Davide Dell’Anna on hybrid intelligence, guidelines for human-AI teams, calibrating trust, and team ethics (AC Ep33)

Transcript

Ross Dawson: Hi Davide. It’s wonderful to have you on the show.

Davide Dell’Anna: Hi Ross, nice to meet you. Thank you so much for having me.

Ross: So you do a lot of work around what you call hybrid intelligence, and I think that’s pretty well aligned with a lot of the topics we have on the podcast. But I’d love to hear your definition and framing—what is hybrid intelligence?

Davide: Well, thank you so much for the question. Hybrid intelligence is a new paradigm, or a paradigm that tries to move the public narrative away from the common focus on replacement—AI or robots taking over our jobs. While that’s an understandable fear, more scientifically and societally, I think it’s more interesting and relevant to think of humans and AI as collaborators.

In this sense, human and AI means a synergy where teams of humans and AI together lead to superior outcomes than either the human or the AI operating in isolation. In a human-AI team, members can compensate for each other’s weaknesses and amplify each other’s strengths. The goal is not to substitute human capabilities, but to augment them.

This immediately moves the discussion from “what can the AI do to replace me?” to “how can we design the best possible team to work together?” I think that’s the foundation of the concept of hybrid intelligence. So hybrid intelligence, per se, is the ultimate goal. We aim at designing or engineering these human-AI teams so that we can effectively and responsibly collaborate together to achieve this superior type of intelligence, which we then call hybrid intelligence.

Ross: That’s fantastic. And so extremely aligned with the humans plus AI thesis. That’s very similar to what I might have said myself, not using the word hybrid intelligence, but humans plus AI to say the same thing. We want to dive into the humans-AI teaming specifically in a moment.

But in some of your writing, you’ve commented that, while others are thinking about augmentation in various ways, you point out that these are not necessarily as holistic as they could be. So what do you think is missing in some of the other ways people are approaching AI as a tool of augmentation?

Davide: Yeah, so I think when you look at the literature—as a computer scientist myself, I notice how easily I fall into the trap of only discussing AI capabilities. When I talk about AI or even human-AI teams, I end up talking about how I can build the AI to do this, or how I can improve the process in this way. Most of the literature does that as well. There’s a technology-centric perspective to the discussion of even human-AI teams.

We try to understand what we can build from the AI point of view to improve a team. But if you think of human-AI teams in this way, you realize that this significantly limits our vocabulary and our ability to look at the team from a broader, system-level perspective, where each member—including and especially human team members—is treated individually, and their skills and identity are considered and leveraged.

So, if you look at the literature, you often end up talking about how to add one feature to the AI or how to extend its feature set in other ways. But what people often miss is looking at the weaknesses and strengths of the different individuals, so that we can engineer for their compensation and amplification. Machines and people are fundamentally different: humans are good at some things, AI is good at others, and we shouldn’t try to negate or hide or be ashamed of the things we’re worse at than AI, and vice versa. Instead, we should leverage those differences.

For instance, just as an example, consider memory and context awareness. At the moment, at least, AI is much more powerful in having access to memory and retrieving it in a matter of seconds—AI can access basically the whole internet. But often, when you talk nowadays with these language model agents, they are completely decontextualized. They talk in the same way to millions across the world and often have very little clue about who the specific person is in front of them, what that person’s specific situation is—maybe they’re in an airport with noise, or just one minute from giving a lecture and in a rush. The type of things you might say also change based on the specific situation.

While this is a limitation of AI, we shouldn’t forget that there is the human there. The human has that contextual knowledge. The human brings that crucial context. Sometimes we tend to say, “Okay, but then we can build an AI that can understand the context around it,” but we already have the human for that.

Ross: Yes, yes. I don’t think that’s what I call the framing. Framing should come from the human, because that’s what we understand—including the ethical and other human aspects of the context, as well as that broader frame. It’s interesting because, in talking about hybrid intelligence, I think many who come to augmentation or hybrid intelligence think of it on an individual basis: how can an individual be augmented by AI, or, for example, in playing various games or simulations, humans plus AI teaming together, collaborating. But the team means you have multiple humans and quite probably multiple AI agents.

So, in your research, what have you observed if you’re comparing a human-only team and a team which has both human and AI participants? What are some of the things that are the same, and what are some of the things that are different?

Davide: Yes, this is a very interesting question. We’ve recently done work in collaboration with a number of researchers from the Hybrid Intelligence Center, which I am part of. If you’re not familiar with it, the Hybrid Intelligence Center is a collaboration that involves practically all the Dutch universities focused on hybrid intelligence, and it’s a long project—lasting around 10 years.

One of the works we’ve done recently is to try to study to what extent established properties of effective human teams could be used to characterize human-AI teams. We looked at instruments that people use in practice to characterize human teams. One of them is called the Team Diagnostic Survey, which is an instrument people use to diagnose the strengths and weaknesses of human teams. It includes a number of dimensions that are generally considered important for effective human teams.

These include aspects like members demonstrating their commitment to the team by putting in extra time and effort to help it succeed, the presence of coaches available in the team to help the team improve over time, and things related to the satisfaction of the members with the team, with the relationships with other members, and with the work they’re doing.

