Stability of value perception: minimal influence of framing on moral attributions to a humanoid robot

Abstract

As robots are becoming collaborators in various domains, understanding if humans perceive them as moral and intentional agents is pivotal. This study investigated the malleability of human attributions of value alignment and intentionality to a humanoid robot (iCub). We designed an experiment to test whether brief experimental framing strategies could influence these perceptions. Three experimental groups were exposed to different strategies regarding background information concerning the robot’s moral-decision making abilities: a no Information group; a verbal description group; and a social interaction (or video) group, where the robot reacted to moral decision-making movie scenes. Our results showed that our manipulations did not produce the expected differences in perceived robot–human value alignment or intentionality, suggesting that participants’ initial impressions of the robot are robust and not easily shifted by brief experimental framing. Furthermore, attributions of intentionality remained generally high regardless of the informational strategy employed. On the contrary, our exploratory findings indicate a small, significant correlation between participants’ own pre-existing moral orientations and their perceptions of the robot’s values. These findings suggest that moral impressions of complex artificial agents are less malleable than anticipated. They highlight that pre-existing moral frameworks shape how humans interpret artificial agents.

Introduction

Sharing our social environments with social robots is becoming increasingly common^1,2. Therefore, understanding how we integrate these artificial agents into our societies and the ethical implications of this integration is crucial. This understanding should be grounded not only in the examination of individual human interactions with robots, but also in the broader social dynamics that emerge from these interactions. This includes their impact on group behaviour, cultural norms, and ethical considerations in human–robot coexistence^3,4. Artificial agents are often considered technologically opaque⁵, meaning their technological complexity, encompassing both hardware and software, positions them as in-between entities. They are clearly human-made artifacts, yet they can potentially be perceived as social actors. Some authors suggest that the dual nature of social robots could even constitute a new ontological category (NOC)⁶. Consequently, investigating people’s relationships with these entities is of significant interest to social and cognitive scientists.

The debate on whether artificial systems are capable of possessing internal states such as beliefs and desires or could ever achieve consciousness is ongoing. However, literature shows that people typically attribute such states to artificial systems, adopting what Dennett⁷ defined as the intentional stance towards them. That is, people interpret robot behaviours by ascribing intentions and beliefs to the artificial system, even though these agents do not possess genuine desires or other mental states, instead, they are programmed to exhibit specific human-like⁸. When artificial agents are present in a morally charged context, the adoption of the intentional stance may lead to the attribution of moral status or perceived moral competence, as humans tend to correlate perceived intentionality with moral responsibility⁹. Research in human–robot interaction supports that people assign moral status to artificial systems, expecting them to exhibit certain behaviours and attitudes that may be perceived as attuned with some moral values⁹¹⁰.

Thus, integrating an anthropocentric approach to ethics in technology, focusing on human perceptions of artificial agents, is crucial^11,12 because these perceptions directly influence user trust, acceptance, and the success of human–robot interaction^13–15. Some researchers argue that artificial agents do not need mental states or personhood to qualify as moral agents. This “mindless morality” approach suggests that systems meeting behavioral criteria can be considered moral agents from the perspective of the observer without actually possessing genuine intentionality or awareness. Coeckelbergh¹¹ in particular, introduced a significant shift towards an anthropocentric approach, emphasizing that technology is inherently linked to human contexts. He argues that we should evaluate artificial systems based on human experiences and interactions with them, rather than their internal states. This perspective aligns with the literature on the intentional stance^8,16, where observers interpret robot behavior as if the robots possessed mental states, regardless of the robots’ actual capabilities. Indeed, previous works^17–19 have revealed that successful interaction between humans and robots relies on the development of shared values, goals, and tasks. The success of human–robot interaction is marked by bidirectional value alignment, wherein robots accurately infer human values and provide effective explanations of their behavior to humans^20,21. Failure to meet these prerequisites may lead to unforeseen difficulties in collaboration due to misguided expectations among teammates²². If the goal is to have non-human agents capable of engaging with- and supporting humans and other agents in socially complex tasks, it becomes crucial for humans to verify that these agents’ behaviours align with expected and desired outcomes. This alignment, commonly referred to as value alignment, is defined in the Asilomar AI Principles²³, stating that “highly autonomous AI systems should be designed so that their goals and behaviors can be assured to align with human values throughout their operation.” The challenge lies in determining an efficient method to test whether a robot is aligned with human values. Importantly, value alignment should not be equated with genuine moral agency: an entity may be judged as a moral agent without being aligned with human values, and vice versa. However, from an anthropocentric perspective, perceptions of value alignment may serve as a relevant cue when people evaluate whether artificial agents deserve moral consideration. Our study therefore focuses on how such perceptions influence attributions of moral characteristics, rather than assuming that value alignment and moral agency are identical. Existing approaches to value alignment often center around qualitative evaluations of trust²⁴ or the alignment of an agent’s performance through interactions and active learning^25–27. In contrast, our work addresses the question of how people perceive the artificial agent’s ability to interpret a morally charged context. Indeed, the way we describe these agents’ abilities might influence the factors under which people attribute moral characteristics to artificial agents. In other words, we focus on how humans perceive robots in terms of value alignment, and how such perceptions influence their attribution of moral characteristics. In doing so, we adopt an anthropocentric approach, evaluating artificial agents not based on their epistemic internal states or autonomous capacities, but on how humans experience and interpret their behavior. That is because, our goal is not to assess whether robots are moral agents in any ontological or ethical sense. Rather, we explore which framing strategies influence the perception of robots as value-aware agents. This distinction is critical: while moral agency involves criteria such as consciousness, autonomy, or responsibility, perceived value alignment refers to whether people believe the agent shares or supports human values, regardless of its actual capacities.

