Human-AI-Robot Teaming Technical Group

I have done some thinking on this question.  I say teammate as defined in the literature - member of a heterogeneous and interdependent team in which members work toward a common goal and have different roles and responsibilities.                                                                                                                                        I can add the following from a recent article published in Human Factors  "As indicated in Question 1 of Table 1, intelligent machines can be considered teammates just as animals can be considered teammates (Phillips et al., 2016). In many ways, we may do better to consider machine teammates as members of another species. Considering a machine as a teammate does not mean that the machines are in control, that machines are human or human-like, or that the machine is not human-centered. In fact, designing a machine to work well with humans as a teammate can increase human-centeredness. In addition, this design can draw on what we know from the team literature (e.g., team composition, team process, team development, and team measurement) to do so."                                                                                                                                Cooke, N. J., Cohen, M. C., Fazio, W. C., Inderberg, L. H., Johnson, C. J., Lematta, G. J., Peel, M., & Teo, A. (2023). From Teams to Teamness: Future Directions in the Science of Team Cognition. Human Factors, 0(0). https://doi.org/10.1177/00187208231162449                                                                                                                                    

I have done some thinking on this question.  I say teammate as defined in the literature - member of a heterogeneous and interdependent team in which members work toward a common goal and have different roles and responsibilities.                                                                                                                                        I can add the following from a recent article published in Human Factors  "As indicated in Question 1 of Table 1, intelligent machines can be considered teammates just as animals can be considered teammates (Phillips et al., 2016). In many ways, we may do better to consider machine teammates as members of another species. Considering a machine as a teammate does not mean that the machines are in control, that machines are human or human-like, or that the machine is not human-centered. In fact, designing a machine to work well with humans as a teammate can increase human-centeredness. In addition, this design can draw on what we know from the team literature (e.g., team composition, team process, team development, and team measurement) to do so."                                                                                                                                Cooke, N. J., Cohen, M. C., Fazio, W. C., Inderberg, L. H., Johnson, C. J., Lematta, G. J., Peel, M., & Teo, A. (2023). From Teams to Teamness: Future Directions in the Science of Team Cognition. Human Factors, 0(0). https://doi.org/10.1177/00187208231162449                                                                                                                                    

Hi there!

To add to your initial interest in "tool vs. teammate" discussion, I wanted to suggest that you look into literature on mind perception/humanlikeness/anthropomorphism (terms used somewhat interchangeably), and social robotics. Nancy and Mike have brought up the concept of humanlikeness in terms of the design/capabilities of the agent, but I wanted to add that the extent to which an agent is perceived as humanlike is highly subjective and has a lot of consequences for human-AI interactions.

Humans have a natural tendency to endow objects with humanlike "minds of their own", capable of having their own beliefs, desires, intentions, etc., and this is (somewhat) independent of the actual capabilities. Merely giving a self-driving car a human name and voice can increase trust and people's perception of the underlying competence (check out Waytz, Heafner & Eply, 2014). Anthropomorphism/ mind perception is complex (see Epley, Waytz, Cacioppo, 2007) and the extent to which it is applied varies a lot and deeply affects social perceptions (even on the level of processing sensory information) and treatment of agents (e.g., humans are punished for "unfairness", but the same actions from computers are not seen as unfair; see Sanfey et al., 2003). Some folks in the social robotics field believe that designing agents which can activate the same social brain regions used in human-human interactions is the key to effective HRI / teamwork with agents (see Wiese, Metta, Wykowska, 2017). To date, there are no AI-agents capable of maintaining a high degree of perceived humanlikeness over longitudinal interactions in the real-world without deception (disagreements welcomed!). Everyone loves to reference LLMs like ChatGPT, but it doesn't take long before it starts to "hallucinate" or for people to jail-break it. As soon as the agent does something "mindless", the illusion is shattered and it's just another dumb bot.

Something I (and others) worry about in this context is the potential to harm human-human interactions once an AI-agent is introduced in the real-world. "Humanization" mechanisms share the same spectrum as "dehumanization" of other humans (see Waytz, Gray, Epley, Wegner, 2010). It is possible that adding anthropomorphic agents (even during one-off interactions) will create unintended opportunities to act unfairly or rudely towards other human teammates.

