GRAIL: Group-AI Interaction Laboratory

• GRAIL is an experimental platform for real-time, multi-user chat studies that facilitates controlled research on group-AI interactions.
• It supports customizable tasks with private stimuli, rich transcript logging, and integrated survey tools to capture participant attitudes.
• LLM facilitation in GRAIL has been shown to boost information sharing and participation equality, though it may not improve final decision accuracy.

The Group-AI Interaction Laboratory (GRAIL) is an open-source experimental platform for studying group-AI interaction in real-time, multi-user chat settings. Introduced alongside a pre-registered randomized experiment on LLM-facilitated group decision making, GRAIL was designed to support dynamic AI or human facilitation, assignment of unique private stimuli to participants, rich transcript-level logging, and integrated onboarding and survey collection. In the source study, it served as the infrastructure for a hidden-profile decision task with 1,475 participants assigned to 281 groups, and it was released to support further research on collaborative decision making and group-AI interaction (Alsobay et al., 11 Aug 2025).

1. Definition, platform scope, and research role

GRAIL was developed as an experimental platform specifically for real-time, multi-user chat studies in which groups interact with either AI facilitators or human facilitators. It is built upon the Empirica framework and supports customizable group sizes and task designs, including assignment of unique, private information sets to individual participants. The platform also supports AI or human facilitation through mechanisms for interventions, message timing and scheduling, and integration with LLMs via API. Its instrumentation includes rich logging of full group transcripts, facilitator interventions, prompts, rationales, and metadata, as well as built-in consent, onboarding, and survey tools for capturing subjective participant attitudes (Alsobay et al., 11 Aug 2025).

Within the study that introduced it, GRAIL functioned as both experimental apparatus and reusable infrastructure. This dual role is important: it makes the platform not merely a chat interface, but a controlled environment for manipulating group composition, information asymmetry, facilitation style, and post-task measurement. The release of GRAIL as an open-source tool positions it as a reusable substrate for replication, extension, and comparative evaluation of group-AI systems in settings where simultaneously coordinating multiple real participants and a live facilitator is otherwise technically burdensome (Alsobay et al., 11 Aug 2025).

2. Experimental embodiment: the hidden-profile decision task

The platform’s initial demonstration centered on a cooperative hidden-profile task in which each group had to choose one of three fictional cities—Eldoron, Myloria, or Cragnio—to host an international sporting event. Each city was described by 10 facts, for a total of 30 facts, with information distributed across public, shared, and private channels. The crucial manipulation was informational asymmetry: each participant’s private information made Myloria appear best individually, but Eldoron was optimal when the group aggregated the full information set. City utility followed a unit-weight linear model,

$$\text{Utility} = (\# \text{ positive facts}) - (\# \text{ negative facts}),$$

under which full-information scores were Eldoron $= 6$, Myloria $= 2$, and Cragnio $= 0$ (Alsobay et al., 11 Aug 2025).

The study procedure consisted of onboarding with a comprehension check, a one-minute chat icebreaker, a ten-minute group discussion in the chat interface, and an exit survey. Participants were randomized into 281 groups, almost all with 5 members each. In the human facilitator condition, a sixth member acted as facilitator. This design used GRAIL to coordinate multi-party interaction, information control, facilitation timing, and outcome logging within a single experimental pipeline (Alsobay et al., 11 Aug 2025).

Condition	Facilitation mechanism	Key specification
None	No facilitation	Group completed task without facilitator or nudges
Message	Static prompt	“People may have different information about what is being discussed… encourage everyone to share all of the relevant information they have.”
Human	Sixth participant	Facilitator had no task-relevant information and could chat and @call individuals
LLM	GPT-4o facilitator	Intervened every 90 seconds and responded to direct @mentioning

The LLM facilitator received only generic instructions, not the city facts. Its prompt included the full prior chat, instructions to encourage information sharing and track pros and cons, and group and time context. Responses used structured JSON containing both a message and a rationale. This design isolated facilitation behavior from privileged task knowledge and thereby focused the intervention on coordination rather than answer provision (Alsobay et al., 11 Aug 2025).

