Discussing an AI Answer Together

Three students designing a cultural educational game had access to the same ChatGPT output. In an observed session, one member pointed to a passage and said, “We can use this part for the second module.” The group discussed that passage and divided subtasks around it. The student had contributed a proposal about where the answer belonged in their work.

That action helps put a collaborative-search study posted November 3 in perspective. Students researching in pairs reported lower demands and easier collaboration with answer-generating assistants embedded in their shared chat than with independent Google searching. Questions and answers appeared where both partners could see them, reducing the need to bring material back from a private search. The game-design episode comes from a separate study, not an explanation experimentally established for those ratings.

For colleagues investigating a question together, shared access changes what one person needs to contribute. Relaying an answer and proposing a use for it are different jobs. An assistant can put material into everyone’s view while leaving people to select a passage, question its relevance and connect it to a decision. The practical choice is when a colleague’s judgment needs to enter that process.

What the shared channel changes

In the newer study, Peiyuan Gong and colleagues recruited 80 students in 40 pairs. Each pair researched four assigned topics, including AI painting and air pollution, with a different search setup each time. Topics, tools and order were counterbalanced across pairs. Partners sat in separate rooms, communicated only through Slack and could not switch freely among search tools. Each wrote an individual report afterward. These were constrained research sessions, not coworkers choosing their normal mix of browsers, calls and documents.

All four setups used Slack for communication. Google required each person to search outside the channel and return with selected findings, excerpts or links. SearchAgent put search-result lists directly into Slack without synthesizing an answer. QAAgent generated answers from its model’s existing knowledge, without retrieving web sources. CoSearchAgent interpreted the conversation, could ask clarifying questions, retrieved web material and returned answers with source references.

The experience ratings favored the answer-generating assistants on several measures. Both lowered perceived mental demand and effort compared with Google. QAAgent had the highest mean on all five collaboration questions, covering awareness of the partner’s activity, information sharing, division of work and smooth communication. Its means ranged from 5.83 to 6.41 on seven-point scales. CoSearchAgent also scored above Google on awareness, sharing and communication.

The conventional search plugin provides an important comparison. Its results were visible in Slack, too, yet QAAgent’s collaboration ratings exceeded its ratings on awareness and sharing. Visibility alone cannot account for all the differences. Ready-made prose, source access, clarification and response delays also varied among the tools. The experiment compared those combinations rather than changing one feature at a time. Participants also rated the conventional plugin as more reliable than QAAgent.

The interface changes how to read another result: participants shared an average of 11.80 URLs per task with Google, compared with 2.08 with SearchAgent and 1.03 with CoSearchAgent. In the embedded setups, sources were already arriving in the shared channel. Posting a URL again could be redundant. Fewer manually shared links therefore do not establish that the partners encountered less evidence together.

Nor does a busier conversation necessarily mean more discussion between colleagues. Participants posted the most human-authored turns with QAAgent. But excluding bot replies leaves human requests to the bot in the count; it does not identify what the partners said to each other. To understand the collaboration, the content and destination of an exchange matter.

Giving an answer a role

The opening passage-selection episode comes from Haowei Xu and colleagues’ study, published in April. They observed 18 students in six trios designing cultural educational games, with about 45 minutes for the main task. Each group used one shared ChatGPT-4 interface. Three groups worked in person with a large display; three used Microsoft Teams and screen sharing. One person at a time controlled the AI input.

These observations show actions that experience ratings cannot. Participants drew attention to relevant output by reading it aloud or highlighting it with a cursor. The student proposing a passage for the second module did something more specific: connected a piece of text to part of the game, giving the others a proposal they could discuss and organize around. The answer became a reference for a particular piece of work.

The distinction also appeared when discussion split. In another trio, one student pursued a topic with ChatGPT while the other two talked together. The researchers report that the discussion converged after the first student invited the others to look at the AI-generated responses. Everyone had access to the shared interface before that invitation. Directing attention to its output was a further action.

That episode does not make independent prompting a mistake. In a different group, a student asked whether the game could use a Japanese role-playing-game format without first consulting the others. The response received positive feedback and became a productive topic for group discussion. Both cases began with individual initiatives; their subsequent place in the human conversation differed.

Partners could also participate before an answer existed. In the group discussing kendo culture, one student proposed the search direction, another helped refine it, and a third noticed an input error. The group helped the operator correct the input and submit the prompt. One person’s keyboard carried a question that several people had shaped.

These cases are possibilities observed in six assigned design sessions, not estimates of typical group behavior. They do not establish why Gong’s participants rated the assistants favorably. They make the work behind a shared answer concrete: people can help form a question, identify a relevant passage or propose how to use it. Seeing the same text leaves those contributions open.

When to bring a colleague in

A shared screen also concentrates control. Although everyone could see the game-design answers, the person at the keyboard could submit a question without consulting the others. In Gong’s Slack setup, either partner could invoke the assistant. Visibility and the ability to act on the system were arranged differently.

Private preparation offers another legitimate way to divide the work. Xu and colleagues interviewed nine additional students about earlier group projects. One described using AI to gather technical literature and suggestions, checking the material through other literature and then discussing the work at a meeting. Another revised an initially inadequate project outline before bringing it to teammates. These were retrospective accounts from different people, not a controlled comparison with the observed shared-screen sessions.

That preparation can include useful filtering. A colleague need not sit through every unsuccessful query to evaluate a selected proposal. But a proposal arrives after some choices have been made. If the group needs to question the scope of an inquiry, joining only after the answer has been selected may be too late for that particular contribution. If the question is already settled, individual exploration may give the later discussion better material. Those are choices about the task, not benefits established by a private-versus-shared experiment.

Some participants in Gong’s study wanted help managing what accumulated in the channel: conversation summaries, outlines and ways to mark content. These were requests for future features, not tested improvements. They point to a practical concern beyond obtaining another answer: keeping useful material available for later discussion.

Shared exploration and private preparation can serve different parts of the same inquiry. What matters is whether the colleague needs to help shape the question, examine the proposed evidence or decide where a passage belongs. In the game-design session, the answer was already on view. The student’s contribution was to give the group a specific use to consider.

Sources

Study findings are condensed and critically interpreted from Peiyuan Gong, Jiamian Li, Erhan Zhang, Fengming Liang, Kexin Liu, Yue Chen and Jiaxin Mao, Collaborative Search Meets Large Language Models: An Empirical Study, arXiv v1, November 3, 2026. Collaboration episodes and interview accounts are adapted from Haowei Xu, Ahmed Kharrufa, Ellis Solaiman and Vasilis Vlachokyriakos, AI as an Agent and Collaborative Space: Exploring the role of Generative AI in Small Group Synchronous and Asynchronous Collaborative Dynamics, CHI 2026. Both works are licensed under CC BY 4.0. The interpretation and practical implications here are not an endorsement by the researchers.