Transparency in AI Design: When to Show Users What the AI Is Doing

AI agents promise to handle complex tasks autonomously - booking trips, managing calendars, analyzing data. But when you hand off a task, what happens next? For 30 seconds (or 30 minutes), the system vanishes. Then it returns with a result. And every user asks the same thing: did it work, or did it hallucinate?

‍

This anxiety pushes designers toward two extremes. The Black Box - hide everything, keep it simple. Or the Data Dump - stream every log line and API call to the user. Neither works.

‍

The Black Box leaves users feeling powerless. The Data Dump creates notification blindness, destroying the efficiency the agent was supposed to provide. At Orizon, designing AI-powered products across fintech, healthcare, and SaaS has taught us one thing clearly: transparency isn't binary. It's about revealing the right information at the right moment.

‍

2026 marks a turning point where the emphasis has shifted from proving AI can work to proving it can be trusted. That shift changes what UX designers need to build.

‍

The AI Transparency Spectrum

‍

Most AI products fall somewhere on this axis:

• Black Box - AI returns a result with no explanation. Works for low-stakes, fast, familiar tasks (autocomplete, spell-check). Fails the moment stakes go up.
• Minimal Transparency - High-level status like "Thinking…" Works for quick tasks. Breaks down after 30 seconds when users have no idea if the AI is close or completely lost.
• Selective Transparency - AI reveals decision points that matter, hides routine operations. Claude's Artifacts shows when it's creating a document without surfacing syntax parsing. This is the sweet spot.
• Data Dump - Every action, API call, and log streamed to the user. Fine for developers debugging. A disaster for anyone trying to get work done.

‍

The goal is Selective Transparency. Transparency in AI depends on three elements: visibility (what the AI is doing), explainability (why decisions are made), and accountability (allowing users to understand and influence outcomes). Show what matters at each layer. Hide what doesn't.

‍

🤖 Building an AI product and not sure how much to show users? Let's design the transparency layer that builds trust →

‍

‍

How to Find Your Transparency Moments: The Decision Node Audit

‍

The most useful framework we've applied is the Decision Node Audit - mapping your AI agent's workflow to pinpoint exactly when to surface information.

‍

Take a travel booking agent. It parses your request, searches databases, filters by price and duration, applies your preferences, selects an option, verifies availability, books, and confirms. Nine steps. Most of them are routine operations users don't need to see.

‍

The ones worth revealing:

• Filtering criteria - what rules is the AI applying?
• Preference application - did it correctly understand aisle seat, no layovers?
• Final selection - why this flight over that one?

‍

For each decision point, ask three questions: Is this high-stakes? Can the AI get this wrong? Does the user need control here? If the answer to any of them is yes, reveal it.

‍

The Right Transparency Pattern for the Right Moment

‍

Once you've identified your transparency moments, pattern choice matters:

‍

Intent Preview - before an irreversible action. "I'm about to book this $1,200 flight. Confirm?" Prevents costly mistakes. Non-negotiable for any action that can't be undone.

‍

Status Update - during long-running tasks. "Checked 127 flights. Filtering by your preferences…" Reduces anxiety without overwhelming. Progress beats silence every time.

‍

Decision Explanation - after a non-obvious choice. "I chose Flight A over B because it matches your preference for no layovers." Builds trust. Helps users course-correct if the AI misunderstood something.

‍

Autonomy Dial - for users who want variable control. Slack's AI reply settings (Low / Medium / High autonomy) let users tune transparency to their own comfort level. The AI features users trust most are ones where they feel in control - not because the AI is passive, but because the user can redirect it at any moment.

‍

‍

What Real Products Get Right (and Wrong)

‍

Claude's Artifacts is Selective Transparency done well. When generating code or documents, it shows a live output panel - you see the document being written without being exposed to tokenization or model inference. Reveals the outcome, hides the process. Trust builds naturally.

‍

ChatGPT's "Thinking…" indicator works for tasks under five seconds. For anything longer, it breaks down - no progress signal, no way to intervene, no sense of whether you're 10% or 90% through. Minimal transparency has a ceiling.

‍

Perplexity's live source citations take a different approach - showing where each piece of information came from as the answer generates. It doesn't explain the AI's reasoning, but it builds trust through evidence. For accuracy-sensitive tasks, showing your work is enough.

‍

Humane AI Pin is the cautionary tale. Constant notifications - every action, every API call. Users hit notification blindness within hours. Urgent alerts buried in routine logs. More transparency made the product less usable, not more.

