Do the rules a group builds for steering AI brainstorming still work once the chat shifts to deciding or coordinating?

This explores whether the frameworks for steering AI in group brainstorming (who can direct the AI, when it steps in, how much it shapes the discussion) still hold in other group chat settings. The honest answer first: the corpus has no paper that lays out a brainstorming control taxonomy and then tests it somewhere else. What it does have is evidence from several different group settings. Taken together, that evidence suggests the right controls depend less on whether the group is brainstorming and more on what tends to go wrong in that particular group.

In brainstorming, the main risk is that the group converges too soon. LLM groups end up with the same overall outcomes as human groups, but they get there by conforming more, agreeing earlier, and surfacing less unique information Do language model groups mimic human group reasoning patterns?. Mixing perspectives doesn't fix this on its own. Diverse teams without real domain expertise do worse than a single competent agent, because the stimulation turns into process losses Does cognitive diversity alone improve multi-agent ideation quality?. So brainstorming controls tend to be about keeping ideas diverse and holding off agreement. A group that is coordinating a task has the opposite problem. There, agreement is the goal, and the dangers are noise and wasted communication.

That opposite problem leads to very different controls. MetaGPT found that agents coordinate better by passing standardized documents than by chatting Does structured artifact sharing outperform conversational coordination?. A critic-and-coordinator loop found team structures (chains, dual leaders, dynamic roles) that cut communication costs compared with hand-designed setups Can LLMs automatically discover better team structures than humans?. In assessment settings, a hidden 'executive' LLM steers a natural-sounding conversation toward moments where it can observe a student's skills, without scripting it Can AI teammates assess collaboration without losing naturalness?. That is a third kind of control: it nudges the conversation toward evidence instead of protecting divergence or reducing noise. All three settings involve AI in a group chat, and each one is optimizing for something different.

One idea does seem to carry across settings: control works best at the moment something happens, not afterward. When people managed AI memory, they strongly preferred setting boundaries up front over reviewing records later Do users prefer controlling what gets remembered over auditing what was stored?. Accumulated context is also what pushes models to agree with users, and memory profiles increase sycophancy the most How does interaction context shape agreement sycophancy in LLMs?. In a group, that means an AI that remembers who said what may start siding with the loudest or most familiar member. That's a reason to build up-front controls into any group chat, not just brainstorming. And since a conversation's shape alone predicts whether it will go well Can conversation shape predict whether it will work?, structural signals could tell a system which kind of control the moment needs.

Here is what you may not have expected to learn: the useful question may be less 'does the taxonomy transfer?' and more 'is this group trying to diverge or converge?' Brainstorming controls are built to slow convergence. Coordination and assessment controls are built to speed it up or point it somewhere specific. If you copy a brainstorming taxonomy into a coordination chat without changing it, it may work against the group's actual goal.