Visualization for Human-Centered AI Tools

• Human-centered AI emphasizes AI that enhances rather than replaces people and addresses human values and agency.Source: §1 · p. 1
• The paper proposes interactive visualization as a key enabling technology for human-centered AI tools.The approach is called visualization-enabled HCAI tools.Source: §1 · p. 1
• The authors refine HCAI tool characteristics using literature analysis and empirical feedback from HCI, AI, and visualization experts.Source: §1 · p. 1
• They review visualization-enabled AI tools and extract design guidelines for visual representations that support HCAI principles.Source: §1 · p. 2

• Intelligence augmentation treats computers as tools that enhance human capabilities, whereas AI traditionally pursued autonomous entities simulating human intelligence.Source: §2 · p. 2
• AI and intelligence augmentation have converged under HCAI in response to ethical, fairness, transparency, accountability, and human-capability concerns.Source: §2 · p. 2
• Visualization has supported both intelligent tools and model development, including neural-network displays, visual debugging, human-agency applications, and interpretability.Source: §2 · p. 2
• Despite these contributions, the role of visualization in HCAI remains vague and poorly understood, motivating this paper’s focus.Source: §2 · p. 3

Method

• The definition is developed through a mixed-method process combining a literature survey with interviews involving HCI, AI, and visualization experts.Source: §3 · p. 3, §3.1 · pp. 3–4

Capabilities

• The four capabilities distinguish strengthening existing abilities, adding new capabilities, increasing user agency, and improving existing abilities or artifacts.Source: §3.3 · pp. 4–5
• Capability boundaries can be difficult to separate, especially between amplify and enhance.Source: §3.3 · p. 5

Scope

• Not every tool addresses every concern, concern categories can overlap, and broader societal considerations remain outside the paper’s interactive-tool-design scope.Source: §3.4 · pp. 5–6, Definition: Human-Centered AI Tools · p. 7

Definition

• HCAI tools are interactive software systems integrating AI to amplify, augment, empower, or enhance human capabilities while addressing specified human concerns.Source: Definition: Human-Centered AI Tools · p. 6

Evaluation and scope

• Experts agreed the definition can analyze existing and future HCAI tools, while noting that its scope covers the tooling aspect of human-centered AI.Source: Definition: Human-Centered AI Tools · pp. 6–7

• Visualization-enabled HCAI tools connect human intentions and AI outputs through an interactive interface that bridges execution and evaluation gaps.Source: §4 · p. 7

DC1–DC5 Design Characteristics

• Overview-plus-detail views, confidence encodings, and visual before-and-after comparisons support tradeoffs, uncertainty awareness, explanations, and consequences of actions.Source: §4.3 · pp. 9–10
• Externalizing data lets users perceive patterns, outliers, relationships, and interaction history without retaining all details in working memory.Source: §4.3 · p. 10
• Shared data representations create common ground between users and AI models, helping users understand model behavior and correct errors.Source: §4.3 · p. 10
• Direct manipulation and persistent visual affordances make available actions explicit and reduce the gap between user intention and action.Source: §4.3 · pp. 10–11

Human Concerns

• Visualization can expose harmful bias by showing performance or disparities across many intersectional social groups.Source: §4.4 · p. 11
• Visualization supports transparency by exposing model structures, training dynamics, provenance, learned representations, internal states, and component relationships.Source: §4.4 · p. 11

Addressing Transparency

• Visualizations can expose model architecture, training processes, internal representations, and data flows to support transparency.Source: H Addressing Transparency · p. 11

Addressing Explainability

• Feature importance, saliency, and counterfactual visualizations can show why particular AI decisions were made.Source: H Addressing Transparency · pp. 11–12

Addressing Understandability

• Interactive visual explanations can improve understanding of AI outputs by offloading memory and presenting information at appropriate abstractions.Source: H Addressing Transparency · p. 12

Addressing Accountability

• Decision-making histories and audit trails provide a clear visualization-based mechanism for AI accountability.Source: H Addressing Transparency · p. 12

