgency of the cause: among the many important issues in today’s world, does the cause you consider belong to those that deserve urgent action?” or “The honesty of the organization to which you donate: do you trust the organization managing the project to be reliable; i.e. do you trust the money will be used as advertised rather than diverted.”

The objective was to elicit their preferences in the sense of which perspective was most important. The rest of the experiment depended on which one of three groups the participants belonged to.

In the first control condition (baseline) they were next asked whether or not they wanted to read the first descriptions again or if they wanted to make their final decision i.e., to choose between supporting Elephant Crisis or Tiger Forever both represented by an image of an adult tiger respectively adult elephant (also delivered in random order on a line).

In the first treatment condition, they were redirected to a screen with general information compatible with the aspect they indicated as determinant to their choice when making a donation. Importantly the information did not favor or

disfavor directly or indirectly any of the two projects. Rather, the Honesty extrainformation dealt with NGOs’ integrity in general, and the Urgency one with global mass extinction. Then they were o ered the opportunity to read again the

descriptions before deciding or directly make their image choice between ECF and TF. Those who responded honesty saw:

“Did you know that most Elephant and Tiger projects are run by NonGovernmental Organizations (NGOs)? But NGOs are not always honest ! NGOs operating in countries with endemic corruption face particular risks. NGOs are created by enthusiastic benevolent citizens who often lack proper competence to manage both internal and external risks. Numerous scandals have shown how even long standing NGOs had been captured by less scrupulous people to serve their own interest.

So a reasonable concern is whether Tiger Forever or Elephant Crisis Fund deserves our trust.”

Those who responded urgency saw:

“Did you know that global wildlife populations have declined 58% since 1970, primarily due to habitat destruction, over-hunting and pollution. It is urgent to reverse the decline! “For the first time since the demise of the dinosaurs 65 million years ago, we face a global mass extinction of wildlife. We ignore the decline of other species at our peril – for they are the barometer that reveals our impact on the world that sustains us.”—Mike Barrett, director of science and policy at WWF’s UK branch. A reasonable concern is how urgent protecting tigers or elephants actually is.”

In the main treatment condition, participants were redirected to a screen with general information on the aspect they did not select as determinant. So those who selected honesty (resp. urgency) saw the screen on global wildlife decline (resp. NGO’s scandals). Then they were o ered the opportunity to read again the description before deciding or directly make their image choice.

Finally, information about their age, gender, education and habits of donation to NGOs was collected before the thank-you message and the end of the experiment.

1371 participants completed the survey on the website Typeform and were recruited through Amazon’s Mechanical T¨urk. They were paid either $1, $0.8 or $0.75 (for the shorter baseline survey) depending on their conditions. They spent on average 1:33 min for the experiment. 49% of them were females, 61% males, the mean age was 35, and their mean education level was undergraduate level.

3.3Predictions

Before getting into the results and their interpretation. Let us remind of what the main predictions are.

First, the predictions of both the Bayesian and the quantum model regarding treatment group 1 (who received compatible information) are similar to the extent that general information should have a marginal or no e ect at all compared with the control group’s choice of whom to donate.7 This treatment group allows rejecting the argument that any additional information upsets people’s beliefs so as to significantly a ect their choice.

In contrast the predictions of the two models regarding the main treatment group are distinct. The Bayesian model predicts that general information on an issue that is not determinant to choice should have no e ect or a very small counter-balancing e ect. That e ect would be due to the fact that even if say “trustworthiness” is determinant, it needs not mean that urgency is irrelevant. In contrast, the quantum persuasion model predicts that the distraction provided to the treatment group could significantly modify the allocation of responses compared to the control group. It should be noted that since we lack information

7In a companion paper, we investigate in details the mechanisms behind those predictions. In the quantum case we do have e ects due to the measurement but they tend to neutralize each other. In the classical case the e ect if not null, is small and depends on the initial conditions.

34 A. Lambert-Mogiliansky et al.

about the exact correlation coe cients between the two perspectives, we do not have precise quantitative predictions.

4 Results

Data were processed, cleaned and analyzed with Stata. Mainly due to missing values, but also to solve a technical misstep8 and in order to equally balance the number of participants in each condition, 471 participants had been removed from the data. Probit regression models were conducted to analyze the impact of the variables of interest.