3. Empirical findings from the initial GRAIL study

The primary outcome was total information coverage: the number of distinct facts, out of 30, surfaced in discussion. GRAIL’s LLM facilitation condition increased information sharing relative to all three comparison conditions. Relative to no facilitation, the LLM condition increased facts shared by $\Delta = 2.9$, $t(138) = 3.58$, $p < 0.001$, $d = 0.61$. Relative to the static message condition, the increase was $\Delta = 1.58$, $t(139) = 2.1$, $= 6$0. Relative to the human facilitator condition, it was $= 6$1, $= 6$2, $= 6$3. By contrast, there was no significant difference between Message and None or Human and None on fact coverage (Alsobay et al., 11 Aug 2025).

These gains did not arise from simple verbosity. LLM-facilitated chats were described as more information-dense, using the measure

$= 6$4

with LLM versus None yielding $= 6$5, $= 6$6, $= 6$7. Information breadth also changed: LLM-facilitated groups shared information more evenly across the three cities, reflected in a lower Gini coefficient, $= 6$8, $= 6$9, $= 2$0. At the participation level, LLM facilitation raised the minimum number of facts contributed by any group member by $= 2$1, $= 2$2, $= 2$3, and 61% of LLM groups achieved full participation, compared with 39–47% in the other conditions, $= 2$4, $= 2$5 (Alsobay et al., 11 Aug 2025).

A mechanism-level observation concerned direct engagement. The LLM addressed individuals in 68% of its messages, compared with 25% for human facilitators. All LLM-facilitated groups received @address, whereas only 77% of human-facilitated groups did. The source study interprets this pattern as reflecting more active prompting of quieter individuals and an ability to intervene assertively without social inhibition (Alsobay et al., 11 Aug 2025).

Despite these effects on discussion process, facilitation did not significantly improve final decisions. Optimal-choice rates for Eldoron were 31% in None, 21% in Message, 30% in Human, and 23% in LLM, with $= 2$6, $= 2$7. Multinomial logistic regression showed that group decisions tracked the utility scores of surfaced information, but Myloria retained disproportionate influence. For the odds of choosing Eldoron over Myloria, the reported model included coefficients

$= 2$8

This indicates that increased information sharing alone was insufficient to overcome the hidden-profile effect (Alsobay et al., 11 Aug 2025).

Subjective measures showed no significant differences between conditions on NASA Task Load Index subscales or on group attitudes such as cohesion and productivity. Participants rated the LLM as more effective than the human facilitator at eliciting and synthesizing information, but also as more distracting. Exposure mattered for future acceptance: participants who experienced LLM facilitation were more open to both LLM and human facilitators later, whereas participants in the human-facilitator condition were less open to future AI facilitators (Alsobay et al., 11 Aug 2025).

4. Position within the broader group-AI interaction literature

GRAIL occupies a methodological niche within a broader shift from individual-centered evaluation toward explicitly group-centered AI research. The provocation “Beyond Individual UX: Defining Group Experience(GX) as a New Paradigm for Group-centered AI” defines Group Experience as “the collective perceptual, emotional, and cognitive responses that emerge specifically when multiple individuals interact cohesively” and proposes Group-centered AI as a mezzo-level framework for mediating and amplifying group processes (Lee et al., 19 May 2025). In that framing, GRAIL provides a concrete experimental substrate for operationalizing group-level constructs such as participation equity, collective sensemaking, and collaborative decision quality.

Adjacent empirical work has emphasized that group-AI outcomes depend not only on model capability but also on how participation is structured and controlled. “Controlling AI Agent Participation in Group Conversations: A Human-Centered Approach” develops a taxonomy of controls over when, what, and where an AI agent responds, who can control it, and how those controls are specified and implemented (Houde et al., 28 Jan 2025). “I Felt Bad After We Ignored Her”: Understanding How Interface-Driven Social Prominence Shapes Group Discussions with GenAI” shows that agent presence in the shared space and the degree of user control systematically alter communication patterns, vigilance, and negotiation of agent influence across Roundtable, Peripheral, and Breakout modes (Johnson et al., 16 Feb 2026). The AI Collaborator platform similarly targets human-AI teaming research through customizable personas, a memory system based on the “Generative Agents” model, and modular Slack integration (Samadi et al., 2024).

