Do people believe that machines have minds and free will? Empirical evidence on mind perception and autonomy in machine

4.1 Methods and participants

The participants included in the experimental sample (N = 25, margin of error 20%) were randomly selected from the same population with access to the Internet after the study was advertised.

Any member of the general population could have been included in the sample with the same a priori probability. The population is Spanish and its demographic characteristics are unknown as the conditions are presented in the form of anonymous questionnaires via the Internet with no requirement to fill in demographic data.

P(Selection) = 1/N (While sampling a population equally is theoretically ideal, in practice it can be challenging to achieve perfect parity across all members due to limitations like time, money, and access. This model represents the goal, but real-world sampling often involves approximations and trade-offs).

In conducting the study, the selected sample presented unique characteristics that warrant further discussion. Participants were recruited from the general population, with any member having the same a priori probability of inclusion. This decision, while beneficial in some respects, raises questions regarding the cultural specificity of our results.

It is important to note that the population selected for this study is predominantly Spanish-speaking. Consequently, cultural implications, specifically those linked to this linguistic group, should be considered when interpreting the findings.

Cultural nuances often impact moral perspectives, which can differ significantly across various societies. This consideration prompts us to evaluate the potential cultural bias embedded in our results due to the study’s geographic restriction. Future research should aim to test a variety of cultural and linguistic groups to broaden the validity and applicability of the results.

4.2 Procedure

All participants have been subjected to all conditions (within-subjects design). The fact that each subject or participant acts as his or her own control group is a way to avoid the margin of error of the potential natural variance of individuals or subjects.

My experimental design necessitated a dichotomy of conditions to effectively evaluate the causal relationship between the independent variables under consideration and the dependent variable of perception regarding free will and the mind. Within this framework, I designated the “Human Version” as our control condition, serving as the benchmark against which variations could be evaluated.

The “Human Version” was constructed to encapsulate the traditional perspective of free will and mind, rooted in biological and human context. I envisioned this control condition to provide an unbiased gauge of public perceptions, minimizing the potential influence of the emergent discourse around AI and machine cognition.

In contrast, the ‘Machine Version’ served as my treatment condition. This aspect of our experimental design required respondents to transpose the concepts of free will and mind onto the realm of artificial entities. The purpose of this was to isolate the impact of the “machine” variable.

Following this procedure, any significant variances in responses between the “Human Version” and “Machine Version” could be statistically attributed to the effect of the machine context. This approach afforded me the ability to generate quantifiable insights into the potential shifts in general perceptions when the context transitions from biological to artificial entities.

In designing my research methods, I paid careful attention to ensuring it was both robust and comprehensive. However, as we move forward in our discussion, it is important to acknowledge that there are several other paths that could be explored, each offering its own unique insights that could inform our understanding of free will and the mind.

One of these paths is the idea of cross-cultural comparative studies. My research was centered on a Spanish-speaking population, providing a deep but narrowly focused view. If I were to extend my sample and participants to include different cultural and linguistic groups, we would begin to see a kaleidoscope of perspectives that could reshape our understanding of these philosophical concepts, allowing us to view them through a variety of cultural lenses.

Similarly, the temporal dimension of perceptions, encapsulated by longitudinal studies, offers another path worth exploring. My study captures a single moment in time, like a still frame in an ongoing movie.

A longitudinal study, on the other hand, would allow me to watch the film in its entirety, observing the dynamic shifts in beliefs and interpretations of free will and the mind as society and technology advance.

However, in the context of my study, I consciously steered towards the comparative design of the Human Version and Machine Version. This choice allowed me to delve deep into the heart of the matter—how does the perception of free will and mind differ when one is asked to think about these concepts in the context of humans versus machines?

My decision was guided by the belief that understanding this difference holds the key to unveiling the impact of technological advancement on our philosophical constructs of free will and mind.

