Design Guidance Towards Addressing Over-Reliance on AI in Sensemaking

This paper proposes preliminary design principles for GenAI-augmented Group Awareness Tools that utilize implicit visualizations to trigger cognitive conflict and autonomous elaboration, thereby mitigating over-reliance on AI and fostering independent sensemaking in collaborative work and learning.

The Gist

Here is an explanation of the paper, translated into simple, everyday language using analogies to help visualize the concepts.

The Problem: The "Over-Attentive GPS"

Imagine you are driving with a group of friends to a new destination. You have a GPS (Generative AI) that is too helpful. Instead of just showing you the map, the GPS starts shouting out every single turn, telling you exactly how to merge, and even critiquing your driving style.

"Turn left now! You're driving too slow! Sarah, you should be navigating!"

At first, this feels great. But soon, you and your friends stop looking at the road, stop talking to each other about the route, and stop thinking for yourselves. You just blindly follow the GPS. If the GPS glitches, you're lost because you never learned how to navigate.

The Paper's Point: In group work and learning, AI is becoming like this over-attentive GPS. It gives "Explicit Instructions" (step-by-step orders), which makes groups lazy and stops them from figuring things out on their own. This is called Over-Reliance.

The Solution: The "Mirror on the Wall"

The authors suggest a different approach using something called Group Awareness Tools (GATs).

Instead of a GPS that tells you what to do, imagine a magic mirror that simply shows you a reflection of your group's behavior.

• The mirror doesn't say, "You are arguing too much."
• It doesn't say, "Sarah is doing all the work."
• It just shows a visual chart where you can see that Sarah is doing 80% of the talking while others are silent.

When you see this, you don't get an order. You get a Cognitive Conflict (a little mental "huh?"). You and your friends naturally start talking: "Wait, why is Sarah doing everything? Let's fix this." This is Implicit Guidance. The tool doesn't tell you the answer; it shows you the problem so you come up with the answer.

The New Idea: Mixing AI with the Mirror

The paper asks: Can we use the smart AI to make this "Magic Mirror" even better without turning it back into a bossy GPS?

They found three golden rules for doing this:

1. Know When to Use the "Smart Brain" vs. the "Calculator"

• The Calculator (Old School): If you just need to count things (e.g., "Who typed the most words?"), a simple computer program is fine. It's fast and accurate.
• The Smart Brain (GenAI): If you need to understand feelings or ideas (e.g., "Did the group actually understand the topic, or were they just pretending?"), you need the AI.
• The Rule: Don't use the AI to give orders. Use it to read the "vibe" of the conversation and show that to the group.

2. The "Color-Code" Trick (How to Show the Data)

The authors suggest a clever way to display the AI's insights so it doesn't look like a command.

• The Old Way: A radar chart showing what the group says they know.
• The New Way: Keep that same chart, but add a background color based on what the AI heard in the actual conversation.
  • Dark Background: "You said you knew this, and your conversation proves you do." (Good alignment).
  • Light Background: "You said you knew this, but your conversation showed you were confused." (A discrepancy).
• Why it works: The group sees the light color and thinks, "Wait, why is this part light? Did we miss something?" They have to investigate. The AI didn't say "You are wrong"; it just highlighted a gap for them to explore.

3. The "Hover to Investigate" Button

When the group sees a gap (like the light background), they shouldn't just accept the AI's word as truth. They need to dig deeper.

• Imagine hovering your mouse over that light spot.
• Pop-up: The AI shows you a specific quote from your conversation that proves the confusion, along with a confidence score.
• The Result: The group reads the quote and debates: "Do we agree with this? Is the AI right? Let's talk about this specific point."
• This turns the AI from a "Teacher" into a "Research Assistant" that provides evidence for the group to discuss.

The Big Picture

The paper is a warning and a guide. It warns us that if AI just gives us answers, we lose our ability to think together. But, if we design AI to act like a mirror that highlights differences and gaps, it can spark deep, meaningful conversations.

The Goal: We want AI to be the spark that starts the fire of discussion, not the firefighter that puts out the thinking process. By using "Implicit Guidance," we help groups learn to solve their own problems, even when the AI is watching.

Technical Summary

Based on the provided extended abstract, here is a detailed technical summary of the paper "Design Guidance Towards Addressing Over-Reliance on AI in Sensemaking."

1. Problem Statement

The paper addresses a critical issue in collaborative work and learning: the over-reliance on Generative AI (GenAI) which erodes autonomous sensemaking.