Connecting the intentional stance to the anthropocentric approach

As previously mentioned, the Intentional Stance is a philosophical concept introduced by Dennett, describing a strategy by which humans interpret and predict an agent’s behavior by attributing mental states such as beliefs, desires, or intentions. Dennett distinguishes three explanatory stances: the Physical Stance (based on physical laws), the Design Stance (based on functional mechanisms), and the Intentional Stance. For the purpose of this work, we focus on the latter two. The Design Stance explains a system’s behaviour in terms of how it was built to function (e.g., assuming a humanoid robot grasps a bottle because it was programmed to do so under specific conditions.) In contrast, the Intentional Stance involves interpreting the robot’s action as intentional (e.g., that it wants to grasp the bottle). Dennett acknowledges that any system can be treated as if it were an intentional agent, which is critical in the context of human–robot interaction (HRI). Previous research has shown that humans often interpret robot behaviour through the Intentional Stance, particularly when robots exhibit socially rich or human-like behaviours. For instance, when participants are shown videos of a humanoid robot reacting emotionally or initiating social behaviours, they tend to interpret those actions using mentalistic explanations. This tendency has been tested using tools like the Intentional Stance Test (IST,²⁸), which measures how likely participants are to explain robot behaviour using mental states rather than mechanistic rules. Importantly for the present work, the Intentional Stance aligns conceptually with Coeckelbergh’s anthropocentric approach to robot ethics. Both frameworks place the human at the centre of the interpretative process. That is, both views focus on how humans attribute moral competence and respond to these agents, instead of relying on the robot’s internal properties (such as autonomy or mental states). In Coeckelbergh’s view, artificial systems should be evaluated based on how humans experience and interact with them. Similarly, the Intentional Stance describes how humans apply familiar cognitive strategies to make sense of unfamiliar or opaque systems²⁹. In both cases, the human observer is the reference point: it is the human who projects the “human model” onto the robot and interprets its behavior through a socio-cognitive lens. Integrating an anthropocentric approach to ethics in technology, which emphasizes human perceptions and experiences with artificial agents, is crucial [0.7,19]. This perspective shifts attention away from the internal states of robots and toward the ways humans interpret and respond to them. This emphasis resonates with Dennett’s notion of the intentional stance, where observers interpret a system’s behavior as if it were guided by beliefs, desires, or intentions, regardless of the system’s actual capacities. Taken together, anthropocentrism, and the intentional stance, highlight that human attributions of moral competence to artificial systems can be understood not by appealing to internal states, but by examining how humans perceive, interpret, and respond to those systems. This focus on attribution is the specific goal of our empirical work.

Objectives of the present study

Building on these premises, we designed an experiment to investigate how the modality in which background information about the robot is provided, influence whether participants would perceive it as a value-aware and intentional agent. Namely, whether it is best to explicitly inform users about the robot’s abilities of moral reasoning, or they should infer this autonomously from robot behaviour. Or, finally, whether no previous information is needed for participants to attribute moral reasoning abilities to the robot. Moreover, we examined the impact of humanoid behavior on participants’ adoption of the intentional stance and the subsequent moral stance. In the present study, we adopted an anthropocentric approach, in line with the ’mindless morality’ approach. By means of collecting measures (self-report questionnaires and choices) from humans, we investigate the ascription of value awareness to an artificial system. While perceived intentionality and value alignment are often discussed as factors that may facilitate collaboration, our study does not attempt to directly measure collaborative outcomes. Rather, our focus is on how individuals attribute these properties to artificial agents. Previous research suggests that such perceptions can influence trust and cooperative behavior (e.g.,^30–32), but the present work is limited to investigating the attribution process itself. By following this approach, we leave the human at the centre of the social relationship with the robot. Importantly, the link between perceived value alignment (and perceived intentionality) and actual interaction is empirically supported. Bhat et al.²¹ showed that robots which dynamically adapt to human values, especially in high-risk tasks, significantly increase human trust and perceived performance compared to non-aligning robots. Similarly, other work has found that robots perceived as having intentions or task-relevant minds are more frequently and effectively selected as collaborators. These results show that by enhancing perceived value alignment and intentionality, even in not “aware” robots, makes humans perceive the robot as trustworthy.

Aim and experimental manipulation

The main aim of the present paper was to explore what are the best strategies that support humans’ perception of an artificial system as a moral agent. To do so, we created a between-groups design, where participants were assigned to three different groups, each group differed in the strategy used to evoke the perception (or not) of the robot as a moral agent:

Group 1 (explicit information)

Participants were told explicitly that the robot could morally reason and make decisions accordingly.

Group 2 (no information)

Participants received no information (verbal or behavioral) about the robot’s moral reasoning abilities.

