Source-linked AI summary
Towards Actionable Visualization: Ten Years Later, What Generative AI Changes and What It Cannot
Leonel Merino, Mohammad Ghafari, Oscar Nierstrasz
TL;DR
The paper asks whether generative AI has made visualization production cheap enough to overturn the field’s earlier matching assumption. Through a self-critical revisit of prior survey findings, it argues that AI expands tractable domains and immersive delivery while elevating human judgment and worsening sustainability risks. The resulting agenda centers on helping people understand complex systems and AI-generated reasoning, not merely producing more artifacts.
Problem
Earlier software visualization research treated adoption as a costly matching problem between developers’ questions and available techniques.
Method
The authors revisit their 2016 survey findings, relate them to generative AI’s changing role in visualization production, and scan VISSOFT titles from 2017 through 2025 for delivery-medium change.
Results
Generative AI makes neglected domains more tractable and visualization production cheaper, while immersive-medium papers grew roughly fifteen-fold and tool sustainability may worsen.
Takeaways & Limitations
Actionable visualization should prioritize human perception, evaluation, and interaction for understanding complex software systems and AI-generated reasoning.
Abstract
from arXiv · showhide
Ten years ago, we surveyed software visualization research under the assumption that the challenge for adoption was matching developers' needs with techniques. Generative AI might have made that assumption obsolete by driving the cost of producing a visualization on demand toward zero. We argue that this mirrors a broader inversion already underway in software engineering: as AI devalues the production of artifacts, it elevates the human work of perceiving and directing them. Looking back at our 2016 research, we found that domains we flagged as neglected, such as rationale, are exactly the ones AI now makes tractable, and a tool-sustainability problem we diagnosed then is one generative AI may worsen before it helps solve. One finding shifted outright: the share of studies delivered through immersive environments grew roughly fifteen-fold, though it remains a small minority of the field. We argue that future research should focus on helping people understand large and complex software systems, including the reasoning processes generated by AI agents alongside the code itself. Humans have always been accountable for what they trust, but AI may lead them to trust without enough scrutiny. We confront the risks that follow, from the reliability of AI-generated visualization itself to the temptation of competing with machines on their own terrain, and conclude that actionable visualization is no longer a distant destination but a standard the field can now meet.
I. INTRODUCTION
Generative AI changes visualization from a costly, scarcity-bound artifact into an on-demand capability, shifting the central challenge toward human perception, judgment, and direction. The authors revisit this inversion and argue that visualization must help people assess what AI-generated views show.
- The original survey treated adoption as a matching problem between developers’ questions and available techniques, constrained by the cost of building specific visualizations.
- A developer can now obtain a disposable, purpose-built visualization for a specific codebase question instead of searching for the closest existing technique.
- Generative AI devalues visualization production while increasing the value of reading, judging, and directing visual artifacts.
- Human vision rapidly compares size and position in a bar chart, giving visualization an evidentiary channel that prose cannot match in speed or attention demands.
- The authors’ revisit finds that AI makes neglected domains tractable, may worsen tool sustainability, and coincides with growth in immersive delivery environments.
II. REVISITING THE 2016 LANDSCAPE, A DECADE LATER
The authors revisit their earlier classification of software visualization studies and identify three decade-later findings: two changed interpretations under generative AI, while one reflects changed data about delivery medium.
- The earlier survey classified 86 design studies across seven dimensions: task, need, audience, data source, representation, medium, and tool.
- The decade-later reassessment distinguishes changed interpretations of under-served domains and tool lifespan from an actual change in delivery-medium data.
A. Neglected domains, and the barrier AI removes: rationale, contracts, policies
Rationale, contracts, and policies were highly needed but under-visualized because preparing their heterogeneous data was too costly. Generative AI removes that preparation barrier, while human perception and judgment remain necessary to evaluate the resulting views.
- Rationale, contracts, and policies ranked highly in developer need but were among the least visualized domains, whereas architecture, concurrency, and dependencies received comparatively more attention than reported importance warranted.
- Answering who wrote a system component and why required manually correlating version-control metadata, commit messages, and source-code structure.
- AI can transform and map heterogeneous data, but people must recognize meaningful patterns and judge whether they are genuine or artifacts of transformation.
- Visualization systems should support movement from overview to confirming detail and back again, with each question guiding the next view.
B. The tool-lifespan problem AI may worsen before solving
Earlier visualization tools were short-lived and rarely open-sourced, reflecting incentives that favored publication prototypes over maintained infrastructure. Generative AI may amplify this sustainability problem by making undocumented, nondeterministic artifacts much cheaper to produce.
- 3.7 years was the average tool lifespan, while only 26% of sampled tools made source code publicly available.
- The authors attribute abandonment primarily to incentive structures that reward validating a publication hypothesis rather than documenting, generalizing, or maintaining a tool.
- Generative AI does not fix these incentives, and its nondeterminism may increase the volume of abandoned, undocumented, and unreproducible artifacts.
- Research evaluation should reward documentation and maintenance, making a smaller well-maintained tool stronger than an ambitious tool unlikely to survive its authors’ graduation.
C. Medium: a prediction that held
The survey’s prediction held: immersive delivery grew substantially as development barriers fell, while smart glasses may define the next medium shift. The authors therefore argue that medium should be measured explicitly in future design studies.