Hope that adds something meaningful. Happy to chat more if desired 😊

-Stephanie

I have done some thinking on this question.  I say teammate as defined in the literature - member of a heterogeneous and interdependent team in which members work toward a common goal and have different roles and responsibilities.                                                                                                                                        I can add the following from a recent article published in Human Factors  "As indicated in Question 1 of Table 1, intelligent machines can be considered teammates just as animals can be considered teammates (Phillips et al., 2016). In many ways, we may do better to consider machine teammates as members of another species. Considering a machine as a teammate does not mean that the machines are in control, that machines are human or human-like, or that the machine is not human-centered. In fact, designing a machine to work well with humans as a teammate can increase human-centeredness. In addition, this design can draw on what we know from the team literature (e.g., team composition, team process, team development, and team measurement) to do so."                                                                                                                                Cooke, N. J., Cohen, M. C., Fazio, W. C., Inderberg, L. H., Johnson, C. J., Lematta, G. J., Peel, M., & Teo, A. (2023). From Teams to Teamness: Future Directions in the Science of Team Cognition. Human Factors, 0(0). https://doi.org/10.1177/00187208231162449                                                                                                                                    

Nancy makes a fine point about why machines CAN be teammates, and I completely agree.  But machines can ALSO be tools.  As I learned in my Intro to Physics class, a lever is a "machine"-- but I don't tend to think of it as a teammate, even though it is independent (of me) and we can work together in relationships with a common goal and different roles and responsibilities.

This probably hinges on WHEN it is useful to think of machines in one way vs. another and that, in turn, rests on complexity (both of mechanism and of behavior), independence of action and, thus, autonomy.  A hammer (or a lever) doesn't have a lot of behavioral complexity or autonomy-- so when I hit my thumb with it, I blame myself ("I hit...") but if a computer or a robot does something adverse for our team interactions, I'm more likely to blame it.  

I am convinced this is all related to the phenomenon of anthropomorphism-- people's propensity, in certain circumstances, to regard complex, autonomously-acting agents/systems/things (e.g., weather, "the gods") using the same set of expectations and interpretations that they have learned for human-human interactions.  This is behind Dennett's Intentional Stance (1989) and Nass's "Computers As Social Actors" paradigm (Reeve's and Nass, 1996) and my own work on human-machine "etiquette" (Hayes and Miller, 2011, Miller, 2004).  

But that ends up being a final, strong reason (IMHO) for HF researchers to at least permit the idea of treating machines as teammates-- because human users do!  Pretty regularly.  And user's interactions are functionally similar (even identical) to how they behave with human teammates.  It's worth looking at the Nass work (and later work in that tradition) for factors and conditions that do and don't encourage anthropomorphic responses-- and maybe designing accordingly.  

In the end, I suspect this is a spectrum of perceptions and mental models that maps onto a range of behavioral interactions.  The agents themselves (humans vs. machines vs. animals) may be different and provide different affordances (e.g., greater range of behaviors and understanding for humans, but also intentionality that can be explicitly at odds and opposed to that of the user), but the interactions are similar across a wide range of what we might call "teaming" behaviors.  

--Chris

Nancy makes a fine point about why machines CAN be teammates, and I completely agree.  But machines can ALSO be tools.  As I learned in my Intro to Physics class, a lever is a "machine"-- but I don't tend to think of it as a teammate, even though it is independent (of me) and we can work together in relationships with a common goal and different roles and responsibilities.

This probably hinges on WHEN it is useful to think of machines in one way vs. another and that, in turn, rests on complexity (both of mechanism and of behavior), independence of action and, thus, autonomy.  A hammer (or a lever) doesn't have a lot of behavioral complexity or autonomy-- so when I hit my thumb with it, I blame myself ("I hit...") but if a computer or a robot does something adverse for our team interactions, I'm more likely to blame it.  

I am convinced this is all related to the phenomenon of anthropomorphism-- people's propensity, in certain circumstances, to regard complex, autonomously-acting agents/systems/things (e.g., weather, "the gods") using the same set of expectations and interpretations that they have learned for human-human interactions.  This is behind Dennett's Intentional Stance (1989) and Nass's "Computers As Social Actors" paradigm (Reeve's and Nass, 1996) and my own work on human-machine "etiquette" (Hayes and Miller, 2011, Miller, 2004).  

But that ends up being a final, strong reason (IMHO) for HF researchers to at least permit the idea of treating machines as teammates-- because human users do!  Pretty regularly.  And user's interactions are functionally similar (even identical) to how they behave with human teammates.  It's worth looking at the Nass work (and later work in that tradition) for factors and conditions that do and don't encourage anthropomorphic responses-- and maybe designing accordingly.  

In the end, I suspect this is a spectrum of perceptions and mental models that maps onto a range of behavioral interactions.  The agents themselves (humans vs. machines vs. animals) may be different and provide different affordances (e.g., greater range of behaviors and understanding for humans, but also intentionality that can be explicitly at odds and opposed to that of the user), but the interactions are similar across a wide range of what we might call "teaming" behaviors.  

