Hybrid Human-AI Conference: Day 1 of Summer in Munich

Our venue, Munich’s Design Studio, is a beautiful celebration of the design itself, from the architecture to the furniture to the artful espresso machines.

Summer in Munich is hot and full of energy. I'm here attending the second International Conference on Hybrid Human-Artificial Intelligence (HHAI), where scholars from around the world gather to examine the latest research on how humans and AIs can work together safely, effectively, and ethically. Given all the hype about AI in the past few months, it might be tempting to write off human-AI interaction as a trend – but it's better thought of as the bedrock for organizational strategy and individual career planning for the rest of our time. The further AI gets integrated into our lives, the trickier the questions we must answer about how our values are reflected in these systems!

Our venue, Munich’s Design Studio, is a beautiful celebration of the design itself, from the architecture to the furniture to the artful espresso machines.

Three Key Learnings

Be responsible!

At BCG, we often advocate for "responsible AI." What does responsibility mean? Take a moment to consider the following statement: "The responsibility for errors or harm caused by hybrid human-AI systems lies with the human operator, not the AI component." Do you agree with it? Let's break it down. First, there is a difference between legal and moral responsibility. Legal compliance is important, but leaders take a more personal assumption of responsibility. Everyone is quick to take responsibility when things go well, but what about when a self-driving car hits a pedestrian or a faulty investing algorithm erases millions of dollars of value? Accidents happen. As we do with automobiles and airplanes, it's critical to hold human designers accountable for releasing faulty and dangerous systems – but also to implement consequences for misuse.

In the first session of the day, organizers asked us to answer provocative questions about ACI by standing on a line between “yes” and “no.”

Are machines intelligent?

I struggled with this question last week in Santa Fe but got new insight today! There seem to be two competing definitions of intelligence in the greater AI community: one describing brains like computers and the other describing intelligence as a hyper-complex interdisciplinary process. This first definition stems from *cognitivism,* an early cognitive science approach that views cognition as distinct mental states governed by rules, like computers. Most current measures of AI capabilities operate in this frame. Meanwhile, 4E Cognition explores how intelligence results from complex interactions with other agents in our social and physical environments. There's a tension between being *thorough* - using 4E Cognition to understand AI – and moving quickly. According to adherents of 4E cognition, AI isn't actually intelligent or engaging with the learning process. And that may be true, but in the meantime, we need something to call it!

Accuracy is not enough

Imagine your team lead proposes using an AI tool with a success rate of 99%. Would that be acceptable? Maybe – it depends on what happens in the other 1%. In "AI Shall Have No Dominion - on How to Measure Technology Dominance in AI-supported Human Decision-Making," Federico Cabitza's model shows that some errors are more impactful than others. How often the AI tool is right is not the same question as "How effectively does it support human decision-making," the latter is much more important for trust and reliance, which are built on reputation and perception. I recommend reading the paper to dig into the model he and his co-authors created, which can increase human trust in AIs by estimating AI tools' sensitivity (ability to avoid false negatives) and specificity (avoiding false positives).

Friendly student volunteers helped me find the right room after several wrong turns.

Three Humans of HHAI

Federico Cabitza: Computer scientist and professor of Human-Machine and Human-Data interaction who has authored more than 120(!) publications on medical informatics and human-AI decision-making support in the hospital sector. I'm especially interested in his human-AI interaction assessment model. https://dss-quality-assessment.vercel.app/step/4 

Wijnand IJsselsteijn is a Cognition and Affect in Human-Technology Interaction professor at Eindhoven University of Technology (TU/e). His work includes how insights from psychology can be used to design better technology models and interfaces.

According to leading cognitive scientists, donuts make you really smart.

Jonne Maas is a Ph.D. student studying the power dynamics underlying the development and use of AI systems, including what political philosophy can teach us about what to expect from (and demand of) these processes. She's giving a workshop presentation tomorrow that I'm quite excited to see!

Two sessions I enjoyed

Exploring Responsible AI by Design was the first of two half-day tutorials featuring multiple-topic lectures, interactive exercises, and small group discussions about which values we embed in our designed systems (i.e., what can/should be deferred to AI). My favorite talk was Wijnand's, whose cognitivist / 4E discussion helped me nail down a critical axis of disagreement in the field.  

Attendees of Exploring Responsible AI by Design debate the extent of privacy risks stemming from hybrid human-AI systems.

How to Assess Human Reliance on Artificial Intelligence in Hybrid Decision-Making, led by Federico, started to answer the difficult questions of how to determine whether human-AI collaboration is effective – his model is fairly complex, and I'm still trying to wrap my head around its potential applications, but it's desperately needed. Like the other attendees at the Santa Fe Collective Intelligence Symposium, Abhishek and I have found measuring intelligence (especially collective intelligence) to be almost prohibitively difficult, so Federico's work is an exciting step forward.

Looking forward to tomorrow

This place is perfect if the designed environment can improve creative thinking.

As the jet lag eases, I'm hungry to find places where the rubber meets the road on these different definitions of "intelligence." What should we do if intelligence has a different meaning than how we currently define it in the industry? How can we 1) stop perpetuating an overly simplistic and faulty conception of our core human capabilities and 2) discover and apply more useful benchmarks to assess AI systems' fitness as collaborators?