- Roughly fifteen-fold, papers explicitly naming immersive 3D, virtual, or augmented reality grew from about one in a hundred to about one in seven.The scan covered every VISSOFT paper from 2017 through 2025 across all tracks.
- Months of specialized toolchain work previously constrained immersive research, including CityVR and the more difficult CityAR.Sparse HoloLens documentation consumed much of the CityAR project effort.
- AI-generated scaffolding and boilerplate could reduce prototype-development friction, redirecting effort toward more robust immersive systems.The authors connect this possibility to the fifteen-fold increase in immersive-medium papers.
- Smart glasses may shift visualization toward ambient, glanceable AI-mediated access rather than another headset-based immersive environment.The authors present this as a medium to watch, not a settled design answer.
- Medium should become a first-class reported dimension because unmeasured shifts cannot be tracked consistently.
III. VISION: TOWARD A PERCEPTUAL RESEARCH AGENDA
Because matching developer needs to visualization techniques may no longer be the bottleneck, the field should reconsider what its research optimizes for. The authors propose three shifts following from this change.
- The authors propose three research shifts that follow from the possibility that AI has removed technique-production as the primary bottleneck.
A. From technique catalogs to perceptual capacity
The authors propose redirecting some VISSOFT effort from individual techniques toward human perceptual and evaluative capacity. The agenda focuses on whether software visualizations are trustworthy and how quickly misleading views can be detected.
- AI can generate competent visualizations from single questions, reducing the case for treating individual techniques as the community’s sole contribution unit.The earlier unit emphasized new layouts, metaphors, and tools evaluated on specific tasks.
- VISSOFT should study what makes software artifacts hard to see and what makes generated views trustworthy at a glance.Examples include dependency graphs, commit histories, and AI agents’ reasoning.
- Research should examine how quickly developers can detect that a visualization is misleading rather than merely unfamiliar.
- The proposed agenda grounds perceptual research in software-specific artifacts and questions rather than importing perceptual psychology wholesale.The authors distinguish this focus from the broader visualization community’s remit.
B. From monolithic tools to disposable, continuous seeing
The authors envision visualization systems organized around continuously examining software from whatever angle a current question demands. Generative AI should translate unarticulated intuitions into specifications, while deterministic machinery handles transformations and mappings.
- AI-facilitated rationale visualization points toward systems that continuously examine software from the angle demanded by each current question.Each disposable view answers the immediate question and is abandoned when the question changes.
- AI-assisted moldable development environments already practice the shift from task-specific artifacts toward continuously adaptable channels.
- AI should translate an initially unarticulated intuition into a specification that deterministic visualization machinery can execute reliably.
- Data transformation and visual mapping should remain deterministic rather than being regenerated by a language model on every request.The authors describe these operations as solved and caution against needless compute and energy use.
- The field’s accumulated knowledge is needed to ground models in choosing which representation fits which question.The authors identify this grounding as VISSOFT’s opening rather than a threat to its relevance.
C. A new object of study: visualizing the reasoning of AI itself
AI-generated code introduces a distinct visualization target: not merely what software does, but how reliably its statistical generation process can be trusted. The paper proposes adapting legacy-comprehension techniques and embedding visual inspection within agent-mediated coding.
- The new object of study: AI-generated code is often readable but not reliably comprehensible because legibility emerges from training rather than deliberate authorship.Its apparent meaning can vary across generated fragments.
- The new object of study: Vibe coding differs from AI-assisted coding because developers accept generated output with little line-by-line review.Such systems may be functionally correct while remaining structurally opaque.
- Research openings: Legacy-comprehension techniques such as dependency graphs, call graphs, and execution traces may be especially urgent for vibe-coded systems lacking familiar human scaffolding.The paper identifies these techniques as directly applicable to AI-generated systems.
- Research openings: Embedding navigable views of agent tool calls, intermediate steps, and proposed diffs can let developers inspect changes before acceptance.This shifts visualization from post-hoc comprehension toward inspection within the coding loop.
IV. RISKS AND A CLOSING REFLECTION
Generative AI makes visualization cheaper but relocates responsibility for verification and judgment to humans and, potentially, checking agents. The paper therefore calls for traceable, actionable views and concludes that visualization should now meet an actionable standard rather than merely approach it.
- Risks: AI-generated visualizations can misrepresent their underlying data, so views should use checkable mappings and remain navigable to source data.Automated, cheap generation means the connection between a view and its evidence must be designed deliberately.
- Risks: Structured, drillable visualizations can let agents query sources and check their own work before presenting changes.Navigability supports both human tracing and agent-side verification.
- Research agenda: Actionable-visualization benchmarks should evaluate data fidelity, source traceability, and support for reliable reasoning rather than visualization production alone.The proposed standard concerns whether the resulting view remains usable for trustworthy action.
- Research agenda: AI relocates verification and judgment instead of eliminating them, making human perceptual and evaluative capacities central to the research agenda.Generation can be delegated, but checking what is shown against what is true cannot.
- Closing reflection: The field argues that actionable visualization is now a standard it can meet, because question-specific visualizations can be generated, inspected, and discarded at low cost.The conclusion reframes the goal from progressing toward actionability to delivering it.