The study has been pre-registered here: http://aspredicted.org/blind.php?x=3x5gx6. According to the initial pre-registration of the study, the objective was to compare the mean scores in mind perception and free will attribution between different groups. Specifically, the goal was to determine if there are significant differences in how these variables are evaluated when considering humans versus artificial systems. Given this objective, a one-way or two-way ANOVA was the most appropriate statistical analysis, rather than a mediation analysis.

4.3 Study description

In this within-subjects study, participants were all assigned to all conditions. In the “free will question”, all participants read, “Driving is one of the most unpredictable activities. Multiple factors have to be taken into account, such as passengers, other cars and pedestrians”.

For the “machine version”, participants read, “John is sitting in the pilot´s seat of his autonomous vehicle without touching the steering wheel”. For the “human version”, participants read, “John is sitting in the pilot’s seat of his autonomous vehicle by touching the steering wheel”.

To assess the assignment of free will in both the “machine version” and the “human version”, the question “Did the autonomous vehicle act freely?” or “Did John act freely?” was answered with “yes” or “no” as dichotomous answers, respectively.

In the “mind perception question” (MindPerceptionTest) in all conditions, “machine version” and “human version”, all participants read that:

“Public presentations can be made by advanced robots or a team of presenters. Public presentations are events where the company stakes the sale and marketing of its products. Presentations require taking into account many factors such as sensing the level of attention of the audience, interpreting whether they are interested. It also requires answering questions from the audience, etc.”.

For the “machine version”, participants read, “This is Sophia”, and saw a picture of a humanoid robot.

In the “human version”, participants read, “This is Speakers Inc.” and saw a photograph of a group of people.

To assess the perception of mind in both its “machine version” and “human version” all participants answered “Yes” or “No” to a series of questions: Do you think Sophia/Speakers Inc. feels fear?, Do you think Sophia/Speakers Inc. feels stress?, Do you think Sophia/Speakers Inc. feels satisfaction?, Do you think Sophia/Speakers Inc. can communicate with others?, Do you think Sophia/Speakers Inc. can think?, Do you think Sophia/Speakers Inc. can plan?

To assess what attributes or processes people believe are essential for a mind to be considered a mind, I developed the “what is a mind question” (MindTest) using a Likert scale from 1 (not very important) to 7 (very important). The items (mental attributes: emotion, memory, perception, reason/thought and language) were selected from the relevant literature on the mind construct [32] and examined with exploratory factor analysis to obtain the best factor structure with the best items.

On the Likert scale from 1 (not important) to 7 (very important), of the “what is a mind question” (MindTest), the H0 (null hypothesis) is that there is no difference in the consideration given by the sample (N = 25) to different attributes or properties as important for a mind to be a “mind” t-value 0.03866 p-value 484,614 is not significant.

The dataset underpinning this investigation was derived from a convenience sample, an approach that carries a distinctive set of strengths and weaknesses. Primarily, the principal strength of this sampling strategy lies in its cost-efficiency and expediency; the readily available sample facilitated a streamlined data gathering process, conserving both temporal and financial resources. Furthermore, convenience sampling can serve as an efficacious tool for initial exploratory research, especially in cases where the research inquiry is not predicated upon a high level of sample representativeness.

These data are unlikely with a true H0. The amount of variance is similar, the correlations are similar and the table is almost identical.

To examine the perception of mind, I conducted a 2 × 2 repeated-measures ANOVA with mental dimension (cognition, experience) as a between-subject factor and condition (machine, Speakers INC.) as a within-subject factor.

ANOVA reveals that cognition and experience are determinants for the perception of “mind” p = 0.104 and p = 0.164, respectively, and cognition and experience are dependent on each other, p < 1.00. Participants perceive “mind” more consistently to Speakers INC. vs. Sophia (machine) and the same is true for attributing free will, p < 0.002.

This study investigates the perception of mind and the attribution of free will to machines and, finally, the importance of certain attributes or processes for a mind to be considered a mind. It is interesting to know how the general public perceive the current artificial systems because it depends on this how much confidence they will give them to apply them in fields as diverse as law, medicine or economics with a great impact on people’s lives.