• The Core Issue: Poorly designed GenAI systems often provide explicit instruction (highly structured, step-by-step directives on what to do next). This leads groups to passively follow AI suggestions, bypassing the cognitive processes necessary for developing their own interpretations and strategies.
• The Contrast: Effective collaborative sensemaking requires implicit guidance. Traditional Group Awareness Tools (GATs) achieve this by visualizing data to reveal differences between members, creating cognitive conflict. This conflict triggers autonomous elaboration and discussion, forcing the group to resolve discrepancies themselves rather than being told the solution.
• The Gap: While GenAI offers the potential to analyze unstructured data (e.g., discussion transcripts) that traditional GATs cannot, there is a risk that integrating GenAI will simply replace implicit guidance with explicit instruction, further diminishing human agency.

2. Methodology

The authors employed a qualitative literature review and design analysis approach to derive preliminary design principles:

• Data Sources: An initial literature search was conducted across the ACM Digital Library, IEEE Xplore, and Scopus, complemented by backward snowballing (reviewing references of relevant papers).
• Scope: The analysis focused on existing GAT systems to understand how they generate, present, and support the exploration of group awareness information.
• Synthesis: The authors analyzed how GenAI could be integrated into these systems without disrupting the mechanism of implicit guidance. They synthesized findings to identify consistent design considerations for GenAI-augmented GATs.

3. Key Contributions & Design Principles

The paper proposes three primary design considerations to integrate GenAI into GATs while preserving implicit guidance:

A. Strategic Deployment (Hybrid Architecture)

• Principle: GenAI should not be used for all data processing.
• Implementation:
  • Rule-based approaches should handle structured, quantitative data (e.g., contribution counts, participation patterns) where objectivity is high.
  • GenAI should be reserved for the qualitative interpretation of unstructured content (e.g., discussion transcripts, document revisions).
  • Goal: Use GenAI to assess semantic nuances, such as whether members are building on ideas, identifying substantive disagreements, or detecting gaps in shared understanding that resist quantification.

B. Presentation Strategy (Augmentation, Not Replacement)

• Principle: GenAI-generated insights must serve as secondary visual encodings that augment, rather than replace, primary quantitative representations.
• Implementation:
  • The system should visualize discrepancies between group self-reports and actual AI-analyzed behavior.
  • Example (Figure 1): A radar chart showing self-reported knowledge levels (primary) is overlaid with background color intensities representing GenAI's analysis of actual discussion alignment (secondary).
  • Effect: Darker backgrounds indicate alignment; lighter backgrounds reveal discrepancies. This creates cognitive conflict by showing the group that a difference exists without explicitly telling them what it means, prompting them to interpret the data.

C. Interaction Techniques (Active Investigation)

• Principle: Groups must have affordances to investigate the evidence behind AI assessments rather than passively accepting them.
• Implementation:
  • Hover-for-details interactions (Figure 2) allow users to inspect the underlying evidence.
  • When a user hovers over a discrepancy, the system reveals the AI's assessment, a confidence score, and specific evidence (e.g., direct quotes from the discussion).
  • Goal: This enables the group to judge the validity of the AI's interpretation, decide if it warrants discussion, and engage in autonomous elaboration.

4. Results (Preliminary Findings)

While the paper is a design proposal rather than an empirical evaluation, the "results" are the three synthesized design considerations that define a new design space:

1. Hybrid Processing: Combining rule-based logic for metrics with GenAI for semantic analysis.
2. Visual Augmentation: Using GenAI to create "richer" cognitive conflict by revealing hidden semantic differences via secondary visual layers.
3. Evidence-Based Interaction: Designing interfaces that require users to engage with the raw evidence (quotes, data points) to validate AI insights, preventing passive acceptance.

5. Significance

• Theoretical Impact: The paper reframes the role of GenAI in collaborative learning from an "instructor" to an "awareness enabler." It argues that the value of GenAI in sensemaking lies not in providing answers, but in enriching the differences revealed to the group.
• Practical Application: It offers a concrete roadmap for HCI designers and developers to build GenAI-augmented GATs that mitigate the risk of automation bias and over-reliance.
• Future Direction: The proposed framework suggests that preserving implicit guidance is essential for maintaining human agency. The authors invite further discussion on how these mechanisms can be transferred to sensemaking contexts beyond collaborative work and learning.

Conclusion: The paper posits that for GenAI to support rather than hinder autonomous sensemaking, it must be integrated as a tool for revealing complexity and conflict (via unstructured data analysis and interactive visualization) rather than a tool for resolving complexity (via explicit instruction).