Here Comes the Bad News: Doctor Robot Taking Over

In their Handbook of Research on Computer Mediated Communication, Kelsey and St. Amant [19] state that to regard robots as social, they need to express emotions, show personality, work with natural cues, and be capable of conducting high-level dialogue (p. 867). Their handbook also states that most robots are still underdeveloped in this respect (p. 865). Part of this omission is that we do not understand too well how, for instance, high-level dialogues take place between humans. To advance a little bit into this direction, the current study reports on the delivery of bad health news by a social robot, using framing techniques and language biases that are recommended to human messengers.

Delivering bad news through robots is the test case we used for the deeper aim of this study: A demonstration of testing the reach of Media Equation [35] and Computers Are Social Actors (CASA [32]) fairly by using Bayes as well as frequentist statistics. From the results, our contribution to the field emerged: A moderation of the said theories that people to some extent apply human social criteria to machines but that robots sometimes perform better according to those criteria than humans actually do and that some criteria do not count for robots. This is unexpected for many theories in media and technology, which simply assume that humans are the standard of performance. In application, this study opens up the serious possibility to apply robots in executing emotionally sensitive tasks, both relieving the messenger and the receiver from relational stress.

To progress towards such study, several hurdles have to be taken: in theory, in methods, and practically. Theoretically, we conceive of a social robot as a humanoid software and hardware system that at least in part can make autonomous decisions and expresses social behaviors by electro-mechanical means. In our case, we used the Robokind R50 “Alice,” designed by David Hanson and equipped it with our own software (Fig. 1). As a starting point, we relied on theorists such as Reeves and Nass [35], who formulated the Media Equation, suggesting that humans hardly distinguish between other humans and machines in the way they behave toward them. Reeves and Nass report that, for example, applications that stimulate extroversion are more appreciated by extroverts, whereas introvert people would prefer applications that demand more introversion. With regard to robots, certain studies indeed find that similar personalities attract (e.g., [42]). However, others find that complementarity attracts (e.g., [23]) but this may still be in line with Media Equation if humans respond similarly to humans who complement them. In the same vein did Nass and Moon [32] assert that Computers Are Social Actors, the CASA view. While interacting with computers, users would automatically or ‘mindlessly’ apply human-inspired heuristics and rules because certain anthropomorphic aspects of the machine remind them of human conduct—even if they are aware of the machine (cf. the emotional expressive systems of [10]). And indeed, Lee et al. [26], for instance, studied a robot moving around an office space while it dispensed snacks to the employees. Over time, the robot became more of a co-worker than a vending machine in the eyes of the employees, inspiring all kinds of social dynamics such as becoming more polite to the robot, protecting it, making social comparisons, and becoming jealous of people that supposedly were preferred by the machine.

There are also authors who believe that humans and robots are not treated equally (e.g., [11, 13]). For example, Takayama and Go [41] observed that certain people categorized mobile telepresence machines as indeed another person, arguing with it and making hand gestures (cf. CASA). Yet, other people saw it as a disabled person because it could not open the door or failed otherwise. A third group saw the telepresence machine as robot, a substitute of the actual person, and the fourth group merely saw an object, referring to it as ‘it’ and ‘thing,’ sitting on it like a chair or leaning against it like a railing (ibid.).

Such studies underscore that not all perceivers transfer the human model to robots without more (some see the machine-side rather than the human likeness) or at least incompletely (robot is a substitute for the real thing) or qualified by different scores (robot is clumsier). For instance, Küster and Świderska [21] reported evidence that humans do attribute a certain amount of morality to humanoid robots but also that systematic differences with humans occurred. For instance, humans were regarded as more experienced and more conscious than robots (p. 341). According to this kind of dissimilarity hypothesis, robots better should be seen as novel social entities with non-human but highly appreciated qualities of their own [14].

Where do we stand then? Do we treat social robots as if they were real people (similarity hypothesis)? There may be some margin assumed but not so large that the difference between treating robots and treating human beings becomes statistically significant. If similar, then all of the rules and subtleties of human communication apply and people supposedly are sensitive to whether the robot complies with those rules or not. Or are our human responses modified by the fact that it is a robot (dissimilarity hypothesis)? In that case, human behavioral rules would roughly apply but not systematically and unequal from responding to a human partner.

So far so good for theory because the comparison between human–human and human–robot interactions has its challenges in methods and analysis as well. From a methods point of view, ideally one would like to compare a physically present human (the ‘confederate’) with a physically present robot, interacting with a test participant. However, human confederates are hardly capable of delivering the same performance time and again across a sample of participants; let alone if that person has to act in two or more different ways (e.g., acting happy vs. sad). Woods et al. [48, 49] recommend the use of video clips in cases where interaction is low (such as unidirectionally delivering a message). Therefore, human–human studies often rely on videotaped materials that can be repeated multiple times without changing the contents. That may sound as yielding to methods at the cost of ecological validity but if we see the movie clips as a proxy of a videoconference (cf. telemedicine), we may not deviate too far from reality, particularly in view of future developments.