What we’ve done was to study the extent to which we could use these dimensions to characterize human-AI teams. We looked at different types of configurations of teams—some had one AI agent and one human, others had multiple agents and multiple humans, for example in a warehouse context where you have multiple robots helping out in the warehouse that have to cooperate and collaborate with multiple humans.

We tried to understand whether the properties of—by the way, we also looked at an interesting case, which is human-animal-animal teams, which is another example that’s interesting in the context of hybrid intelligence. You see very often in human-animal interaction—basically two species, two alien species—interacting and collaborating with each other. They often manage to collaborate pretty effectively, and there is an awareness of what both the humans and the animals are doing that is fascinating, at least for me.

So, we tried to analyze whether properties of human teams could be understood when looking at human-AI teams or hybrid teams, and to what extent. One of the things we found is that some concepts are very well understood and easily applicable to different types of hybrid teams. For example, the idea of interdependence—the fact that members in the team, in order to be a team, need to be mutually dependent, at least to some extent. Otherwise, if they’re all doing separate jobs, there’s a lack of common goal.

There are also things related to having a clear mission or a clear objective as a team, and aspects related to the possibility of exhibiting autonomy in the operation of the team and taking initiative. Also, the presence and awareness of team norms, like a shared ethical code or shared knowledge about what is appropriate or not. These were things that we found people could easily understand and apply to different configurations of teams.

Ross: Just actually, one thing—I don’t know if you’re familiar with the work of Mohammad Hussain Johari, who did this wonderful paper called “What Human-Horse Interactions May Teach Us About Effective Human-AI Interactions.” Again, these are the cases where we can have these parallels—learning how to do human-AI interactions from human-human and human-animal interactions. But again, it comes back to that original question: what is the same? I think you described many of those facets of the nature of teams and collaboration, which means they are the same. But there are, of course, some differences. One of the many differences is accountability, essentially, where the AI agents are not accountable, whereas the humans are. That’s one thing. So, this allocation of decision rights across different participants—human and AI—needs to take into account that they’re not equal participants. Humans have accountability, and AI does not. That’s one possible example.

Davide: Yeah, definitely. I totally agree, and I remember the paper you mentioned. I agree that human-animal collaboration is a very interesting source of inspiration. When looking at this paper, we looked at the case of shepherds and shepherd dogs. I didn’t know much about it before, but then I started digging a little bit. Shepherd dogs are trained at the beginning, but over time, they learn a type of communication with the shepherd. Through whistles, the shepherd can give very short commands, and then the shepherd dogs—even in pairs—can quickly understand what they need to do. They go through the mountains, collect all the sheep, and bring them exactly as intended by the shepherd, with very little need for words or other types of communication. They manage to achieve their goals very effectively.

So, I think we have a lot to learn from these cases, even though it’s difficult to study. But just to mention differences, of course—one of the things that emerged from this paper is the inherent human-AI asymmetry. Like you mentioned, accountability is definitely one aspect. I think overall, we should always give the human a different type of role in the team, similar to the shepherd and the shepherd dogs. There is some hierarchy among the members, and this makes it possible for humans to preserve meaningful control in the interactions. This also implies that different rules or expectations apply to different team members.

Beyond these, there is asymmetry in skills and capabilities, as we mentioned earlier, and also in aspects related to the identity of the members. For instance, some AI could be more easily replaceable than humans. Think, for example, of robots in a warehouse. In a human team, you wouldn’t say you “replace” a team member—it’s not the nicest way to say you let someone go and bring someone else in. But with robots, you could say, “I replace this machine because it’s not working anymore,” and that’s fine. We can replace machines with little consequence, though this doesn’t always hold, because there are studies showing that people get attached to machines and AI in general.

There was a recent case of ChatGPT releasing a new version and stopping the previous one, and people complained because they got attached to the previous version. So, in some cases, replacing the AI member would work well, but in others, it needs to be done more carefully.

Ross: So one of the other things looked at is the evaluation of human-AI teams. If we’re looking at human teams and possibly relative performance compared to human-AI teams, what are ways in which we can measure effectiveness? I suppose this includes not just output or speed or outcomes, but potentially risk, uncertainty, explainability, or other factors.

Davide: Yes, this is an interesting question, and I think it’s still an open question to some extent. From the study I mentioned earlier, we looked at how people measure human team effectiveness. There are aspects concerning, of course, the success of the team in doing the task, but these are not the only measures of effectiveness that people consider in human teams.

People often consider things related to the satisfaction of the members—with their teammates, with the process of working together, and with the overall goals of the team. This often leads to reflection from the team itself during operation, at least in human teams, where people reassess and evaluate their output throughout the process to make sure satisfaction with the process and relationships goes well over time.

In general, there are aspects to measure concerning the effectiveness of teams related to the process itself, which are often forgotten. It’s a matter, at least from a research point of view, of resources, because to evaluate a full process over time, you need to run experiments for longer periods. Often people stop at one instant or a few interactions, but if you think of human teams, like the usual forming, storming, norming, and performing, that often goes over a long time. Teams often operate for a long time and improve over time. So, the process itself needs to be monitored and reassessed over time. This is a way to also measure the effectiveness of the team, but over time.