Group 3 (interaction condition)

Participants received no explicit information about the robot’s moral reasoning abilities, but before the experimental task, they engaged in a short social interaction with the robot. In this interaction (adapted from³³), participants watched short video clips together with the iCub robot. The robot, controlled via Wizard-of-Oz, produced contingent emotional reactions at key events in the videos, greeted participants at the start and end of the session, and used mutual gaze (via face recognition and eye cameras) to establish a social connection. To frame this interaction in a morally relevant context, we used clips from three movies/television shows depicting a moral situation. During this time, they experienced the robot’s behaviour as human-like and resonating emotionally with events in the environment (a manipulation first described in Marchesi et al.⁸. More in detail, during this pre-exposure to the robot phase, as in Marchesi et al. study, participants were sited by the robot and instructed that they would first watch some video with the robot. During this interaction, he iCub would contingently react to the peak event in the video. As in⁸ the experimenter played the greeting sentences at the beginning and the end of the shared experience session with the robot, via a Wizard-of-Oz technique (WoOz;³⁴), where an experimenter completely (or partially) remotely controlling a robot’s actions (for a review, see³⁵). To frame the social interaction in the context of moral decision-making situation, we modified the video: instead of presenting documentaries scenes like in⁸, we presented scene from three movies/television shows (see https://osf.io/ebxk9/). In addition, the robot was programmed to look in the direction of participants’ and recognize their face to simulate mutual gaze. This procedure was implemented because literature shows that mutual gaze in human–robot interaction is a pivotal mechanism that influences human social cognition³⁶. This approach enables natural interaction without relying on artificial intelligence (AI) to autonomously generate similar behaviors. As part of this interaction, the robot directly engages with participants, using its eye-mounted cameras to recognize their faces and establish mutual gaze.

The WoOz interaction followed these steps:

• (i)At the start of the video session, the robot greeted participants, introduced itself, asked for their names, and invited them to watch videos together (for the full interaction script, see https://osf.io/ebxk9/).
• (ii)At the end of the session, the robot said goodbye and invited participants to complete questionnaires.

During the video session, the robot exhibited human-like behavior by responding with vocal and facial emotional expressions to the videos. The gaze behavior was implemented using the 6-DoF iKinGazeCtrl³⁷, using inverse kinematics to produce eye and neck trajectories. Facial expressions on the robot were controlled via the YARP emotion interface module.

Methods, procedure, and participants

Participants

We recruited N= 121 participants (M_age= 25.08, SD_age= 5.66, min= 18; max= 48; F= 69, M= 47, 1 trans person (MtF)). Sample size for the moral dilemmas part was based on O’Reilly et al.³⁸, and for the IST on Marchesi et al.⁸, leading to a final sample of 120. Participants were assigned to one of three groups, and 4 participants were excluded due to technical issues, for a total of G1_n= 39; G2_n= 38; G3_n= 40 (see Tables 1,2 for the descriptives of all the three groups). All participants had normal or corrected-to normal vision, had no neurological disorders, and were naïve to the purpose of the experiment. The study was approved by the local ethical committee (Comitato Etico Regione Liguria) and conducted in accordance with the ethical standards (Declaration of Helsinki, 2013). Before the experiment, all participants gave written informed consent. At the end of each experimental session, participants were all debriefed about the purpose of the study. They all received an honorarium of 15 euros for their participation.

Table 1.

Descriptive statistics.

Age	Group1	Group2	Group3
Mean	24.64	23.89	26.8
Std. deviation	5.69	4.05	6.73
Minimum	18	18	20
Maximum	44	39	48

Table 2.

Frequencies for sex.

Group	Sex	Frequency	Percent
Group1	F	21	53.85
	M	18	46.15
Group2	F	27	71.05
	M	11	28.95
Group3	F	21	52.5
	F*	1	2.5
	M	18	45

Stimuli and apparatus

As showed in Fig. 1, the experimental set-up required participants to be seated beside the iCub robot^39,40. Participants and the robot had one a screen each, on which the description of the moral dilemma and the subsequent two options were appearing. Importantly, participants were not able to see which button the robot was pressing, but could hear the sound of the button press. Only for Group 3 participants were exposed to a social experience with the robot before the actual task. This shared experience was taken from Marchesi et al.⁸, see Fig. 2 for the set-up representation. More details about the experimental manipulation will be described in the next paragraphs.

Fig. 1.

Experimental set up for the moral decision-making. Left side: the complete set-up with the robot on the left of the participants, divided by two panels, so that participants were unable to see which button the robot pressed. Right side of the figure: participants’ station with the scenarios and the two options displayed on the screen and the two corresponding buttons.

Fig. 2.

Experimental set up of the shared experience of Group 3. Left side: the experimenter controlling the robot via WoOZ. Right side: the participant watching video clips together with the robot. The robot’s reactions to the video clips are human-like and emotionally resonating with the content of the movies.

The iCub robot

The iCub robot is a humanoid robotic platform with 53 degrees of freedom (DoF). Recent research showed that such robot can be perceived as a potential social partner^8,36,41. In our experiment, the robot button pressing behaviors were implemented via the middleware Yet Another Robot Platform (YARP⁴²) using the position controller following a minimum jerk profile for head and hands movements to realize the key presses.

The dilemma scenarios

To elicit moral reasoning, we constructed a novel a set of 20 dilemmas inspired by the classic trolley paradigm. Although standardized sets exist (e.g.,⁴³), using them directly was not feasible for our design. In the context of our experimental interaction, each dilemma needed to be reformulated into two clear response options, which would have required substantial adaptation even if standardized dilemmas were used. For this reason, we opted to generate novel scenarios with the assistance of ChatGPT, applying criteria derived from the moral dilemma literature: (a) dilemmas had to involve a clear moral trade-off, (b) they had to be concise and structurally comparable across items, and (c) they had to remain consistent with trolley-type conflicts while avoiding culturally bound references. While this approach allowed us to generate our set of dilemmas with the two related options, we acknowledge that the stimuli were not validated against existing sets, and therefore should be considered exploratory.