--Chris

Hi James,

Those that advise against conceptualizing technologies as teammates seem to be most focused on reminding us that machines are not humans, and should not be conceptualized as HUMAN teammates. You also make a number of good points about what machines are (at least currently) unable to do. Some make a distinction that these machines can be team players, but not teammates, and some eschew the team moniker entirely.

It is also dangerous to conceptualize these as tools, as this designation masks the effect that these technologies have (again, at the very least) on the cognitive processes of their human counterparts. This is true of what you termed automation AND autonomous technologies. Once a machine processes input data and presents an output (sometimes referred to as making an inference), it is in a different league than technologies that do not, and design considerations should change accordingly. 

I have found that the most reliable perspective in directing my design is that of joint human-machine activity. Designing for joint activity focuses on how all of the machine and human agents come together to jointly perform the activities necessary to perform a task, or a mission. It also implies that there are numerous joint activity architectures that can be enabled in the design directions that are chosen. I have found that most activities map onto the standard set of described macrocognitive functions, or can be decomposed into them (e.g., detecting events, sensemaking, responding, coordination). When thinking about how machines contribute to joint activity, we can more readily bring everything we know about how these machines influence joint cognition, including how they can induce framing effects, how they enable or hobble detection, or how they induce increased coordination or stymie it. 

To answer what capabilities a technology that supports joint activity must have, it somewhat depends on the activity. At the most basic, though, I'll borrow from one of the more read "team player" papers: they must observe the Basic Compact, facilitate (or at least not obscure) Shared Common Ground, and must be observable, predictable, and directable. We are developing 20 more detailed Joint Activity Design (JAD) heuristics right now, which I would be happy to share when we're a little further along.

Is that helpful?
Mike

Hi Jim, 

It's been fun to listen in on the perspectives from the HART TG on this very interesting topic. I wanted to jump in and offer two things: 1) Given the amount of interest on this topic I would recommend that this could be a highly attended discussion panel event or symposium at the next HFES conference (as the HART TG Program Chair there's a shameless plug). 2) My colleague Kevin Wynne and I wrote a brief conceptual paper on this topic in 2018, I've shared the citation below. I believe machines "can be" teammates particularly when you endorse the idea that Nancy has spoken about that they need not be the same kind of teammate as your human counterparts. I also believe that many of our machines are tools but may, at times, be "teammate-like" (pulling on a thread that Guliz mentioned earlier). Very interesting discussion thread, thank you for initiating it. Reference below:

Wynne, K.T., & Lyons, J.B. (2018). An Integrative Model of Autonomous Agent Teammate-Likeness. Theoretical Issues in Ergonomic Science, 19, 353-374.  

Hi Jim, 

It's been fun to listen in on the perspectives from the HART TG on this very interesting topic. I wanted to jump in and offer two things: 1) Given the amount of interest on this topic I would recommend that this could be a highly attended discussion panel event or symposium at the next HFES conference (as the HART TG Program Chair there's a shameless plug). 2) My colleague Kevin Wynne and I wrote a brief conceptual paper on this topic in 2018, I've shared the citation below. I believe machines "can be" teammates particularly when you endorse the idea that Nancy has spoken about that they need not be the same kind of teammate as your human counterparts. I also believe that many of our machines are tools but may, at times, be "teammate-like" (pulling on a thread that Guliz mentioned earlier). Very interesting discussion thread, thank you for initiating it. Reference below:

Wynne, K.T., & Lyons, J.B. (2018). An Integrative Model of Autonomous Agent Teammate-Likeness. Theoretical Issues in Ergonomic Science, 19, 353-374.  

Hi Jim,

We debated this issue on the recent NAS study on human-AI teaming.  We decided that the framework of a "team" was appropriate for several reasons. First, as Nancy said, because it meets the definition of a team. Secondly, because the concept of team was important for providing improved adaptability and capacity for mutual support and backup that is important. And thirdly because the need for coordination with people increases as the capability of technology increases. More details can be found in that report.  

National Academies of Sciences Engineering and Medicine. (2021). Human-AI teaming: State-of-the-art and research needs. Washington, DC: National Academies Press.