Given that humans should be videotaped to deliver the same performance, to keep things comparable, robots also should be videotaped. That of course, does something for the aspect of interaction (as with humans) and for autonomous decision making (as with humans). Nonetheless, we should start somewhere and explore the ‘simpler’ situations first before entering the methodological complications of real-life interactions. And admittedly, using movie clips is way more practical and cost-efficient than hiring professional actors and far less work than handling a robot, which commonly shows all kinds of technical flaws that need to be repaired on the fly [48, 49]. Fortunately, there are situations where little interaction and decision making is required, namely in one-way monologues where someone tells the other about a state-of-affairs, such as notifying employees of their layoff or breaking bad news to a patient, which is the topic of the current study.

Rejecting the null actually means that in the data significant differences were found between two groups, confirming the H1. In other words, frequentist analyses favor in the case of Media Equation and CASA, the dissimilarity hypothesis (H1: robots and humans are perceived as significantly different). Because if H1 is rejected and the H0 remains, it merely says ‘we found no difference’ rather than ‘we found similarity.’ Hence, to be fair to the similarity hypothesis as well, analyses should also be performed with Bayesian statistics. With Bayes, all unknown parameters can be defined by a probability distribution. Thus, Bayesian statistics do not result in a point estimate, but rather in an interval with a certain probability that the true coefficient is part of that interval [12, 44]. Bayesian analysis exists of three parts: the prior distribution, the data (the likelihood), and the posterior distribution. The posterior distribution is a combination of the prior distribution and the data, an updated understanding of the theory under question (ibid.). In other words, with Bayes the H0 can be tested and if H1 (dissimilarity) is refuted and H0 remains, we can assume that robots are treated similarly to humans as Media Equation and CASA would have it.

Different from Kelsey and St. Amant [19, p. 867], to regard robots as social, they do not need to express emotions too much, do not have to show a specific personality, nor work with natural cues, and do not have to be capable of conducting high-level dialogue. We repeated Experiment 2 of [6] with positive-negative framing and affirmative-negative language, confronting 134 participants with a robot doctor that brought bad news about Bekhterev’s rheumatic disease. H0 expected effects to be quite similar to the assorted effects of a human doctor. H1 predicted mediocre scores for robots with human doctors obtaining the highest and lowest scores dependent on the way they communicated the health message.

At three out of four dimensions (Doctor Evaluation, Message Evaluation, and Medical Adherence), the robot beat the human doctor (main effects). At Expected Quality of Life, the two tied. To increase Expected Quality of Life, the robot had to use negations, whereas the human had to use negative frames. The human also had to use negative frames to enhance Message Evaluation. Looking into the stimulus materials, Expected Quality of Life was said to (not) progress or (not) deteriorate. If a robot is to use negations (not progress and not deteriorate), the robot actually is preferred when things remain stable; when nothing changes compared to how it was. If for a human the emphasis is on the negative outcome but not so much on the language (not progress and deteriorate), then a human doctor seems the preferred option; when things get serious and worsen. Note, however, that there were some issues with the measurement of Expected Quality of Life (Online Resource 3), so that results should be taken with caution.

Putting it differently, robot and human did equally well at times, supporting H0; or the robot surpassed the human, even in cases where H1 did not expect it. Ergo, the human doctor never outperformed Doctor Robot except (if we trust the measurement) when someone’s quality of life was expected to worsen.

There may be some nuance to these conclusions, dependent on the type of analysis one adheres to. If the reader follows the frequentist analyses, proper framing seemed more of an issue for humans (Message Evaluation, Expected Quality of Life) than for robots. If the reader follows Bayesian analysis, robots had similar restrictions on how to convey the news. At one point (Expected Quality of Life), language was important for robots (i.e. negation use) but not for humans. In other words, robots could frame it and say it any way they wanted with sometimes an exception. Nevertheless, the robot outperformed the human at Doctor Evaluation (robot was nicer) and Message Evaluation (robot delivered a better message), and performed about equally well as a human on Medical Adherence (if you do not believe Bayes, then the robot performed even better). The main effects on Expected Quality of Life were insignificant so that may count as a tie.

We also measured how the robot was experienced in terms of Ethics, Affordances, Involvement, Distance, and Use Intentions [45]. Doctor Robot was not seen as distant but rather involving, morally good, skillful, and evoking the willingness to see her again.

