AI products need to show their work — but showing everything turns transparency into yet another thing you have to manage.
How do design systems balance transparency with minimizing cognitive load?
This explores how AI products show users what the system is doing (its state, its reasoning, its AI involvement) without overwhelming them with information they have to process.
This explores how AI product design shows users what the system is doing without turning that visibility into one more thing to manage. The corpus doesn't settle the trade-off with experiments. What it offers is a set of design strategies, plus a warning that transparency can create new work of its own.
The most common strategy is layering: show a little by default and let people ask for more. IBM's Carbon design system puts all AI transparency into one reusable component. A light-inspired visual style signals that AI is involved, and an AI label opens an explainability popover only when someone clicks it Does consistent AI labeling actually reduce automation bias in practice?. Explanation becomes something you can reach rather than something you're shown every time. The style guide reports no usability testing, though, so we don't know whether users open the popover or whether the label reduces over-trust. Atlassian's Rovo guidelines make a similar move: they list 'transparent system state' next to 'flow preservation', treating visibility and uninterrupted focus as goals to balance rather than opposites How should AI design connect individual work to team knowledge?.
The less obvious finding is that transparency can create work. When a design agent showed its progress step by step, designers spotted moments where they could contribute, which is the upside. But when they edited the canvas mid-task, the agent misread or undid their changes, and they gave up on collaborating Does visible agent progress help designers know when to intervene?. Showing the process is like an invitation: it raises cognitive load and frustration if the system can't respond to people who accept it. Lam and colleagues argue the opposite way, for keeping the system's design decisions visible at every interaction so that human design judgment stays part of the loop Does visible design structure keep HCI expertise necessary as AI advances?. Taken together, the lesson is that visibility only helps when the system can act on what users do with it.
A second strategy is to split the work so users only see the part that needs them. MemoAnalyzer has users state their privacy preferences in plain terms and lets the AI carry out the actual memory edits Does splitting privacy work between users and AI reduce cognitive burden?. Users stay in charge of intent without supervising every operation, though the claim that this lowers cognitive load hasn't been tested. Larsen-Ledet gives a useful way to see the cost: much current AI forces 'mundane articulation', meaning users restate context the system should already know. Good design should cut that while keeping the 'productive articulation' that sharpens a person's own thinking Can generative AI reduce overhead work that adds no value?. On this view, transparency that just makes people re-explain things is overhead, while transparency that helps them think more clearly is worth its cost.
One idea carries over from a different field. AI models struggle with overloaded tasks too. GPT-4V failed at GUI tasks when it had to work out what icons meant and choose actions at the same time, and pre-parsing the screen into labeled elements fixed this Why do vision-only GUI agents struggle with screen interpretation?. The same principle applies to human interfaces: structure that is interpreted in advance and shown in small pieces is easier to use than raw complete exposure. Overall, the corpus offers good design ideas but little evidence that they change how users behave.
Sources 7 notes
IBM's Carbon for AI guideline assigns AI transparency to a single reusable component: light-inspired visual styling plus an AI label that triggers an explainability popover. The pattern operationalizes explainability as tiered access rather than default exposure, but the style guide reports no usability testing or evidence of actual user behavior change.
Atlassian's published design guidelines for its Rovo AI product specify that AI should connect individual work to team knowledge through context-awareness, flow preservation, transparent system state, and fluid human-AI multiplayer collaboration, operationalizing abstract collective productivity principles as concrete UI commitments.
Designers used step-by-step agent execution to spot contribution opportunities, but when they edited the canvas during agent work, the agent misinterpreted or reverted their changes, creating coordination conflicts that discouraged further attempts to collaborate.
Lam and colleagues argue that just-in-time objectives architecture makes design decisions between AI and interfaces visible, ensuring HCI expertise remains critical. By surfacing generation-and-evaluation steps at every interaction rather than hiding them, the architecture structures ongoing design judgment into the system itself.
MemoAnalyzer's design divides memory-privacy labor: users state preferences, AI executes edits. This role split is proposed as a way to reduce cognitive load and improve perceived agency, though validation metrics and conflict-resolution processes remain undescribed.
Show all 7 sources
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.
OmniParser demonstrates that GPT-4V fails when forced to simultaneously identify icon meanings and predict actions from raw screenshots. Pre-parsing screenshots into structured semantic elements with descriptions lets the model focus solely on action prediction, removing the composite-task bottleneck.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Guidelines for Human-AI Interaction
- "When to Hand Off, When to Work Together": Understanding Concurrent Human-Agent Interaction in Shared Co-Creative Workspaces
- Privacy Leakage Overshadowed by Views of AI: A Study on Human Oversight of Privacy in Language Model Agent
- What does Generative UI mean for HCI Practice?
- Design Principles for Human-Agent Interaction
- Carbon for AI
- OmniParser for Pure Vision Based GUI Agent
- Just-In-Time Objectives: A General Approach for Specialized AI Interactions