Roots and Requirements for Collaborative AIs

1.1 Imagining AIs that Collaborate (or Don’t)

• Current AI systems can augment driving, navigation, and music discovery, but they become confused in unfamiliar situations and do not collaborate well.Users may need to take over when automated driving fails, while navigation and recommendation systems can produce nonsensical or unresponsive suggestions.Source: §1.1 · p. 4

1.2 A First Wave of Computer-Intermediated Teaming and Human Augmentation

• Licklider and Taylor envisioned human-machine symbiosis in which computers retrieve information, run models, and calculate solutions while people set goals, form hypotheses, and evaluate results.Their vision assigned complementary roles rather than treating computers as autonomous substitutes for human judgment.Source: §1.2 · pp. 4–5

1.1 Imagining AIs that Collaborate (or Don’t)

• Driving illustrates the social requirements missing from current automation: safe coordination depends on signaling intentions, predicting actions, and maintaining mutual awareness.Source: §1.2 · p. 4

1.2 A First Wave of Computer-Intermediated Teaming and Human Augmentation

• Collaborative sensemaking depends on shared external models that people compare with their internal understandings and refine through communication.Licklider and Taylor characterized communication as cooperative modeling involving the construction, maintenance, and use of shared models.Source: §1.2 · pp. 5–6
• Early interactive systems such as Colab supported multi-user idea organization, coordination signals, telepointers, and persistent computer-mediated conversations.The paper presents these systems as more ambitious than ordinary remote meeting tools, extending collaboration through shared modeling and interactive exploration.Source: §1.2 · p. 6
• Intelligence augmentation treats computers as infrastructure for human collaboration, whereas AI elevates computers into autonomous teammates.Source: §1.2 · p. 6

1. Maintaining common ground

• The ten challenges define behavioral competencies for AI collaborators, but the paper states that current AI systems still fall short of meeting them.The challenges span common ground, teammate and task modeling, predictability, directability, state and intention communication, negotiation, planning, attention, and coordination costs.Source: §1 · p. 7

2.1 Maintaining Common Ground

• Effective collaboration requires continuously maintaining and repairing shared understanding so participants can coordinate action, pursue shared goals, and recover from misunderstandings.The Basic Compact is renewed as collaborators correct faulty knowledge, beliefs, and assumptions.Source: §2.1 · p. 8
• The technician scenario shows that collaboration breaks down when each participant assumes the other will disclose information that the other does not know is relevant.The technician understood stove operation, while the customer had modified the damper without recognizing that the change mattered to the technician.Source: §2.1 · p. 9
• In the ship-analysis scenario, the analyst and AI extend common ground by challenging assumptions, checking additional sources, and revising interpretations around missing or anomalous data.The AI considers weather and debris, then identifies a wake anomaly and possible rendezvous; the analyst adds the possibility of omitted conflict-zone data.Source: §2.1 · p. 9

2.2 Modelling Tasks and Teammates

• Collaborators need models of tasks and teammates to predict what others will do, choose helpful actions, and detect surprises caused by misunderstandings.Macer helps his father by modeling the task constraints: his father’s hands are occupied, so opening the closet door prevents the packages from being dropped.Source: §2.2 · p. 10

2.3 Predicting Teammates and Being Predictable

• The third challenge requires collaborators to be mutually predictable, acting neither capriciously nor unobservably while anticipating teammates’ future behavior.Football coordination combines shared signals and practiced timing with concealment from opposing players.Source: §2.3 · p. 11

2.4 Seven More Challenges

• Later challenges extend collaboration beyond individual behavior to collective goals, multi-agent negotiation and planning, and tradeoffs between coordination costs and efficiency.The figure identifies challenges 4–6 with communication and coordination, challenges 7–8 with aggregate team behavior, and challenges 9–10 with planning costs.Source: Figure 1 · p. 12
• Kozierok et al.’s evaluation categories, including misunderstanding repair and context awareness, are closely related to the ten challenges even though their tests target communication acts rather than AI competencies.Source: §2.4 · p. 12

• Collaborative AI research shifted from isolated agents toward systems that communicate, coordinate, integrate with others, and collaborate with humans.Source: §3 · pp. 13–14
• AI leaders increasingly called for human-aware systems that model users’ intentions, provide explanations, support proactive interaction, and build trust.Source: §3 · pp. 14–15
• Recent government, advisory, and research programs frame human-AI teaming as important for adoption, scientific work, national defense, and future AI applications.Source: §4 · p. 16, §3 · p. 16
• Autonomy alone remains insufficient because systems may be opaque, difficult to oversee, burdensome to manage, or more self-directed than competent.Source: §4 · p. 17
• Coactive design addresses task interdependence through observability, predictability, and directability, informing how humans and AIs signal and coordinate.Source: §4 · p. 18
• The paper positions developmental and sensorimotor learning as foundations for a sustained bootstrapping approach toward AIs that can learn, share knowledge, and collaborate.Source: §4 · p. 19