In addition, you may be interested in a discussion I had with Ben Shneiderman on this topic (attached) that explores the ins and outs of considering them as a team. As you know, he is very much of the "AI cannot be a teammate, only a tool" opinion. And in fairness, there is very little in the way of AI that acts like a teammate at present. He published this discussion in his Human-Centered AI email posting that goes out once a week to the community. 

best wishes,

Mica

Hi Jim, 

It's been fun to listen in on the perspectives from the HART TG on this very interesting topic. I wanted to jump in and offer two things: 1) Given the amount of interest on this topic I would recommend that this could be a highly attended discussion panel event or symposium at the next HFES conference (as the HART TG Program Chair there's a shameless plug). 2) My colleague Kevin Wynne and I wrote a brief conceptual paper on this topic in 2018, I've shared the citation below. I believe machines "can be" teammates particularly when you endorse the idea that Nancy has spoken about that they need not be the same kind of teammate as your human counterparts. I also believe that many of our machines are tools but may, at times, be "teammate-like" (pulling on a thread that Guliz mentioned earlier). Very interesting discussion thread, thank you for initiating it. Reference below:

Wynne, K.T., & Lyons, J.B. (2018). An Integrative Model of Autonomous Agent Teammate-Likeness. Theoretical Issues in Ergonomic Science, 19, 353-374.  

Hi! Jim,

This is a great topic. I recently submitted a small article to the journal "ACM Interactions" titled " Applying human-centered AI in developing effective human-AI teaming:  A perspective of human-AI joint cognitive systems". Please note that the audience of this journal is CS/HCI professionals. With a limit of 2800 words, I cannot go too deep, it is not an academic paper for the HFES community. The key points I tried to make:

•    AI has led the emergence of a new form of human-machine relationship: human-AI teaming (HAT), a paradigmatic shift in human-AI systems   
•    We must follow a human-centered AI (HCAI) approach when applying HAT as a new design paradigm 
•    We propose a conceptual framework of human-AI joint cognitive systems (HAIJCS) to represent and implement HAT for developing effective human-AI teaming  

The pre-print can be accessed via https://www.researchgate.net/publication/372246925_Applying_human-centered_AI_in_developing_effective_human-AI_teaming_A_perspective_of_human-AI_joint_cognitive_systems

Because I am heavily involved in the HCI area, so I regularly wrote papers there. I had talked to Shneiderman a couple of times. As context, I wrote one of the first papers for Human-Centered AI (HCAI) in the journal back in 2019 (widely downloaded/cited), and Prof. Shneiderman and I had some collaborative work to promote HCAI. As Mica mentioned, he is strongly against the teaming metaphor (instead, he prefers using super tools...), concerning that the teaming metaphor is against HCAI. In my small article, I intend to argue that AI as a teammate does not mean AI can be a leader to control humans (there are diff roles on a team), and let's apply HCAI for developing human-AI team systems. It is an attempt, but not sure if that works -:)  I reference the NAS report, which Mica mentioned and she led the effort, to the CS/HCI community. 

I feel the CS/HCI community needs to hear the voice of the human factors community, which is my motivation to write papers there. 

Also, it is good to know that the National AI Research and Strategic Plan 2023 Update specifically calls out research needs for human-AI teaming in a section. The report can be found on the Internet.            

Thanks

Wei

Hi Jim, 

It's been fun to listen in on the perspectives from the HART TG on this very interesting topic. I wanted to jump in and offer two things: 1) Given the amount of interest on this topic I would recommend that this could be a highly attended discussion panel event or symposium at the next HFES conference (as the HART TG Program Chair there's a shameless plug). 2) My colleague Kevin Wynne and I wrote a brief conceptual paper on this topic in 2018, I've shared the citation below. I believe machines "can be" teammates particularly when you endorse the idea that Nancy has spoken about that they need not be the same kind of teammate as your human counterparts. I also believe that many of our machines are tools but may, at times, be "teammate-like" (pulling on a thread that Guliz mentioned earlier). Very interesting discussion thread, thank you for initiating it. Reference below:

Wynne, K.T., & Lyons, J.B. (2018). An Integrative Model of Autonomous Agent Teammate-Likeness. Theoretical Issues in Ergonomic Science, 19, 353-374.  

Hi Jim, 

It's been fun to listen in on the perspectives from the HART TG on this very interesting topic. I wanted to jump in and offer two things: 1) Given the amount of interest on this topic I would recommend that this could be a highly attended discussion panel event or symposium at the next HFES conference (as the HART TG Program Chair there's a shameless plug). 2) My colleague Kevin Wynne and I wrote a brief conceptual paper on this topic in 2018, I've shared the citation below. I believe machines "can be" teammates particularly when you endorse the idea that Nancy has spoken about that they need not be the same kind of teammate as your human counterparts. I also believe that many of our machines are tools but may, at times, be "teammate-like" (pulling on a thread that Guliz mentioned earlier). Very interesting discussion thread, thank you for initiating it. Reference below:

Wynne, K.T., & Lyons, J.B. (2018). An Integrative Model of Autonomous Agent Teammate-Likeness. Theoretical Issues in Ergonomic Science, 19, 353-374.