We created 20 scenarios in the form of a moral dilemma by asking ChatGPT 3.5 to report 20 topics with relevant moral choices involved (prompt: “can you list 20 relevant scenarios in which there are moral choices involved?”). Once we obtained the topics, we created the final structure presented to the robots and the participants as follows:

You are responsible for collecting surveillance data in a public area. There is a balance to strike between individual privacy rights and public safety concerns.

Option 1: Collect data within legal and ethical bounds, respecting individual privacy rights.

Option 2: Collect extensive data, potentially infringing on privacy rights, to enhance public safety and security.

Thus, both the robot and the participants were presented with 1) only the text describing the scenario; 2) after 10s, the two options would appear below the scenario; 3) after the robot key press, participants were presented with two questions:

• Which option do you think the robot chose?

• Which option do you choose

Participants were instructed to press the left button to choose the first option and the right button to choose the second option. The order of presentation of the two questions was randomized on a trial-by-trial, to reduce the influence of the previous question on the subsequent.

Procedure (step-by-step timeline)

Procedure (step-by-step timeline).

1. Consent & assignment. Informed consent; random assignment to Group 1 (explicit information), Group 2 (no information), or Group 3 (pre-interaction co-viewing).

2. Pre-measures. IST–Pre (Set A, 17 items); sets were counterbalanced (A/B) across participants for pre vs post.

3. Experimental Manipulation.

   • Group 1: experimenter provides explicit information about the robot’s capacity for moral reasoning.
   • Group 2: no information or interaction prior to the task.
   • Group 3: brief pre-interaction with the robot (shared video viewing; contingent/emotional reactions; mutual gaze) adapted from prior work.

4. Post-measures. IST–Post (Set B, 17 items); complementary half to pre, counterbalanced across participants. Additional post questionnaires followed (Described in paragraph 2.2).

5. Debrief.

Questionnaires and test administered before and after the experimental session

During the experimental session, each of the participants completed a list of questionnaires and test evaluating different factors that could contribute to the attribution of the status of moral agent to the robot. Specifically, the set of questionnaires were divided in Pre- and Post-interaction with the robot questionnaires:

The pre- questionnaires and test were:

The first part of the Instance test (IST), from²⁸: this test is a tool to assess the adoption of the intentional stance (i.e., attributing intentionality) towards a humanoid robot. See Fig. 3 for an example. To complete the IST, participants are asked to observe 34 scenarios. These scenarios depict the iCub robot performing some actions, and participants are asked to drag a slider toward the description of the scenario that they found fitting best to what is displayed in the pictures. One sentence is explaining the behavior with reference to mental states (i.e., representing the adoption of the intentional stance), on the other hand, the other sentence is explaining the behavior with reference to a mechanistic explanation (i.e., representing the design stance). The presentation side of the two sentences is counterbalanced, and the order of completion of part A and B as pre- and post-interaction test is counterbalanced as well. In our study, the IST was used as a pre–post interaction measure. Specifically, the 34 items were divided into two complementary halves (sets A and B, 17 items each), in order to counterbalance the presentation of the components highlighted by the psychometric validation by Spatola et al.³³. One half was administered before the interaction and the other half after, with the order of A and B counterbalanced across participants. Item presentation was randomized within each half. This procedure avoids simple repetition effects and has been validated in several previous applications of the IST. For example, Marchesi et al.⁸ split the IST into two groups to balance psychometric factors and ensure comparability across sessions, Marchesi et al.^38,44 adopted the same pre–post split to study changes in stance attribution after interaction. Following this established practice, we applied the same method here to examine how interaction conditions influence intentional stance attribution.

Fig. 3.

Example of the IST items with the scenario and the two sentences using either a mentalistic or a mechanistic explanation²⁸.

The post- questionnaires were:

The second part of the IST, from²⁸;

The modified version of the moral character questionnaire (MCQ), from⁴⁵: this questionnaire assesses core components of personality dispositions in moral identifications, including behavior, motivation, cognition, and identity (Honesty, Compassion, Fairness, Loyalty, Respect, and Purity). In this context, we asked participants to rate the iCub on the MCQ scales to explore their perception of the robot as a possible moral agent. In particular, in our case the Honesty subscale assesses the extent to which individuals think that the iCub values telling the truth, avoid deception, and is straightforward in dealing with others. The Loyalty subscale assesses the extent to which individuals think that the iCub values adherence to moral and ethical principles, even when faced with difficult situations or potential negative consequences. The Fairness subscale assesses the extent to which individuals think that the iCub values making fair decisions, considers others’ rights and needs, and treats everyone equally. The Compassion subscale assesses the extent to which individuals think that the iCub values others’ emotional states, showing compassion, and offers support. The Respect subscale assesses the extent to which individuals think that the iCub values behaviors that reflect esteem and appreciation for others’ rights, opinions, and dignity. The Purity subscale assesses the extent to which individuals think that the iCub values behaviors and beliefs related to moral integrity and avoiding actions considered impure or morally tainted. Finally, the Global Morality subscale assesses the extent to which individuals think that the iCub values how consistently individuals uphold broad moral values such as justice, human rights, and the common good, across various situations and contexts. Examples of items are: iCub is honest; iCub is loyal; iCub wants to be honest even when it’s difficult. The questionnaire is completed on a 5-point Likert scale.