Ross: Interesting point, because as you say, the dynamics of team performance with a human team improve as people get to know each other and find ways of working. They can become cohesive as a team. That’s classically what happens in defense forces and in creating high-performance teams, where you understand and build trust in each other. Trust is a key component of that.

With AI agents, if they are well designed, they can learn themselves or respond to changing situations in order to evolve. But it becomes a different dynamic when you have humans building trust and mutual understanding, where that becomes a system in which the AI is potentially responding or evolving. At its best, there’s the potential for that to create a better performing team, but it does require both the attitudes of the humans and well the agents.

Davide: Related to this—if I can interrupt you—I think this is very important that you mentioned trust. Indeed, this is one of the aspects that needs to be considered very carefully. You shouldn’t over-trust another team member, but also shouldn’t under-trust. Appropriate trust is key.

One of the things that drives, at least in human teams, trust and overall performance is also team ethics. Related to the metrics you mentioned earlier, the ability of a team to gather around a shared ethical code and stick to that, and to continuously and regularly update each other’s norms and ensure that actions are aligned with the shared norms, is crucial. This ethical code significantly affects trust in operation.

You can see it very easily in human teams: considering ethical aspects is essential, and we take them into account all the time. We respect each other’s goals and values. We expect our collaborators to keep their promises and commitments, and if they cannot, they can explain or justify what they are doing. These justifications are also a key element. The ability to provide justifications for behavior is very important for hybrid teams as well. Not only the AI, but also the human should be able to justify their actions when necessary.

This is where the concept of hybrid teams and, in general, hybrid intelligence requires a bit of a philosophical shift from the traditional technology-centric perspective. For example, in AI, we often talk about explainability or explainable AI, which is about looking at model computations and understanding why a decision was made. But here, we’re talking about a different concept: justifiability, which looks at the same problem from a different angle. It considers team actions in the context of shared values, shared goals, and the norms we’ve agreed upon. This requires a shift in the way we implement AI agents—they need to be aware of these norms, able to learn and adapt to team norms, and reason about them in the same way we do in society.

Ross: Let’s say you’ve got an organization and they have teams, as most organizations do, and now we’re moving from classic human teams to humans plus AI teams—collaborative human-AI teams. What are the skills and capabilities that the individual participants and the leaders in the teams need to transition from human-only teams to teams that include both humans and AI members?

Davide: This is a complicated question, and I don’t have a full answer, but I can definitely reflect on different skills that a hybrid team should have. I’m thinking now of recent work—not published yet—where we started moving from the quality model work I mentioned earlier towards more detailed guidelines for human-AI teams. There, we developed a number of guidelines for organizations for putting in place and operating effective teams.

We categorized these guidelines in terms of different phases of team processes. For instance, we developed guidelines related to structuring the teamwork—the envisioning of the operations of the team, which roles the team members would have, which responsibilities the different team members should have. Here, I’m talking about team members, but I’m still referring to hybrid teams, so this applies to both humans and AI.

This also implies different types of skills that we often don’t have yet in AI systems. For example, flexible team composition is a type of skill required to make it possible at the early stage of the team to structure the team in the right way. There are also skills related to developing shared awareness and aspects related to breaking down the task collaboratively or ensuring a continuous evolution of the team over time, with regular reassessment of the output.

If you think of these notions, it’s easy to think about them in terms of traditional organizations, but when you imagine a human-AI team or a small hybrid organization, then this continuous evolution, regular output assessment, and flexible team composition are not so natural anymore. What does it mean for an LLM agent to interact with someone else? Usually, LLM architectures rely on static roles and predefined workflows—you need to define beforehand the prompts they will exchange—whereas humans use much more flexible protocols. We can adjust our protocols over time, monitor what we’re doing, and reassess whether it works or not, and change the protocols.

These are skills required for the assistants, but also for the organization itself to make hybrid teaming possible. One of the things that emerges in this recent work is a new figure that would probably come up in organizations: a team designer or a team facilitator. This is not a team member per se, but an expert in teams and AI teammates, who can perhaps configure the AI teammates based on the needs of the team, and provide human team members with information needed about the skills or capabilities of the specific AI team member. It’s an intermediary between humans and AI, with expertise that other human team members may not have, and could help these teams work together.

Ross: That’s fantastic. It’s wonderful to learn about all this work. Is there anywhere people can go to find out more about your research?

Davide: Yeah, sure. You can look me up at my website, davidedellanna.com. That’s my main website—I try to keep it up to date. Through there, you can see the different projects I’m involved in, the papers we’re working on, both with collaborators and with PhD and master students, who often bring great contributions to our research, even in their short studies.

That’s the main hub, and you can also find many openly available resources linked to the projects that people may find useful.

Ross: Fantastic. Well, it’s wonderful work—very highly aligned with the idea of hybrid intelligence, and it’s fantastic that you are focusing on that, because there’s not enough people yet focusing in the area. So you and your colleagues are ahead, and I’m sure many more will join you. Thank you so much for your time and your insights.

Davide: Thank you so much, Ross. Pleasure to meet you.