Addressing Provenance and Privacy

• Scalable graph, hierarchical, and timeline representations can present provenance histories, while visualizations can help users audit, secure, and configure privacy-sensitive data.Source: H Addressing Transparency · p. 12

Corpus and Analysis

• The analyzed tools were selected from a 76-tool literature corpus, with eight papers examined in depth to compare capabilities, human concerns, evaluation criteria, and design insights.Four selected papers came from the authors’ prior work and four from other researchers.Source: §5 · p. 13, §5.1 · p. 13

TimeFork

• TimeFork lets analysts directly set one stock’s future trajectory, then uses temporal and conditional models to predict responses across other visible stocks.This amplifies predictive scope from a handful of mentally tracked stocks to potentially thousands and reveals multivariate dependencies through historical correlations.Source: §5.2 · p. 14
• TimeFork users achieved the best portfolio value and investment strategy when combining human control with AI support rather than using fully manual or automatic decisions.The evaluation involved 13 participants investing virtual money at four points during a six-month period in 2014 and 2015.Source: §5.2 · p. 14
• The system exemplifies support rather than supplanting the user: human-AI collaboration outperformed fully manual and fully automatic conditions, especially when predictions were unreliable.Source: §5.2 · pp. 14–15
• TimeFork uses familiar stock line charts and immediate visual feedback to make model responses understandable without exposing model internals or feature-level reasons.Its explainability is therefore shallow: users see how other stocks respond but not why the model produces those predictions.Source: §5.2 · pp. 14–15

Tool overview

• HaLLMark records and visualizes writers’ interactions with an LLM, distinguishing human writing, AI writing, prompts, responses, and edits across the document’s evolution.The timeline supports exploration of when writers relied on the LLM and how contributions were incorporated.Source: §5.3 · pp. 15–16

Evaluation

• Evaluations with 13 creative writers found that HaLLMark encouraged early reflection on AI assistance, increased control and ownership, and could improve transparency about AI co-writing.Participants also saw the provenance record as useful for complying with publication policies requiring AI disclosure.Source: §5.3 · pp. 15–16
• Nine of 13 participants reported more control than with ChatGPT, and writers made more editing passes over AI-generated than self-written sections.These results indicate that provenance visibility supported conscious control and iterative refinement.Source: §5.3 · p. 16

Gaps in Coverage

• HaLLMark does not address fairness, explainability, or privacy: it shows AI involvement but not why suggestions were made, does not detect bias, and sends prompts to GPT-4’s API.These gaps constrain the tool’s coverage of human-centered AI concerns.Source: §5.3 · p. 16

Design Insights

• A single provenance visualization addresses multiple concerns by supporting agency, transparency, accountability, and ownership through color coding, interaction histories, and document timelines.The design uses simple linear timelines and colored rectangles rather than a multi-view visual analytics dashboard.Source: §5.3 · p. 16

Tool overview

• uxSense preprocesses video and audio with computer-vision and action-labeling models, generating synchronized streams for emotions, actions, sentiment, gaze, and other behavioral signals.This makes multimodal correlations accessible while retaining the original video and audio for inspection.Source: §5.5 · p. 19

Evaluation

• Five UX professionals reported that uxSense could ease daily work by providing high-level overviews and pinpointing areas of interest, but questioned whether the models were trustworthy enough for wholesale acceptance.The evaluation used an actual commercial software usability session, giving the assessment ecological validity.Source: §5.5 · pp. 19–20
• Preprocessing can help analysts identify critical incidents faster than scrubbing raw video, but doubts about model accuracy require verification against the video and can eliminate efficiency gains.The practical benefit therefore depends on appropriate trust in automatically flagged events.Source: §5.5 · p. 19

Gaps in Coverage

• The tool does not address fairness, accountability, explainability, or privacy, including demographic performance risks, absent audit trails, opaque model internals, and unprotected sensitive video processing.Without visibility into model confidence and error modes, analysts cannot calibrate appropriate reliance.Source: §5.5 · p. 20

