PAPER
The Generative Pause: Designing UI States for Optical Honesty
Published: March 10, 2026
Authors: Erin Lentz — Executive Director of Design; Goran Paun — Principal, Creative Director
Table of Contents
Abstract:
Part II of The Generative Pause moves beyond principle into practice. While Part I1 argued that velocity‑first digital systems compress human judgment, Part II defines how to operationalize human oversight through design itself. Its central concept, Optical Honesty, asserts that an interface should visually reflect the real certainty, consequence, and review status of an AI‑generated output. Oversight should not appear as an external checkpoint but be embedded into the interaction surface, allowing the environment, not modal friction, to signal when deeper human review is required. Judgment becomes an integral step in the work, not an afterthought. To ensure the principle is actionable, we specify state‑based UI standards, componentized variants, and governance checks that render Optical Honesty measurable, auditable, and repeatable across products.
1. Optical Honesty
One of the central risks in generative systems is not only inaccuracy, but presentation. Outputs often arrive with the visual confidence of finished work, even when their underlying reliability remains uneven. Formatting, hierarchy, spacing, and polish can create a false sense of resolution. The user is not only evaluating content. They are also responding to the interface’s implied confidence.
Optical Honesty begins with a simple premise—the interface should look as certain as the system actually is.
Optical Honesty reframes this problem:
- The interface should be only as “finished” as the underlying certainty merits.
- Provisional material should feel provisional—legible as draft, incomplete, or requiring human review.
- Visual solidity should grow as human review becomes substantive, not by default.
That does not mean making uncertain outputs undesigned or intentionally awkward. It means resisting the design habit of giving provisional material the same composure as verified material. A suggestion should read as a suggestion. A model-generated summary should feel reviewable, not settled. A high-consequence recommendation should carry visible weight before the user approves it.
This is not merely a question of styling. It is a question of epistemic legibility. The interface should help the user sense whether they are looking at a draft, a recommendation, a likely conclusion, or a verified outcome. When every state looks equally complete, the system encourages premature trust.
A more responsible pattern is to let the interface gain visual solidity as human review becomes more substantive. The design does not begin with finality. It earns it.
1.1 Operational Standard for Optical Honesty
To move from principle to practice, Optical Honesty must be rendered as a system, not a style. We use three scaffolds that make the concept repeatable across teams and platforms:
1.1.1 Certainty → UI Mapping
Interfaces render different visual states based on model metadata (e.g., confidence bands, uncertainty flags, or provenance completeness). A suggestion is not styled like a conclusion, and a conclusion is not styled like a verified outcome. This mapping is deterministic: given the same certainty inputs, teams produce the same state.
1.1.2 Componentized Variants & Tokens
Design systems expose Optical Honesty via named component variants—Draft, Reviewable, Resolved—and certainty tokens (e.g., certainty.low, certainty.medium, certainty.high). Designers select variants; they do not “interpret” tentativeness.
1.1.3 Semantic Requirements
Each state must communicate a specific semantic:
- Draft communicates incompleteness and prompts inspection.
- Reviewable communicates inspectability with visible sources and rationale.
- Resolved communicates finality with accountability (including provenance or human sign‑off where appropriate).
1.2 Optical Honesty: Minimum Visual Standards
To ensure epistemic legibility, each state must meet objective, checkable criteria.
Draft (Low Certainty / Low Consequence) — must:
- Use reduced visual solidity (e.g., lighter contrast, softer elevation, subdued chroma).
- Avoid fully “finished” hierarchy (e.g., lighter heading weight, placeholder affordances).
- Include an uncertainty signifier (badge, label, or hint) and a one‑click path to sources.
Reviewable (Moderate Certainty and/or Consequence) — must:
- Reveal source→output relationships inline (expandable citations, side‑by‑side inspection).
- Increase local density only around the claim and its evidence (“elastic density”).
- Provide structured actions for acceptance, annotation, or revision.
Resolved (High Certainty and/or Human‑Verified) — must:
- Present full typographic and layout composure (final hierarchy, stable alignment).
- Display provenance or review trail appropriate to the risk class.
- Offer reversible actions where feasible; require elevated confirmation where irreversible.
These requirements are objective checks, not aesthetic preferences.
Accessibility Note: Optical Honesty variants must meet WCAG contrast thresholds and avoid conveying state by color alone; any motion‑based emphasis should offer reduced‑motion equivalents.
Goran Paun (left) and Erin Lentz (right) in a working session reviewing the VERSIONS paper, aligning content and structure around human-centered design principles.