Taken together, these works situate GRAIL as infrastructure rather than a single facilitation doctrine. Its significance lies in enabling controlled comparison of facilitation modes, intervention schedules, information asymmetries, and subjective responses in a reproducible group-chat environment. This suggests that its scientific value extends beyond the specific hidden-profile task used in its initial study.

5. Research questions and extensions suggested by adjacent work

Several nearby literatures identify variables that a platform such as GRAIL could plausibly be used to investigate. Socially grounded proactive generation is one example. “Social-RAG: Retrieving from Group Interactions to Socially Ground AI Generation” introduces a workflow that retrieves prior group interactions, selects social signals, and feeds them into a LLM to generate socially aligned messages; its deployment in PaperPing showed that proactive recommendations could fit group practices while fostering common ground (Wang et al., 2024). This suggests a natural comparison between generic facilitation and retrieval-grounded facilitation in group chat studies.

A second axis concerns operator expertise and process-level vigilance. In “LLM-Mediated Human-AI Interaction in Search and Rescue: Impact of Expertise on Attentional Allocation,” LLM guidance improved task efficiency and rewards but not total victims saved, while expertise moderated reliance: novices exhibited passive AI reliance, whereas experts maintained a “verification loop” through persistent environmental scanning (Oveisi et al., 17 Jun 2026). A plausible implication is that GRAIL-style experiments can distinguish facilitation effects on process quality from effects on end outcomes, and can test whether certain interface or intervention policies encourage verification rather than passive compliance.

Other adjacent results point to alternative group outcomes beyond decision accuracy. “From Dyads to Groups: Rethinking Emotional Support with Conversational AI” reports that group AI support increased perceived support efficacy relative to single-agent support and that connectedness mediated this effect (Hu et al., 28 Feb 2026). “Inclusive AI for Group Interactions: Predicting Gaze-Direction Behaviors in People with Intellectual and Developmental Disabilities” argues that group-aware AI systems trained on neurotypical data can fail in neurodiverse settings, introduces the MIDD dataset, and shows that fine-tuning improves performance while leaving important limitations (Huang et al., 15 Mar 2026). These findings suggest that GRAIL’s experimental logic could be extended to study connectedness, inclusion, and distributional fairness in group-AI systems, provided the relevant sensing and annotation infrastructure is added.

6. Terminological ambiguity and disambiguation

The acronym “GRAIL” is polysemous in the arXiv literature and refers to multiple unrelated systems. Besides the Group-AI Interaction Laboratory, it names “GRAIL: A Deep-Granularity Hybrid Resonance Framework for Real-Time Agent Discovery via SLM-Enhanced Indexing” (Xu, 4 May 2026), “GRAIL: Gradient-Reweighted Advantages for Reinforcement Learning with Verifiable Rewards” (Pala et al., 3 Jun 2026), “GRAIL: Generating Humanoid Loco-Manipulation from 3D Assets and Video Priors” (Xie et al., 3 Jun 2026), “GRAIL: Post-hoc Compensation by Linear Reconstruction for Compressed Networks” (Tang et al., 27 Feb 2026), and “GRAIL: Goal Recognition Alignment through Imitation Learning” (Elhadad et al., 15 Feb 2026). There is also an older automata-theory lineage around Grail.7ex and the “Grail+ Visualizer” for converting textual automata into TikZ diagrams (May et al., 2024).

For disambiguation, the Group-AI Interaction Laboratory is specifically an open-source platform for experimental research on LLM-facilitated group interaction, not an agent-discovery framework, RL algorithm, robotics pipeline, compression method, goal-recognition model, or automata visualization tool. In current arXiv usage, its identity is therefore best fixed by its full name rather than by the acronym alone (Alsobay et al., 11 Aug 2025).