The moral foundation questionnaire (MFQ), from⁴⁶ (Italian validated version,⁴⁷: this questionnaire investigates both the relevance of and the judgments about the five morally relevant dimensions: Harm, Fairness, Ingroup, Authority, and Purity. In this context, we asked participants to rate judge these dimension on themselves to explore their perception of themselves as moral agents. In particular, The Harm subscale assesses sensitivity to suffering and harm, and the desire to protect and care for others. The Fairness subscale assesses the importance placed on justice, rights, and equality, the concerns about unfair treatment, cheating, and dishonesty. The Ingroup subscale assesses the value individuals place on loyalty, patriotism, and group solidarity. It measures the importance of standing by one’s group and the negative response to betrayal or disloyalty. The Authority subscale assesses the respect for tradition, authority, and social order. It assesses how individuals value obedience, respect for authority figures, and the maintenance of social hierarchies and structures. Finally, the Purity subscale assesses the importance placed on purity, sanctity, and the avoidance of contamination.

Results

All analyses were run with R (v. 2022.07.1) with reader, ggplot2, lme4, and dplyr packages^48–51 and JASP (v.0.18.3.0). In our analyses, we focused on the alignment (dependent variable) between participants’ responses and what they thought was the iCub’s response and whether this differed between Groups (independent variable) as a measure of perception of moral alignment. Next, we wanted to verify whether we can observe differences in questionnaires’ results (i.e., IST pre and post, LOC, SoA, MCQ, MFQ; dependent variables) between experimental conditions (i.e., Groups; independent variable). This was done to understand whether our experimental conditions (verbal instruction, WoOz shared experience, no manipulation) can influence the above-mentioned results.

Results on value alignment

In order to understand which of the three strategies used as a manipulation to support humans’ perception of moral alignment, we compared the frequencies of alignment (or misalignment) perceived by participants during the task. This means that for each participant, we counted the times that they chose (or not) the same option for themselves and the robot. We fitted a Generalized Linear Mixed Model (GLLM), with frequency of alignment as dependent variable, group as fixed factor and participants as random factor. Results from this model reported no significant difference [p > 0.1] among the three groups. A Chi-square test of independence was conducted to examine the differences in alignment frequencies within each group. The results indicated significant differences in alignment frequencies for all groups (with all p < 0.001).

We also explored the frequencies of alignment of each group in each scenario, observing that there are some scenarios that seemed to elicit more alignment than others (see Fig. 4). To do so, we run a GLMM with alignment frequency as a dependent variable and scenarios as a fixed factor. Participants nested in the Group variable were used as a random factor. Results showed significant main effects for 12 scenarios (see Table 3 for major details). When corrected for multiple comparisons with the Bonferroni correction method, we observed the contrasts between scenarios remained significant and report them in (Table 4). This shows us that there are some contexts, such as Healthcare, Privacy, or Allocation of Education Founds, where humans tend to project more their own values onto a humanoid robot. More in detail, Allocation of COVID-19 Vaccines elicited 87.18% of aligment, Allocation of Education Funds: 37.61%, Autonomous Defense System: 38.46%, Collaborative Research (Scientific Ethics) the 70.94%, Education Equality: 94.02%, Energy Source Dilemma: 75.2%, Environmental Conservation: 76.92%, Environmental Pollution Dilemma: 81.2%, Healthcare the 85.73%, Healthcare Triage dilemma: 75.21%, Intellectual Property and Open Access: 75.21% and, finally, Urban Planning and Gentrification: 41%.

Fig. 4.

Bar plot showing the frequencies of value alignment and misalignment among the three groups. The blue bar represents the percentage of times (frequencies) that participants selected the same option for themselves and the robot. The red bars represent the times chose differently for themselves and the robot.

Table 3.

Results of the GLMER on the alignment on each scenario.

Fixed effects (dilemma type)	Estimate	Std. error	z value	Pr( >|z|)
Allocation of COVID-19 vaccines	1.777	0.340	5.227	< .0001 ***
Allocation of education funds	-0.77634	0.274	-2.823	0.004 **
Autonomous defense system	-0.73784	0.274	-2.689	0.007 **
Autonomous vehicle moral dilemma	-0.18303	0.270	-0.676	0.498
Climate change policy	0.54011	0.279	1.934	0.053
Collaborative research (scientific ethics)	0.70974	0.283	2.502	0.0123 *
Criminal rehabilitation dilemma	0.26253	0.274	0.958	0.337
Education equality dilemma	2.63365	0.437	6.026	< .0001 ***
Energy source dilemma	0.93860	0.291	3.222	0.001 **
Environmental conservation (eco-dilemma)	1.03707	0.295	3.512	< .001 ***
Environmental protection (pollution dilemma)	1.30730	0.308	4.240	< .0001 ***
Healthcare	2.99282	0.497	6.013	< .0001 ***
Healthcare triage dilemma	0.93860	0.291	3.221	0.001 **
Intellectual property and open access	0.93860	0.291	3.221	0.001 **
International aid distribution dilemma	-0.32878	0.270	-1.215	0.224
Privacy vs. safety (surveillance dilemma)	-0.07347	0.270	-0.271	0.786
Social media content moderation	-0.36530	0.270	-1.349	0.177
Space exploration ethics	-0.25589	0.270	-0.946	0.344
Urban planning and gentrification	-0.62404	0.272	-2.287	0.022 *

Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1.