2. From Pause to Calibration
Optical Honesty supplies the states; calibration supplies the when. The system selects Draft, Reviewable, or Resolved based on two inputs: (1) model certainty and (2) decision consequence. This pairing removes designer subjectivity and prevents “pretty by default” patterns at precisely the wrong moments. If Part I established the value of slowing down at consequential moments, Part II asks a harder question: what should that slowing down actually feel like?
The answer is calibration, not interruption. Generic confirmations (“click to continue”) create ritual, not responsibility; users learn to dismiss them reflexively. Instead, the system should scale its demands proportionally to (1) the model’s certainty and (2) the risk or irreversibility of the decision. Low‑risk, high‑certainty moments may require very little. High‑risk or low‑confidence moments should visibly thicken the interaction.
The problem with many compliance-oriented systems is that they rely on generic acts of confirmation. Check this box. Click continue. Confirm that you reviewed the material. These gestures create the appearance of oversight without proving that attention ever occurred. Users quickly learn their shape and move through them automatically. Repetition drains meaning from the act of review.
A more effective model is to design the pause into the task itself. The system should vary its demands based on two factors: how uncertain the model is, and how consequential the decision will be if the output is accepted. Low-risk, low-uncertainty situations may require very little. High-risk or low-confidence situations should visibly thicken the interaction.
That is where calibration becomes more useful than friction. The goal is not to block the user. The goal is to make the system proportionate.
- Not every AI-assisted action requires the same kind of pause. The intensity of review should change with the certainty of the system and the consequence of the decision.
At a practical level, this creates three distinct calibration tiers:
| Calibration Tier | Contextual Trigger | Visual / Interaction Implementation |
| Tier 1: Ambient | Low-risk, high-certainty action | Invisible flow. The interface remains light, continuous, and unobtrusive. No explicit pause is introduced because the consequence is low and the system confidence is high. |
| Tier 2: Guided | Medium-risk, moderate-certainty action | Elastic density. The interface expands around the decision point to surface source-to-output relationships, supporting evidence, and review cues without breaking workflow continuity. |
| Tier 3: Deliberate | High-risk, low-certainty, or irreversible action | Optical shift. The workspace visibly changes state to slow habitual action, increase inspectability, and require meaningful human review before authorization. |
3. Invisible Validation
The most credible forms of oversight are often the least theatrical. They do not announce themselves as control points. They are embedded into the way work is done.
This is where Invisible Validation becomes useful. Rather than asking the user to perform a symbolic act of responsibility, the interface can be designed to register whether they are actually engaging the underlying evidence.
One pattern is selective ghosting. Critical values, source-linked claims, or high-consequence fields do not arrive in fully resolved form. They remain visually light, partially recessed, or incomplete until the user moves through the relevant source material and returns to the generated conclusion. The point is not concealment for its own sake. The point is to ensure that review leaves a trace in behavior.
Another pattern is what might be called elastic density. As the stakes of a decision rise, the interface becomes more information-dense in the places that matter. Supporting evidence, provenance, exception logic, or conflicting signals become easier to inspect and harder to bypass. The workspace does not simply grow more crowded. It becomes more concentrated around what deserves human attention. In low-stakes moments, the system stays light. In high-stakes moments, it gains informational gravity.
The pause then stops feeling like an added task. It becomes a shift in the character of the environment. The interface is quietly telling the user: this moment requires more of you.
Auditability Note. Invisible Validation patterns are paired with explicit telemetry: which sources were opened, which rationale panes were expanded, and whether critical fields were reconstructed. We log the fact of engagement (not its content) to preserve privacy while producing an auditable trail that a review actually occurred.
4. Designing for Attention
Much of contemporary enterprise UX still treats oversight as ceremony. A box is checked. A disclosure is acknowledged. A warning modal appears. The flow resumes. These patterns may satisfy governance on paper, but they often fail in practice because they do not measure attention. They measure procedural completion.
Designing for attention means asking a more demanding question: what interaction would make it difficult to proceed without actually looking?
Examples:
- Comparison: requiring side‑by‑side inspection of source and summary.
- Reconstruction: requiring the user to manually input a critical value after reviewing evidence.
- Sequence: withholding polished summaries until the supporting logic has been visited.
The aim: preserve authorship and prevent users from becoming mere approvers of machine‑generated conclusions.
These are not obstacles inserted to slow people down indiscriminately. They are ways of preserving authorship at the exact points where automation can create the illusion of authorship without the substance of it.
The distinction matters. In a healthy human-AI system, the person should not be reduced to a passive approver of machine-shaped conclusions. They should remain the accountable interpreter.
Attention Design Heuristics (Pass/Fail):
- The user cannot finalize without encountering the supporting logic (viewed or expanded).
- At least one inspection action (compare, reconstruct, or sequence) precedes authorization in Tier 3.
