Should we group human-AI teamwork by the tools used, or by who's actually in charge of the work?

This explores two ways of sorting human-AI collaboration: by the form the work takes (talking, sharing documents, using a generated interface) versus by who is in charge (who picks the task and who decides how to do it). The corpus has no single paper that compares the two directly, but it has solid material on each side, and putting them next to each other shows that they answer different questions.

The clearest control-based typology comes from a study of 244 BCG consultants Do humans and AI divide labor differently in knowledge work?. It found three patterns: Centaurs, who split the work cleanly with the AI; Cyborgs, who weave their work tightly into the AI's; and Self-Automators, who hand off most of the job. The key finding is that the tool and the task did not decide which pattern people fell into. What decided it was who chose the task and who chose the method. So a control typology tells you where judgment and authority sit, and it makes no claim about how the collaboration looks or feels.

Modality frameworks look at the channel instead, and the corpus shows the channel has strong effects of its own. In a controlled experiment, changing only the way humans and agents communicated changed how much people trusted the agent and how well they kept track of the shared workspace. These were the same patterns researchers had already seen in human teams How do communication modalities shape human-agent collaboration patterns?. Between agents, MetaGPT found that passing standardized documents back and forth worked better than open-ended conversation Does structured artifact sharing outperform conversational coordination?. On the interface side, users preferred AI-generated dashboards and tools to chat in over 70% of cases Do generated interfaces outperform text-based chat for most tasks?, although those interfaces become rigid when someone wants to change direction partway through Do generated analysis UIs really work better than chat?. Even inside conversation, matching someone's wording and matching their emotional tone do different jobs. The first helps get the task done and the second builds warmth and trust Do different types of alignment serve different conversational goals?.

The less obvious point is that modality quietly moves control around. Chat leaves the method open, so the human keeps steering, which suits Cyborg-style work. A generated interface or a fixed document format builds the method into the medium, so whoever designed it has already made part of the 'how' decision. LLM Programs push this furthest: an outside algorithm controls the flow and shows the model only what each step needs Can algorithms control LLM reasoning better than LLMs alone?. Here the modality is the control structure. The same thing happens as AI moves from chatbot to colleague. Persistent memory and reusable skills change who holds the procedure over time What makes an AI system feel like a colleague rather than a chatbot?.

So the two frameworks fit together, but they are not independent. Control typologies describe who owns the decisions, and modality frameworks describe the medium those decisions pass through. Because the medium can carry decisions inside it, choosing a modality is often a hidden choice about control. A team that switches from chat to a structured interface may be turning Cyborgs into Centaurs without noticing.