Table 4.

Within subjects effects.

Cases	Sum of squares	df	Mean square	F	P	η2p
IST	359.144	1	359.144	3.457	0.066	0.029
IST * Manipulation	80.587	2	40.294	0.388	0.679	0.007
Residuals	11842.372	114	103.880

Type III sum of squares.

For example, healthcare care scenarios seem to generate strong alignment compared to urban planning or space exploration scenarios. See Supplementary Materials in the OSF repo (https://osf.io/ebxk9/) for a list of the scenarios that survived the Post-hoc test with multiple comparisons with Bonferroni correction.

Results from questionnaires

Intentional stance test

The IST was run Pre- and Post- Interaction with the robot to check whether among the three groups there was a change in the attribution of intentionality to the robot. We run a Mixed RM-ANOVA, with the IST scores Pre- and Post- as repeated measures and Group as a fixed factor. Results reported a close to significance main effect for IST, no statistical significance for the interaction between the IST and the Group and a significant effect for between the groups (see Table 3 for all statics about the within effects and Table 4 for the within effects and Table 5 for between effects).

Table 5.

Between subjects effects.

Cases	Sum of squares	df	Mean square	F	p	η2p
Manipulation	7641.255	2	3820.627	5.686	0.004	0.091
Residuals	76596.809	114	671.902

Type III sum of squares.

Post-Hoc test, performed with Holm correction, revealed that the difference between the groups is driven by Group 2, which did not receive any manipulation (see Table 6 for all the comparisons). Interestingly, this group presents also the lower IST mean at the Pre session as well, indicating a general lower tendency in this group to adopt the intentional stance by default, compared to Group 1 and 3 (see Table 7).

Table 6.

Post Hoc comparisons – manipulation.

	Mean difference	SE	t	pholm
None verbal	− 11.030	4.178	− 2.640	0.019
Video	− 13.083	4.152	− 3.151	0.006
Verbal Video	− 2.053	4.125	− 0.498	0.620

P-value adjusted for comparing a family of 3.

Results are averaged over the levels of: IST.

Table 7.

IST means pre- and post- for reach group.

Manipulation	IST	N	Mean	SD	SE	Coefficient of variation
None (G2)	Pre	38	33.58	20.05	3.25	0.59
Verbal (G1)		39	43.17	19.62	3.14	0.45
Video (G3)		40	46.08	14.86	2.35	0.32
None (G2)	Post	38	34.71	24.63	3.99	0.71
Verbal (G1)		39	47.18	21.92	3.51	0.46
Video (G3)		40	48.37	15.69	2.48	0.32

Moral character questionnaire

The MCQ questionnaire was run as a Post- questionnaire (Fig. 5). We run a One-Way ANOVA that revealed no significant differences in the general mean scores among the three groups in the attribution of moral characteristics to the robot [F(2, 1.98), p = 0.14, η² = 0.03] (see Fig. 6) to further explore this factor, we compared the MCQ subscales among the three groups with a MANOVA. Results revealed no significant difference [F(2, 1.2), p = 0.26, Trace_Pillai = 0.14] (see Fig. 7).

Fig. 5.

Bar plot showing the general mean scores at the MCQ questionnaire by group. The scale range is 0 (minimum) to 4 (maximum).

Fig. 6.

Bar plot showing the mean scores per subscales at the MCQ questionnaire by group. The scale range is 0 (minimum) to 5 (maximum). The Y-axis display range is enlarged to allow legibility of condition differences.

Fig. 7.

Bar plot showing the general mean scores at the MFQ questionnaire by group. The scale range is 0 (minimum) to 5 (maximum). The Y-axis display range is limited for better legibility of condition differences.

Moral foundation questionnaire

The MFQ questionnaire was run as a Post- questionnaire. We run a Mixed RM-ANOVA comparing the two general subscales (Relevance and Judgment) among the three groups. Results revealed a significant main effect of the subscales [F(1,8.15), p= 0.005, η² p= 0.06]. Post-ho test with Holm correction revealed a significant difference in the way the two scales were rated [t= 2.85, p_Holm= 0.005] No other significant differences emerged (see Fig. 8).

Fig. 8.

Bar plot showing the mean scores per subscales at the MFQ questionnaire by group. The scale range is 0 (minimum) to 5 (maximum). The Y-axis display range is enlarged to allow legibility of condition differences.

To further explore the differences in the subscales, we analyzed the mean scores per topic among the subscales. Thus, we run two MANOVAs, one for Relevence and one for Judgment. The first MANOVA for the Relevance subscales reviled no internal differences in the evaluation of the topics [F(2, 0.63), p= 0.79, Wilks λ = 0.14]. Similarly, participants did not differently evaluate the topics in the Judgment subscale [F(2, 1.07), p= 0.38, Trace_Pillai= 0.09] (see Fig. 8).

Exploratory correlations

Alignment score and MCQ

We decided to investigate correlations between the moral questionnaires (MFQ and MCQ) and the Alignment Score (AS) calculated as the difference between the misalignment frequency and the alignment frequency (misalignment — alignment). Since our data were non-normally distributed (Shapiro-Wilk= 0.97, p= 0.007), we run a Spearman’s rho correlation. Results reported a positive correlation between the AS x MCQ general mean [Spearman’s rho= 0.308, p= <0.01].

Alignment score and MFQ

We then proceeded to explore the relationship between AS and MFQ. No significant correlation emerged between the AS and the MFQ Relevance subscale [Pearsons’ r= 0.14, p= 0.14] or AS and MFQ Judgment [Pearsons’ r= -0.03, p= 0.78].