- Any irreversible action requires the Resolved state and a visible review trail.
These heuristics convert “attention” from aspiration into release criteria.
5. Measuring Whether the Pause Is Working
The effectiveness of this approach cannot be judged only through speed, completion, or funnel efficiency. Those metrics tell us whether a flow is fast. They do not tell us whether judgment occurred.
Leading Indicators (Process‑Level):
- Increased revision rate when uncertainty is surfaced (without a proportional rise in rework).
- Higher engagement with evidence panes at Tier 2/3 relative to baseline.
- Reduction in “blind approvals” (approvals with zero evidence interaction).
Lagging Indicators (Outcome‑Level):
- Fewer downstream corrections, escalations, or regulatory exceptions.
- Lower incidence of post‑hoc reversals on high‑stakes actions.
- Improved human‑recorded confidence in decisions at sign‑off.
These metrics reward judgment quality, not just flow velocity.
These are not traditional growth metrics, but they are closer to the real value of oversight. A system that prevents avoidable errors, reduces false confidence, and protects people from unexamined approvals is performing well even if it introduces a small amount of additional time.
The measure of success in these moments is not raw speed, but whether the interface helped the user reach a more considered decision.
These metrics privilege judgment quality over raw velocity, aligning product success with reduced downstream corrections and more deliberate decisions.
5.1 Evaluation & Governance of Optical Honesty
To keep Optical Honesty consistent at scale, we introduce three controls:
- Design‑System Enforcement — Component variants and certainty tokens are the only approved means to render state. Product teams consume, not reinvent.
- Linting & QA — Pre‑release checks verify that UI state matches model metadata and consequence class; high‑risk surfaces require the Reviewable → Resolved progression.
- Periodic Calibration Reviews — Quarterly audits sample real decisions to confirm that Optical Honesty states were accurate and that inspection patterns were engaged before authorization.
Together, these controls shift Optical Honesty from a subjective aesthetic to a repeatable standard.
6. Conclusion
Some will argue that Optical Honesty is too subjective to standardize. Our approach removes subjectivity by binding interface states to model certainty and decision consequence, exposing these states as governed component variants, and testing them against objective heuristics. In practice, teams do not “make it look tentative”; they select the appropriate state—and the system enforces the rest.
Part I argued that digital systems need moments that preserve human judgment rather than compress it. Part II extends that argument by showing that those moments do not need to feel like external interruptions. They can be built into the visual and interactive logic of the interface itself.
Optical Honesty is one way to do that. It asks interfaces to stop overstating certainty. It asks review states to feel reviewable. It asks consequential decisions to carry visible weight. And it asks designers to think less about how to accelerate approval and more about how to support attention.
In that sense, the next frontier of responsible interface design may not be frictionless flow at all. It may be the careful shaping of environments where people can still tell the difference between a suggestion, a conclusion, and a decision they are truly prepared to own.
Appendix — Implementation Audit
The following checklist is designed for design and product teams evaluating an existing high-stakes workflow against the Optical Honesty standards in this paper. It is not a compliance instrument. It is a diagnostic. Use it to locate where your current system diverges from the framework and where the highest-priority design work lies.
Work through it against one specific workflow — the authorization of an AI-generated output that carries meaningful organizational, financial, or compliance consequence.
State Mapping
- Does every AI-generated output in this workflow appear in one of three named states: Draft, Reviewable, or Resolved?
- Is state assignment driven by model metadata, or is it left to designer interpretation at implementation time?
- Can a designer on your team articulate the visual difference between a Draft and a Reviewable output without consulting documentation?
Calibration
- Does your workflow contain at least one Tier 3 moment — a decision point where the interface visibly changes character before authorization?
- Is that change tied to consequence and certainty inputs, or is it applied uniformly across all authorization steps?
- Could a user reach final authorization in a Tier 3 flow without encountering any supporting evidence?
Design System
- Does your design system expose named component variants for each Optical Honesty state?
- Are certainty tokens defined and available to product teams, or is visual tentativeness being expressed through one-off styling decisions?
- Is there a pre-release check that verifies UI state against model metadata before high-risk surfaces ship?
Telemetry
- Do you currently have any signal on blind approvals — authorizations completed with zero evidence interaction?
- Can you distinguish, in your data, between a user who opened a source pane and a user who did not before approving?
- Is evidence engagement tracked anywhere in your reporting, or only completion and velocity?
A score of zero on the telemetry section is the most common finding. Most organizations cannot currently answer those questions. That is the starting point, not a failure condition.
If this audit surfaces patterns — either gaps or implemented solutions — we are building a practitioner dataset around Optical Honesty deployment. Share your findings.