Large Language Models (LLMs) are fundamentally trained to be helpful, polite, and thorough. In standard chat interfaces, a paragraph-long explanation or a conversational filler phrase is harmless—perhaps even desirable. However, when an AI assistant operates in a real-time, high-stakes, or assistive environment, this inherent chattiness introduces a severe hidden cost: cognitive overload.
When a user is navigating a physical space, managing a medical emergency, or piloting an aircraft, their working memory is already operating near capacity. Forcing them to read or listen to verbose, unstructured text creates an information bottleneck. Recent research across aviation, medicine, and accessibility highlights a critical paradigm shift: AI assistants must optimize for minimum viable information to remain safe and effective.
The Conversational Tax on Working Memory
To understand why verbosity degrades performance, we must look at how humans process real-time instructions. In dynamic environments, every additional word generated by an AI acts as a cognitive tax, requiring time to parse, filter, and translate into action.
In an early exploration of this dynamic within virtual environments, Koller et al. demonstrated that overly verbose instructions directly hinder physical execution. When a system generates dense text, users spend critical seconds processing the information, which slows down their physical movements and lowers overall task success rates.
This friction is amplified in voice-based or eyes-free interfaces. As Polifroni and Walker observed in dialogue systems, there is a strict trade-off between providing comprehensive data summaries and maintaining manageable response lengths. In audio channels, users cannot skim past conversational fluff; they must listen to it sequentially, leading to rapid cognitive fatigue.
When Verbosity Becomes Dangerous
In high-pressure operational environments, the cost of a chatty AI scales from mere inefficiency to catastrophic failure.
Consider single-pilot aviation operations. Fan Li et al. developed a virtual co-pilot to retrieve emergency procedures from flight manuals. They noted that while LLMs are technically accurate at retrieval, their tendency to output conversational fillers and paragraph-long explanations distracts pilots from primary flight tasks when every second matters.
The danger of this design flaw was explicitly demonstrated by Arnold et al.. By testing standard, naive LLM interfaces against specialized, concise agents in flight simulations, the authors found that paragraph-length explanations during a crisis actively cause cognitive overload. Instead of supporting the pilot, the verbose AI impeded rapid decision-making, ultimately causing catastrophic errors and simulated fatal accidents.
[Verbose AI Output]
"To handle the engine failure, you should first check the fuel selector valve, then ensure the mixture is rich, and next..."
--> High Cognitive Load: User must parse syntax and sequence under stress.
[Concise AI Output]
"1. Fuel Selector: CHECK.
2. Mixture: RICH."
--> Low Cognitive Load: Immediate, actionable execution.
This pattern repeats across other high-stakes domains:
- Acute Medical Care: In a randomized controlled trial of a generative AI tool for gastrointestinal bleeding, Chung et al. found that clinicians overwhelmingly reject conversational depth. In fast-paced emergency rooms, long-form natural language explanations cause information overload; experts heavily prefer concise, button-based, or structured summaries that deliver immediate insights.
- Hazardous Material Response: Evaluating LLMs on chemical emergency response, Surana et al. discovered that models frequently inject unprompted peripheral advice. In a hazmat crisis, these dense text blocks delay first responders from taking critical, time-sensitive actions.
The Information Bottleneck Solution
To solve this, researchers are shifting from generic text generation to frameworks that explicitly model the human information bottleneck.
Chen et al. formalized this approach with VIBE, an annotation-free evaluation framework for video summarization. VIBE scores AI outputs based on a strict balance of task relevance and brevity. The authors proved that minimizing summary length prevents cognitive fatigue, slashing human decision-making time by over 75% while simultaneously improving task accuracy. By treating text length as a constraint rather than a feature, the system ensures the user receives only the information necessary for the immediate decision.
Cognitive Load in Assistive and Live Workflows
The necessity of concise communication extends beyond life-or-death emergencies into daily human-AI collaboration, where verbosity introduces social and emotional friction.
Human-in-the-Loop Verification
In live customer support environments, Lu et al. deployed generative models to suggest responses for human agents. They discovered that agents consistently prefer shorter, more concise suggestions. The reason is simple: an agent must read, verify, and potentially edit a suggestion before sending it. Minimizing the text length minimizes the cognitive load required to validate the AI's work during a live interaction. This operational delay is also noted by Mo et al. in behavioral health settings, where long-form AI responses delay real-time decision-making during live peer support.
Wearable and Spatial Assistive Tech
For wearable devices, verbose outputs create social disruption. Wazeer Zulfikar developed Memoro, an audio-based memory assistant that utilizes a "Queryless Mode" to cut device output length by 85%. By eliminating full-sentence responses and conversational fillers, the system delivers core information instantly, minimizing the duration of social distraction during live conversations.
Similarly, in physical navigation for the visually impaired, Han et al. demonstrated that standard vision-language models produce chatty, overly descriptive spatial text. In a crosswalk or busy intersection, this descriptive verbosity actively obscures critical safety instructions, overwhelming the user when they need clear, immediate spatial guidance.
High-Stress Psychological States
When users experience acute emotional distress, abstract jargon and long lists of options become counterproductive. Çarik et al. explored how autistic individuals interact with AI to cope with negative self-talk. The authors found that standard, wordy AI responses exceed the user's immediate cognitive capacity during distress. To be effective, the AI must shift from complex therapeutic explanations to concise, direct, and actionable guidance. This matches findings by Zhuochun Li et al. in dementia caregiving, where dense, redundant AI medical advice introduces confusion rather than clarity for stressed users.
Designing for the Bottleneck
The consensus across these diverse domains is clear: conversational fluency is not user experience. For an AI assistant to be truly cooperative in real-time settings, it must treat human attention and working memory as strictly finite resources. Designing effective real-time interfaces requires moving away from open-ended text generation and toward highly structured, constrained, and dense information delivery.
Key Foundational Work
Report on the Second Second Challenge on Generating Instructions in Virtual Environments (GIVE-2.5) (2011) — contributed 2 concept elements to this direction.
Intensional Summaries as Cooperative Responses in Dialogue: Automation and Evaluation (2008) — contributed 1 concept elements to this direction.