AI tools don't do exactly what you click — they guess what you meant, and the guess changes each time. How should asking for things change?
What role does generative variability play in shifting how users specify system behavior?
This explores how the unpredictability of generative AI (the same request can produce different outputs each time) changes the way people tell software what they want: from step-by-step commands to descriptions of the result they're after.
This explores how the unpredictability of generative AI changes the way people tell software what to do. The corpus's core claim is that variability doesn't just add noise to the old interaction model. It replaces it. With traditional software you specify *how*: click this, set that field, run this command. With generative systems you specify *what you intend*, and the system picks the method. Because outputs vary from run to run, this breaks a long-standing design rule that the same action should always produce the same result. One line of work proposes six principles for this new setting, including co-creation, tolerance for imperfect outputs, and helping users build a working mental picture of the system How should users control systems with unpredictable outputs?.
The surprising part is that several notes treat variability as a basic property of the medium, not a bug to engineer away. AI output changes with sampling, with prompt wording, and even with who is reading it. That makes it closer to a performance than a manufactured product, and it resists traditional quality assurance Why does AI output change with every prompt and context?. The 'context' that shapes each output (prompt, conversation history, retrieved documents, hidden state) also keeps shifting, unlike a fixed menu a user can memorize. One argument is that this calls for a new discipline of context engineering alongside traditional interface design How does AI context differ from conventional software context?. Once specification moves from commands to intent, what the system actually receives becomes a mix of your prompt and interaction constraints. One framework calls this 'Augmented Instruction' and finds twelve recurring ways tools package it, across 66 products How do interaction signals reshape what generative AI actually receives?.
The shift has a cost. If an agent behaves differently each time and doesn't show it, users can't learn what to expect. The corpus says designers have two options: reduce the variation or make it visible How does inconsistent agent behavior affect user understanding?. Specifying intent also doesn't guarantee the intent gets carried out. A benchmark of generative UI tools found that about a quarter of stated design rationales went unimplemented, rising to a third for functional requirements Do generative UI tools actually implement their stated design rationales?. The gap runs the other way too: models fill in a lot that users never said, with every one of 12 tested LLMs over-inferring user attributes in 35–49% of its claims Do large language models fabricate user attributes beyond available evidence?.
The most interesting tension is about how much specifying users should have to do at all. One argument is that generative UI helps precisely the people who *can't* describe what they want. The AI acts as the designer, so users don't need the interface vocabulary that 'vibe coding' assumes Can generative UI serve users who cannot specify interfaces?. Another argument separates useful articulation, which sharpens your thinking, from rote articulation, where you restate context the system should already know. It holds that AI only pays off at work when it removes the second kind Can generative AI reduce overhead work that adds no value?. Meanwhile, a layer of specification sits out of users' reach entirely. System prompts set by developers form a hidden chain of authority, and small edits there can shift model behavior abruptly Who actually controls what AI systems say?.
The takeaway: generative variability changes *who* specifies, not just *how*. Users trade precise control for stating intent. The system takes on design decisions. Developers keep a hidden layer that users can't see. The open design question is how to make that split visible enough that users can still steer.
Sources 10 notes
Generative AI shifts interaction to intent specification rather than method specification, creating unpredictable outputs that violate traditional consistency heuristics. Six design principles—including co-creation, imperfection tolerance, and mental model support—address this novel paradigm.
AI outputs exhibit essential mutability—they vary with sampling, prompt wording, and audience interpretation. This is not a defect but a defining feature of tokens as media, making them fundamentally different from fixed commodities and resistant to traditional quality assurance.
AI interactions operate on a substrate of constantly shifting context—prompt, history, retrieved data, hidden state—that users cannot internalize like traditional UIs. This structural mutability demands a new design discipline centered on context engineering rather than interface design.
The IAI model formalizes human-GenAI interfaces as a six-entity graph where Augmented Instruction—machine-readable encoding of prompts plus interaction constraints—is what GenAI actually receives. Applied to 66 tools, it distills twelve recurring, composable interaction paradigms and enables design comparison where prior taxonomies could not.
Stochasticity in foundation models and environment-reactive orchestration logic cause agents to behave unpredictably across runs. Users cannot develop reliable mental models unless variation is either minimized or made visible.
Show all 10 sources
A benchmark of 24 tasks across five tools found roughly 25% of design rationales go unimplemented, rising to 34% for functional requirements. Tools recognized only half the UX principles embedded in prompts.
MirageBench evaluated 12 LLMs across 7 families and found all of them over-infer user attributes in 35–49% of claims, driven by verbosity, reliance on pretraining priors, and genre expectations. Models that self-assess as over-inferring less actually over-infer more when judged independently.
Moran argues genUI's advantage is that AI acts as designer, deciding what interface to build rather than requiring users to specify one—addressing the gap where most people lack the interface literacy that vibe coding presupposes.
Larsen-Ledet argues that current AI tools impose mundane articulation—restating context the system should infer—while productive articulation sharpens how work is conceptualized. True workplace value requires design that anchors to genuine context rather than relying on surface cues.
System prompts operate as a confidential three-tier control structure where developers set foundational rules users cannot see or influence. Evidence from 2025 incidents shows small prompt edits can abruptly shift model behavior, yet users remain excluded from this layer.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Evaluating Generative AI in the Lab: Methodological Challenges and Guidelines
- Prototyping with Prompts: Emerging Approaches and Challenges in Generative AI Design for Collaborative Software Teams
- What does Generative UI mean for HCI Practice?
- Designing a Dashboard for Transparency and Control of Conversational AI
- Design Theater: Evaluating the Gap Between User-Facing Design Reasoning and Implementation in Generative UI Tools
- Interaction-Augmented Instruction: Modeling the Synergy of Prompts and Interactions in Human-GenAI Collaboration
- Who Controls the Conversation? User Perspectives On Generative AI (LLM) System Prompts
- Design Principles for Generative AI Applications