Design Insights

• uxSense preserves analyst agency by presenting raw model outputs as visual data streams rather than automatically generating interpretations or decisions.Analysts retain responsibility for contextualizing emotion classifications, action labels, and sentiment scores.Source: §5.5 · p. 20

5.6 FairVis

• FairVis enables users to discover intersectional bias by selecting and comparing demographic subgroups through interactive visualizations.Users choose features and filter values to isolate subgroups, then compare their metrics and model behavior.Source: §5.6 · pp. 20–21
• The system combines feature distributions, subgroup overviews, detailed comparisons, and ranked subgroup suggestions to focus investigation on the most biased areas.Suggested subgroups are ranked according to worst observed performance for a selected metric.Source: Figure 6 · p. 21, §5.6 · pp. 21–22
• By moving from identifying anomalous subgroups to comparing their characteristics, FairVis makes intersectional bias understandable, explainable, and actionable for users.The system first exposes patterns and then provides comparative data to characterize why biases differ across groups.Source: §5.6 · pp. 20–21
• FairVis lacks model transparency, analysis provenance, accountability mechanisms, and inherent privacy protection.It does not expose the underlying model architecture, preserve analysis history, or mitigate privacy issues in source data.Source: §5.6 · p. 21
• FairVis grounds bias analysis in concrete classification differences, such as false positive rates, rather than only abstract fairness scores.Aligned comparisons across protected groups make disparities visually salient and support direct inspection of individual prediction patterns.Source: Figure 6 · p. 21, §5.6 · p. 22

5.7 RuleMatrix

• RuleMatrix extracts rules from black-box model outputs and displays their confidence and fidelity to approximate model behavior.The method perturbs inputs, induces rules from resulting outputs, and evaluates how well those rules represent the original model.Source: §5.7 · pp. 22–23
• The tool enhances model analysis by presenting fidelity and confidence both numerically and through visual encodings.These measures help users assess the reliability of extracted rules.Source: §5.7 · pp. 22–23
• RuleMatrix supports model interpretation by showing how outputs change in response to different input conditions without revealing internal model mechanisms.This provides users with behavioral explanations that do not require deep technical knowledge of the architecture.Source: §5.7 · pp. 22–23
• RuleMatrix lacks accountability and provenance features for tracking responsibility, data lineage, rules, or model versions.Users can understand model behavior but cannot determine who is responsible for decisions or where contributing artifacts originated.Source: §5.7 · p. 23
• A rule-based matrix translates opaque classifier decisions into human-readable patterns that users can inspect directly.Matrix cell size and color encode rule quality so users can scan many rules more efficiently than numerical tables.Source: §5.7 · pp. 22–23

5.8 Defogger

• Defogger uses reinforcement learning to recommend exploration strategies that maximize a limited differential-privacy budget.The system evaluates candidate pairs or triples from the user’s initial intent rather than automatically selecting the exploration path.Source: §5.8 · pp. 23–25
• The interface combines intent specification, ranked recommendations, and an adjacency-matrix-like display of uncertainty in privatized data.Custom matrix cells encode lower and upper bounds produced by noise added for differential privacy.Source: §5.8 · p. 23, Figure 8 · p. 24
• Case studies and a follow-up study found that users could gain significant insights from obfuscated data using Defogger.The studies involved two case-study participants and eight additional participants exploring insurance and hepatitis B datasets.Source: §5.8 · pp. 23–24
• Defogger augments exploratory analysis by helping users form hypotheses under noisy data and query constraints while preserving their control over exploration.Its recommendation engine optimizes the privacy budget but presents strategies for users to choose rather than making the decision automatically.Source: §5.8 · p. 24
• Defogger supports privacy and understandability but does not address human concerns beyond privacy.It prevents analysts from accessing raw protected data and uses uncertainty visualization to support interpretation.Source: §5.8 · pp. 24–25