Yet, it is strange that we obtained these results although the robot lacked interactive capacity and just delivered a pre-recorded message. Imagine what the robot would be capable of if it adapted its tone of voice and facial expression to the gravity of the diagnostic, show empathy, or adaptive behaviors [8]? Would that improve the appreciation of the robot even more? The question that arises is why the robot doctor performed so well at her communication tasks in spite of its lack of interaction and why, moreover, it was regarded as ‘ethical’ (although she made no moral statements), not cold and distant but rather warm (although no emotion simulator was running), and involving (although it spoke with monotonous voice)? Doctor Robot was considered ‘skilled’ (but merely recited a predefined message) and she invited future use although the electronic girlish machine hardly resembled a realistic physician.

In view of the Media Equation [35] and Computers Are Social Actors, the CASA paradigm [32], our results may be interpreted that indeed people listen to a robot as if it were human. However, only occasionally do they take the same factors into consideration and evaluate the performance as high. The as if is important here, because people are not mistaken by who the source of communication is (i.e. a non-human) and hence apply a different yardstick [9] as they do with pet animals for instance. In other words, human-like behaviors evoke human-like, not human-equal, evaluations.

Why was Doctor Robot a better messenger of bad news than the human doctor? Doctors prefer to sweeten the bitter pill by using negations such as “The news is not so good,” whereas patients wish to know what doctor actually thinks [7]. Could it be that the robot is not suspected of doubt and reservation? That the robot is not expected to want to regulate its own emotions nor those of its patient?

Perhaps that, reversely, the patient does not want to regulate the emotions of the doctor. After all, delivering bad news is as demanding for the messenger as it is for the patient. If the patient has to deal with more emotions than his or her own, less cognitive capacity will be available to process the information of the message. Perhaps that a doctor’s well-meant empathy and his sorry face are too much. Maybe professional distance should be restored so that the patient can share her emotions with the ones she loves, not with the one who brings her the bad news. In facing a robot, you do not have to feel embarrassed if you start crying nor will it have second thoughts about you.

Robots are not humans and perhaps people assume that robots have no desires of their own. Therefore, people probably believe that robots have no hidden agendas. They are not judgmental, have no critique. They merely deliver the message, without further ado. And without these negative features, they may gain more credibility. The absence of negative qualities may heighten the acceptance of the message (cf. [33]).

For example, only twice were there effects of framing for the robot (according to Bayes) and at one time did language play a role. It seemed that how a message was framed was important when the source of the message was a human, less so when it was a machine. In that case, predictions based on Prospect Theory [17] are limited by who the sender of the message is (i.e. humans).

Buhlmann and Gisler [5] observed that messengers receive more credibility when they are direct, brief, and result-oriented, things a robot does naturally. The absence of goals of its own may perhaps count as an indicator of ethical behavior of the robot (e.g., honest, sincere, trustworthy). Based on this assumed integrity, receivers would tend to believe the message. And even if they did not remember the message content too well, they probably did recall how the messenger made them feel (cf. [5]); in the case of the robot doctor, as an ethical entity, which reduced the receiver’s affective distance (Fig. 2).

Would it be ethical, then, to employ a robot for bad news conversations? For now, this question comes too early because we do not know the ins and outs of what the robot is allowed to do morally. Yet, certain people may regard the robot as more ethical than humans because it has no hidden agenda. For others, it would not be a sign of ‘good care’ to leave sensitive matters to machines. However, what is always good is that people have the opportunity to choose between a human and a machine as the bearer of bad news. And for the human messenger, that might be a relief as well.

It seems, then, that direct, almost impolite, communication works best for robots. That is to say, Torrey et al. [43] found that when robot as well as human cooking assistants communicated with or without ‘hedges’ (words that mitigate the impact of an utterance), respondents found them more friendly, empathic (i.e. considerate), and less controlling. Robots even outdid human helpers on these dimensions when they modulated their communication through hedges. On the one hand, this supports our finding that robots can be better communicators than humans. On the other hand, the robot in the cooking study achieved this result by using subtle communications whereas in our study, it did so in the absence of communicative modifications. Perhaps that the different tasks explain the different communications required: If you help a novice to achieve a modest goal (e.g., to bake muffins), it is good to be gentle but if bad news about serious matters (i.e. illness) has to be delivered, perhaps that is seen as beating around the bush and people prefer to get to the point straight away.

In all, the Media Equation thesis holds that people treat computers as social actors. Similarly do people treat robots as social actors; not like humans but as species of their own. People are not as simple as Reeves and Nass [35] thought. That the robot shows no signs of a personal ego with its selfish goals and concerns may make it an ethical social entity. Perhaps that in addition, reduced emotional expressions puts the locus of attention on the information (here, health) and less so on the inner struggle of the sender. In view of our results, telemedicine with a human may not always be the preferred way of communication: A humanoid robot on a screen may work just as well, sometimes even better. Media do not equal real life but they come close and in the case of social robots their near-humanness is what makes them stand out against ‘conventional’ humans with all their personal issues.