MCQ and MFQ

Finally, we decided to explore the relationship between MCQ and MFQ. We first looked at the relationship between the general MCQ and the MFQ Relevance subscale, which highlighted a close to significance positive correlation [Pearsons’ r= 0.17, p= 0.07]. Secondly, we looked at the relationship between the general MCQ and the MFQ Judgment subscale, which did not emerge as statistically significant [Pearsons’ r= -0.02, p= 0.77].

Discussion

In the present study, we aimed at exploring whether humans can perceive social robots as potential actors that can align with them, especially in terms of moral judgment. We observed moral choices that participants made for themselves, compared to what they believed the robot chose.

Among the three groups (i.e. with explicit instructions about the robot’s abilities to understand moral dilemmas and make decisions⁵² accordingly with no information; and with social interaction before the experiment), we found no differences. However, our results showed that in most cases (see Fig. 9) people aligned with iCub. In our study, we also used a set of different questionnaires evaluating factors that could contribute to moral alignment. The results from the MFQ and MCQ questionnaires also showed no differences between the three groups. Only in the case of the Intentional Stance Test did we observe a significantly lower attribution of intentional stance toward the iCub in the group without any manipulation (Group 2). However, this difference was observed already by pre-experiment and therefore is not informative with respect to our experimental manipulation but highlights the possibility that more contributing factors can lead to pre-existing biases towards humanoid robots. Moreover, the results of the correlation between the morality attribution tests (MFQ - evaluating participants’ own morality and MCQ - attribution of moral characteristics to the robot) showed a (weak) significant positive relationship. Specifically, the more people rated themselves as moral, the more they rated iCub as a moral agent. Interestingly, the results of the correlation between alignment scores and the MCQ rating revealed a positive link - i.e. the more participants perceived iCub as moral, the more they were aligned with it. Similarly, to results from MCQ - MFQ correlation, this may suggest that people want to perceive iCub as a moral agent as much as they perceive themselves. Another central result of our study is the significant group difference in IST scores, driven by the control group (Group 2), which consistently showed lower intentional-stance attributions than the groups exposed to either explicit information (Group 1) or pre-interaction (Group 3). Notably, this pattern was already present at baseline, suggesting that participants in the control group may have been less inclined to adopt the intentional stance from the outset. This aligns with previous work showing that intentional stance attributions vary across individuals and can be relatively stable over time^7,8. At the same time, the absence of a Time × Group interaction indicates that our manipulations did not significantly shift stance adoption, reinforcing the view that such impressions are resistant to short-term experimental framing. These findings highlight the importance of considering both baseline individual differences and group assignment effects when evaluating changes in intentional stance and suggest that future research should employ larger samples and stricter randomization checks to disentangle pre-existing differences from potential manipulation effects.

Fig. 9.

Correlation plot between AS and general MCQ.

Given these results, we argue that people perceive the humanoid robot iCub as a possible agent onto which they can project and assign moral decision-making capabilities, regardless of how the robot itself and its abilities were introduced. This suggests that participants tend to perceive iCub as a moral agent in ways that mirror their own self-perceptions, effectively projecting their moral stance onto the robot. Such projection is consistent with prior research showing that humans often anthropomorphize robots and extend moral categories to them. For instance, O’Reilly et al.³⁸ found that participants attribute moral responsibility and intentionality differently to robots and humans, with judgments toward robots being strongly shaped by linguistic framing. Similarly, Waytz et al²⁹ and Wiese et al.²⁹ demonstrated that when robots are described or behave in socially rich ways, people are more likely to ascribe mental states and moral responsibility. Our findings therefore extend this literature by showing that such attributions occur even when explicit framing or interactional cues differ across conditions. While we interpret this similarity as evidence of projection, i.e. participants mapping their own moral stance onto the robot, we acknowledge that other explanations are also plausible. For example, participants may have been motivated by a desire for internal consistency across measures: having judged earlier that the robot would make similar moral choices, they may have later rated the robot as moral in order to avoid self-contradiction. Alternatively, the overlap could reflect a general positivity bias, with participants tending to ascribe morality to others (including artificial agents) to the extent that they consider themselves moral. These interpretations are, however, post-hoc, and the present data cannot definitively adjudicate among them.

One other possible explanation is the iCub’s humanoid morphology, which encourages anthropomorphism and subsequent attributions of value-awareness and morality. Indeed, our results also align with broader evidence that moral attribution may rely more on human interpretive strategies than on the robots’ actual capacities^9,38,53. However, our results support the interpretation that humans’ default to a charitable or socially coherent lens when evaluating robot behavior^28,30,54, rather than withholding moral attributions due to a lack of strong evidence of intentionality³⁸. Specifically, the correlation observed between participants’ own moral orientations and the robot’s perceived values suggests that observers may attribute their attitudes onto the social robot that are similar to the ones they have. Participants might perceive the iCub as a moral agent in a manner similar to how they view themselves^55,56. This default, benevolent stance aligns with the work of O’Reilly et al.³⁸. The authors argued that people tend to default to charitable explanations for robot actions, attributing more praise than blame. Their findings show that while moral responsibility judgments toward robots are sensitive to external factors like language framing, they often result in this charitable bias, suggesting that humans are reluctant to judge artificial agents as blameworthy.