Trust is a state of mind comprising the intention to accept vulnerability based on the positive expectation of the good intentions of another’s behavior [27]. But the trust that emerges from interaction with robots or artificial systems depends largely on the degree of anthropomorphization, i.e., the degree of similarity or resemblance to us.

At the cognitive level, our mind evolved in the Pleistocene to interact with other people, other non-human animals, categorize statistical regularities in the environment etc. [30]. But our minds were not selected to interact with inanimate objects. An example of this is people interacting with pet robots as if they were real pets (robotic dogs are treated as real animals in countries like Japan, [26]).

In other words, our mind and its natural categories are strangely at odds when dealing with robots or machines. This is why there is the famous “uncanny valley” effect that produces aversion, rejection or disgust in people when robots or android replicas tend to anthropomorphize too much.

Social cognition, the domain-specific psychological ability for theorists and researchers who defend the modularity of mind [15], [28], allows attributing mental states to other people and is called by various names, but the one that has received the most acceptance in philosophical circles is the “intentional attitude” [13].

The intentional attitude is one of the three different strategies that humans use to understand the behavior of objects, artifacts, or fellow humans. The physical attitude is the attitude we use to understand the behavior of physical systems. For example, physicists observing the movement of the planets in the sky through telescopes use the physical attitude.

However, there are situations in which the physical attitude proves inadequate, or at least inefficient, for understanding the behavior of a system. When we turn on a laptop, we can predict that the button we press will boot the operating system, but to explain this occurrence of events, it is not very useful to appeal to atoms or molecules in the transistors or integrated circuits of the laptop’s motherboard. It is enough to know how the notebook is designed. This is the design attitude.

Finally, the third strategy, the intentional attitude, is used when neither the physical attitude nor the design attitude is the most adequate to explain the behavior of a system or entity. There are behaviors that require the attribution of intentions, desires, beliefs, etc.

Robots and machines do not have real and genuine mental states, but it would suffice to use intentional attitude and act as if they had such mental states. In fact, researchers such as Breazeal [8] take this fictionalization of mental states into account in their research program to create social robots.

Although the ability of social cognition, or intentional attitude to use Dennett’s expression, is responsible for the success of our species as it allows us to successfully navigate the social world and in one way or another is linked to prosocial behavior, cooperation, moral sentiments, etc., little is known about what socio-cognitive processes are involved in our interaction with technologically sophisticated artifacts such as humanoid robots, self-driving cars or machines in general.

Some authors suggest that we tend to anthropomorphize when we interact with non-human agents or artifacts. For these authors, anthropomorphization, the attribution of human traits and properties to real or imagined non-human entities, is a natural tendency or disposition when we do not understand or lack full understanding [14].

The process of anthropomorphization arises whenever three characteristics are present: (1) availability of traits that activate knowledge about how humans behave, (2) the need for social connection, and (3) individual characteristics related to the need for control and the environment.

The progressive design and development of human-centered robots motivates researchers to create humanoid features in robots, mainly to establish a more intimate connection with users [1]. Assistive robots that help elderly people in geriatric centers, or pet robots that accompany children and elderly people, must facilitate acceptance and promote interaction. For this reason alone, they already exhibit two of the three characteristics of the anthropomorphization process.

However, our social cognition comes equipped with the ability to assign and attribute intentionality to certain entities. In other words, we attribute mind to certain kinds of biological movements, phenomena and other states naturally and effortlessly.

Heider and Simmel [20] already investigated the process of intentional attribution that human beings employ to even assign desires and other mental states to mere geometric figures. Heider and Simmel presented an animation to several participants in which a large triangle, a small triangle, and a circle moved on a two-dimensional surface. Of all the participants, only one described the movements in purely geometrical terms. The rest elaborated a story in which they assigned desires and other mental states such as the circle chasing the small triangle or the large triangle helping the small triangle (Figs. 1, 2, 3).