5.9 Loops

• Loops automatically captures notebook edits and presents version history through a compact timeline with color-coded cell states and detailed diffs.Users can inspect changes to code, text, tables, and visualizations across versions.Source: §5.9 · p. 25
• Case studies and interviews found the color-coded provenance view intuitive and the detailed diffs useful for communicating analysis processes to collaborators.The demonstrations covered Austrian concert data and lung cancer analysis, with interviews involving four notebook users.Source: §5.9 · p. 25
• The provenance view helps users revisit abandoned approaches and recognize patterns in trial-and-error exploration that linear histories can hide.Direct access to every version supports fluid exploration without requiring commit hashes or tree navigation.Source: §5.9 · p. 26
• Automatic versioning preserves exploration history without requiring commented-out code or duplicate backup cells, balancing reproducibility and notebook readability.The working notebook remains clean while the full sequence of changes is retained.Source: §5.9 · p. 26
• Loops focuses on individual provenance and does not support collaborative interleaved edits or explain causal relationships between changes and outcomes.It shows what changed, but not why a change improved or degraded results.Source: §5.9 · p. 26

DG1 – Show, Don’t Tell

• Visualizations should show data without steering users toward predetermined conclusions, preserving users’ agency to interpret evidence and make decisions.This includes presenting raw data, neutral data landscapes, and non-directive color schemes.Source: §6.2 · p. 27

DG2 – Simple Is Plenty (Almost Always)

• Designers should favor simple, familiar, understandable visualizations while using progressive disclosure when multiple coordinated views are necessary.Additional views should serve distinct purposes rather than redundantly encode the same information.Source: §6.2 · p. 27

DG3 – Tackle Human Concerns Directly

• Visualization-enabled HCAI tools should address human concerns directly, including ethics, privacy, fairness, transparency, and understandability.Visualizations should communicate how data is used, how decisions are made, and relevant biases and limitations.Source: §6.2 · p. 27

DG4 – Encourage Interaction

• Interactive elements let users explore data and AI outputs, probe uncertainties, ask questions, and manipulate the interface on their own terms.This reduces the gulf of execution between users and AI models while supporting engagement and understanding.Source: §6.2 · p. 28

6.3 Limitations and Open Problems

• The guidelines are derived from eight exemplar tools, but four were created by the authors and came from domains whose constraints may limit transfer to healthcare, transportation, or manufacturing.The authors note that this concentration may bias the guidelines toward rationalizing their own design choices.Source: §6.3 · p. 28
• The framework’s generalizability is also constrained by limited expert diversity, venue-focused literature coverage, visualization complexity, interdependent human concerns, and emphasis on supervised-learning tasks.The authors specifically note limited representation of Asian perspectives, practitioners, and AI specialists, alongside uncertainty about applying the framework to multimodal foundation models and agentic systems.Source: §6.3 · pp. 28–29

Bridging gulfs through visual mediation

• Visualization can bridge Norman’s gulfs of execution and evaluation by making predictions manipulable, provenance tangible, and causal relationships explorable.The examples use line charts, visual timelines, and node-link diagrams to mediate interaction between humans and AI models.Source: §7 · p. 29

Progressive disclosure of complexity

• Progressive disclosure should provide simple defaults while enabling deeper investigation as AI models and interfaces become more complex.Future tools should move between high-level summaries and detailed explanations for novice and expert users.Source: §7 · pp. 29–30

Multimodal and embodied interaction

• Embodied interaction, tangible interfaces, and data physicalization could make AI parameters manipulable through physical action, spatial reasoning, and haptic feedback.These approaches extend visualization beyond screen-based interfaces.Source: §7 · p. 30

Summary

• The paper envisions visualization as a shared medium that makes AI-generated insights accessible across domains and user populations.This proposed lingua franca could support a common vocabulary for human-AI collaboration while addressing concerns such as fairness and transparency.Source: §7 · p. 30