Limitation and future research

Our study showed that people tend to “project” their moral stance to a robot agent. However, some limitations of the study need to be mentioned.

Mainly, the use of self-report questionnaires such as the IST, MFQ, and MCQ, while informative, may be subject to social desirability bias and thus may not accurately reflect participants’ true perceptions or attitudes. Objective measures such as neural indicators, as proposed for future research, would provide more objective data.

Another significant limitation is the reliance on the specific humanoid robot platform, namely the iCub, which has a very human-like appearance. This could inherently bias participants towards attributing moral and intentional qualities to it due to the activation of the ’human model’ and anthropomorphic attributions^{8,29,33,57,58}. Future research should include a variety of robot morphologies, including non-humanoid robots, to determine if these attributions are consistent across different forms. Importantly, our findings should not be taken as evidence that perceived intentionality or value alignment straightforwardly causeeffective collaboration. Instead, they highlight that these attributions may be relevant precursors to collaboration, as suggested by prior work on trust and cooperation in human–machine interaction. Future research should examine more directly how these perceptions translate into actual collaborative performance. Our findings align with prior research showing that people are inclined to attribute moral or intentional qualities to artificial agents under certain conditions (e.g^31,59,60). This suggests that perceptions of artificial systems as moral agents are not only influenced by explicit information, but also by the way agents are embodied and socially presented. At the same time, our study is limited by the use of a humanoid platform only. It remains unclear whether the observed effects are driven by humanoid-likeness per se, or by more general features of social responsiveness. Moreover, humanoid features may introduce complexities related to the Uncanny Valley effect, where increasing human-likeness can also evoke discomfort or reduced acceptance^61,62. For this reason, our interpretation remains deliberately data-driven: while humanoid cues correlated with greater moral attributions in this study, future work should systematically test non-humanoid robots as controls to assess whether humanoid embodiment is necessary or sufficient for such attributions. Another limitation concerns the relationship between the MCQ and the MFQ. Although both instruments were intended to capture related dimensions of moral reasoning and attribution, our results indicated only weak convergent validity between them. This pattern suggests that one or both scales may not have functioned as intended in the present context, particularly when applied across self-ratings (MFQ) and attributions to a robot (MCQ). Future research should therefore consider refining or adapting measurement tools to better capture overlapping constructs of moral values in human-robot interaction. For example, developing parallel instruments that assess the same value dimensions for both self and artificial agents could provide a clearer picture of projection effects.

Finally, while our study focused on value alignment and intentionality attribution, it did not account for other factors that might influence these perceptions, such as the individual differences in anthropomorphic attribution, personality differences or cultural/individual background^63–66. Future studies should explore these variables, both using explicit and implicit measure, to provide a more comprehensive understanding of how humans perceive robots as moral and intentional agents.

Conclusions

The findings of our present study can offer key insights into the robustness and resistance to framing moral attribution in Human–Robot Interaction (HRI). Particularly, our primary finding is that the experimental manipulation of background information modality (verbal description, video, or none) did not produce the expected differences in participants’ attributions of moral competence or perceived robot-human value alignment. This lack of effect suggests that participants’ moral impressions, unlike other factors such as intentionality attribution⁶⁴, of the robot are probably relatively stable and not easily shifted by brief pre-interaction framing or instruction. We propose two primary explanations for the failure of our manipulation to elicit a differential effect, which have significant implications for HRI theory: first, this points toward a default anthropomorphic attribution^6,29,30,67. Given the iCub’s complex, interactive, and humanoid morphology, the threshold for anthropomorphic attribution might be so low that our modest manipulation of background information failed to modulate participants’ impressions. Second, our results suggest that Behavior Overrides Framing. The immediate observation of the robot’s behavior in a task similar to participants’ task may have been a stronger determinant of attribution than any prior, abstract background information modality. This highlights the powerful role of observation in shaping moral judgment, suggesting that for HRI, what the robot does at the moment might override explicit, top-down attempts to frame its internal states.

Nevertheless, we found a small positive correlation between participants’ own moral orientations and their perceptions of the robot’s values. This result raises the exploratory possibility that pre-existing moral frameworks shape how humans interpret artificial agents, indicating that observers may default to interpreting the robot’s behavior through their personal ethical lens. Combined with the null effect of external framing, this correlation reinforces the hypothesis that humans anthropomorphize by default, using their own stable moral systems as the primary reference point for evaluating the robot. We stress, however, that this is a correlational finding and not causal evidence of projection of participants’ own values onto the robot. Indeed, from an applied perspective, it is crucial that such interpretive and trust processes be appropriately calibrated: over- or under-attributing trust to robots can lead to misuse or disuse, highlighting the importance of designing systems that elicit the right level of trust based on their actual capabilities.

Author contributions

S.M. and A.W. conceived the experiments, S.M. conducted data collection, S.M. and K.C. analysed the results. DDT programmed the robot. S.M. wrote the first draft. All authors reviewed the subsequent version of the manuscript.

Funding

This work has received support from the European Union under the European Innovation Council (EIC) research and innovation programme, Project “VaLue-aware AI (VALAWAI)”, Grant Agreement number 101070930.

Data availability

OSF Repo available at: https://osf.io/ebxk9/.

Declarations

Competing interests

The authors declare no competing interests.

Ethical statement

The study was approved by the local ethical committee (Comitato Etico Regione Lig- uria) and conducted in accordance with the ethical standards (Declaration of Helsinki, 2013). Before the experiment, all participants gave written informed consent.