4.1Descriptive Statistics

As shown in Table 2, overall, 72,7% of the participants valued the Honesty of the NGO more than the Urgency of the cause, 87,2% made their final decision without reading the descriptions a second time, and 54,6% voted for Elephants Crisis Fund (ECF). Furthermore, after further distinctions, we observe that while most of participants preferred elephants to tigers in the control condition and in the compatible one (59% and 56% respectively), the tendency reverses for the incompatible condition (51% chose tigers). Regarding the revealed preferences, overall, the majority of participants who preferred Urgency chose Tigers (52%), whereas the majority of those who preferred Honesty chose Elephants (57%).

Table 2. Descriptive statistics

Notes. ChoiceHU-choice between Urgency (=0) and Honesty (=1); DecisionRead-decision to read the descriptions again (=0) or not (=1); FinalChoice-final choice between Tigers (=0) and Elephants (=1); Gender-females (=0) and males (=1); Education-highest level of formal education between secondary school (=0), high school (=1), undergraduate (=2), graduate and over (=3); NGO-donation of either nothing (=0) or something (=1) in the last 3 years.

8Some participants were likely to have taken the questionnaire twice and so were deleted.

4.2Data Analysis

As Table 3 shows, the first set of results establishes that incompatible information has a statistically significant impact on the final choice (p-value = .011), whereas the compatible information did not significantly lead to di erent results compared to the baseline (p-value = .314). The e ect of the incompatible condition on the final decision thus seems to be as expected i.e. it significantly reverses the direction of the choice observed in the control condition. More precisely,

everything else being constant, the predicted probability of choosing Elephants is 10.52% (marginal e ect ) lower for an individual in the incompatible condition.

Table 3. Regression matrix for Final Choice

p-values in parentheses

= p ≤ 0.05, = p ≤ 0.01, = p ≤ 0.001

Not surprisingly there is also a significant impact on the final decision of the choice of determinant (p-value = 0.024) i.e., honesty versus urgency which captures the preferences. In other words, those who claimed to prefer Honesty

chose Elephants significantly more than those who preferred Urgency regardless of the condition in which they have been assigned(p-value = .025). Everything

36 A. Lambert-Mogiliansky et al.

else being constant, the predicted probability of choosing Elephants is 8.44% (marginal e ect ) higher for an individual who preferred Honesty. All of these

e ects remained significant when the other variables were included or removed from the regression. None of the control variables (order of the descriptions included) seemed to significantly influence the final decision, which means that the di erence in decision-making were essentially not due to the sample heterogeneity. Similarly, the decision to reread descriptions did not significantly impact

the final choice (p-value = .740). Both the compatible and incompatible condition significantly lead to a tendency to read the descriptions again (p-value =

.001 and p-value = .006, respectively). In other words, the more information, the more one reads previous information again. We are currently working on further investigation of the data in a companion paper.

4.3Interpretation

In line with our hypotheses, our results show with no ambiguity that incompatible information - that is “distraction” - has a significant impact on the final choice by inducing some extent of switch as compared to both the control group and the compatible information group.

These results are fully consistent with the predictions of the quantum persuasion model and contradict the predictions of the Bayesian model with respect to the impact of incompatible information. Moreover the fact that general compatible information had no impact also supports the view that it is not merely “information” that a ects the choice because the person is slightly “upset”. Instead it is when information induces a change in perspective that something happens even though nothing of relevance is learned.

In addition, the participants’ age, gender, level of education or experience with NGOs had no e ect on the decision to vote for ECF or TF. The final choice seemed to depend only on the descriptions, the conditions and participants’ own beliefs and preferences. We can therefore conclude that our distraction e ect – or change of focus – is quite stable among individuals. This supports the hypothesis that the quantum-like structure is a general regularity of the human mind.

The importance of elicited preferences i.e., the answer to “what is determinant to your choice” to the final choice underlines that the initial texts were well-understood. The description of the Elephant project was designed to suggest more trust to the NGO, while the Tiger project aimed at suggesting higher level of urgency. That explains why respondents who declared Honesty (resp. Urgency) to be determinant were significantly more likely to support the Elephant Crisis Fund (resp. Tiger Forever).

The average time to respond to the questionnaire was between 1 and 2 min, which is rather quick. In addition only a tiny proportion of participants (15%) actually used that opportunity to reassess their understanding of the project by rereading the projects descriptions. These two facts support the idea of an absence of conscious reasoning, that is, the respondents did not take time to reflect and reacted spontaneously to the distraction. This is particularly interesting for us since the quantum working of the mind is not rational reasoning: