Designing a New AI
A DISCUSSION OFAn AI That’s Not Artificial at All
Read Responses From
Immediately upon securing the Normandy beachheads in World War II, American forces were faced with a task for which they were wholly unprepared. The breakout from Normandy required Allied forces to penetrate the hedgerow country that was ideally suited for German defensive operations seeking to stall the offensive. Neither infantry nor tanks could easily penetrate these hedgerows, and the defenses quickly decimated American tanks and infantry when they operated independently. “Technology” solutions soon emerged when soldiers used metal from the German beach defenses to create “teeth” on the front of American tanks. Infantry and armor soldiers soon developed new tactics jointly maneuvering upon breaching a hedgerow to overcome the German defensive kill zones. The leadership, courage, and innovation of these American forces enabled them to break out from the landing sites before German reserves could decisively respond and push the Allies back into the English Channel.
Innovation often appears intuitive in hindsight, and the American success can appear preordained. This would be a revisionist view of the situation at the time as well as a misleading indication for future success. As organizations incorporate growing numbers of cyber-physical systems, artificial intelligence reasoning, and diverse and potentially distributed human teams, broad and rapid innovation as witnessed in the “Norman Bocage” is not guaranteed. In “An AI That’s Not Artificial at All,” (Issues, Fall 2021) John Paschkewitz, Bart Russell, and John Main lay the foundation for innovation and learning in such future organizations through their proposed concept of liminal design.
Soldiers at Normandy first attempted to use explosives to penetrate the hedgerows, but there were not enough explosives to broadly employ this strategy. Only when soldiers abstracted the problem of “hedgerow penetration” as the function could they identify the German steel beach defenses as a viable source. One can view this simply as “looking at the problem in a different way,” but future organizations consisting of AI decision aids will struggle to perform similar composition without a structured language and design methodology to do so. Cyber-physical systems may provide a broad set of services within current organizations, but new situations require these organizations to mediate among these broadly available resources within the organization to address a local problem.
The most crucial aspect of the liminal design framework that Paschkewitz and his coauthors propose is the need for hybrid organizations to learn. Returning to the example of Normandy, tankers and infantrymen did not trust each other because they had not trained previously to work together. They had to form new hybrid teams with new tactics on the fly in the face of a stiff German defense. Imagine the challenge of seeking to create novel, hybrid human-machine teams that don’t even share the same native language. The concept of digital twins that the authors describe would allow current organizations to experiment and breed the trust needed in the face of dynamic environments is compelling and essential.
Finally, liminal design should not be considered orthogonal to human-centered design or systems thinking, but rather a modern complement to structured problem-solving for hybrid organizations faced with pressing problems in dynamic environments. This is just as true today as it was in Normandy 75 years ago.
Acting Director, Defense Sciences Office
Defense Advanced Research Projects Agency
John Paschkewitz, Bart Russell, and John Main propose a new design approach that they term liminal design for collaboration between human and artificial intelligence agents. As a design method it mirrors many of the exciting ideas found in the best product innovation methods. These methods include human-centered design that includes a wealth of sometimes conflicting stakeholders. Examples are found in medical equipment, where the insurance provider focuses on costs, the physician on care, and the patient on comfort; functional modeling, which raises the search process to a more abstract what does it need to do level; configuration design, which provides a formalism for composing complete solutions from constituent components, resulting in the composition of a product that meets that abstract functionality; the concept of slack in supplier-to-product developer and producer negotiations that results in a sweet spot of resolution that trades off different individual goals (a mediated solution); negotiations that result from perceptual gaps between those that must contribute to a design solution (another approach to mediation); and systems that are designed to learn in the context of Industry 4.0.
The liminal design approach raises the need to explicitly address mediation between domains with divergent needs, goals, and problem representations. The formality necessary to manifest this concept may be achievable through market-based mediation, an approach that balances the maximization of outcomes (such as profit) for different participants in the design process. This formal approach could raise the negotiations that occur between players to a more optimal or at least better-resolved finality. Mediation is a current challenge that appears in domains such as infrastructure (the need to mediate between construction and engineering), automobile design (the need to mediate between engineers and studio designers), and additive manufacturing (the need to mediate between design engineers and those intimately familiar with the limitations of specific processes and printers). Mediation is perhaps one of the most pervasive and important perceptual gaps, or liminal spaces, that requires dedicated attention from both industry and academia.
Liminal design also invites the exploration of how artificial intelligence can serve the human team, especially in this mediation process. For instance, AI can help the designers achieve more articulated solutions through tracking how designs are progressing and how team members focus or fixate on different aspects, providing incentive or suggestion to shift the problem-solving direction. AI can also assess trade-off scenarios that lead to mediated solutions faster. Moreover, the possibility of elevating AI from simply a tool used by problem solvers to more of a proactive, adaptive, and responsive partner in the design process offers additional potential for supercharging the search for high-performance solutions, an area of research that we, the authors, currently focus much of our own effort and interests on. The creation of AI tools and partners that advance solutions to these and other wicked problems may ultimately have a transformational impact on the many liminal spaces in which we live, work, and create.
George Tallman and Florence Barrett Ladd Professor
Department of Mechanical Engineering
Carnegie Mellon University