‍

Practical Rules for Getting This Right

‍

Based on our AI UX design work with clients across SaaS, fintech, and healthcare:

• Default to minimal, scale up for high-stakes - don't front-load information. Reveal on demand.
• Always preview irreversible actions - sending emails, making purchases, deleting data. Always.
• Explain surprises - if the AI skips a step or makes an unexpected choice, say why. Surprises erode trust fast.
• Use progressive disclosure - "Analyzing 500 documents" by default, "Show details" for users who want to dig in.
• Test with real users - what feels transparent to your team may feel opaque (or overwhelming) to users. Watch where confusion and anxiety appear, then adjust.

‍

The test is simple: if a user wants to know why the AI did something, can they find out in under 10 seconds? If yes, the transparency UX is working.

‍

‍

Where This Is Going

‍

The next evolution is adaptive transparency - systems that adjust what they reveal based on context. More detail when confidence is low. Less friction during routine tasks. Personalized over time to each user's preferences.

‍

Context-aware modes allow users to align AI behavior with the intent, stakes, and nature of the task at hand. Transparency should follow the same principle - show what helps users reach their goal, nothing more.

‍

Great AI UX isn't about making the AI invisible. It's about making it understandable at exactly the moments that matter. That's what we build at Orizon.

‍

Ready to design an AI product users actually trust? Contact us today 🚀

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

FAQs

‍

What is the black box problem in AI design?

The black box problem occurs when AI systems provide results without explaining how they reached those conclusions. This creates user anxiety - especially for high-stakes decisions like financial transactions or medical advice. Users have no way to know if the AI hallucinated, made an error, or followed correct reasoning.

‍

What is a data dump in AI UX?

A data dump happens when AI systems stream every log, API call, and internal operation to users in real time. While it maximises transparency in theory, it creates notification blindness - users ignore the constant flow until something breaks, then lack the context to fix it. More transparency made the product less usable.

‍

When should AI show users what it's doing?

AI should reveal its decision-making during irreversible actions (sending email, making purchases), high-stakes choices (financial, medical, legal), unexpected behaviour (skipping steps, changing plans), and long-running tasks where silence creates anxiety. Routine operations can and should stay hidden.

‍

What is a decision node audit?

A decision node audit is a design process where teams map AI workflows and identify exactly which moments require user transparency. It involves four steps: mapping the workflow, identifying decision points, assessing user impact at each one, and choosing the appropriate transparency pattern - intent preview, status update, or decision explanation.

‍

How does Claude's Artifacts feature demonstrate good AI transparency?

Claude's Artifacts shows users the output being created - code, documents, diagrams - in a live panel, while hiding low-level operations like syntax parsing. This selective transparency reveals what matters to users (the result) without overwhelming them with implementation details. It's the sweet spot in practice.

‍

What is an autonomy dial in AI design?

An autonomy dial is a UI control that lets users adjust how much the AI acts independently versus how much it asks for permission. Slack's AI offers Low, Medium, and High autonomy settings - low means the AI previews every action before taking it, high means it works with minimal interruption. It puts transparency control in the user's hands.

‍

Why did the Humane AI Pin's transparency approach fail?

The Humane AI Pin streamed constant notifications for every action and API call - a textbook data dump. Users experienced notification blindness within hours because urgent alerts were buried alongside routine logs. The excessive transparency defeated the product's core promise of reducing friction.

‍

What is progressive disclosure in AI interfaces?

Progressive disclosure means showing summary information by default, with detail available on demand. For example: "Analysing 500 documents" upfront, with a "Show details" option for users who want to see the full list. It prevents information overload while keeping everything accessible to users who want it.

‍

How does Perplexity build trust through transparency?

Perplexity displays live source citations as it generates answers, showing where each piece of information came from. It doesn't explain the AI's reasoning process - but it builds trust through verifiable evidence. For accuracy-sensitive tasks, showing your sources is often enough to reduce hallucination concerns.

‍

What is outcome-oriented design for AI?

Outcome-oriented design - a framework from Nielsen Norman Group - focuses on what users want to achieve rather than how the AI achieves it. Applied to transparency, it means surfacing information that helps users reach their goal (decision points they need to control) and hiding implementation details that don't affect the outcome.

‍

‍Header image: ‍Photo by Vera Erm / Dribbble – “AI Design” (Shot #27057725).