Source-linked AI summary
Ethical Dimensions of Visualization Research
Michael Correll
TL;DR
The paper addresses the unclear ethical duties associated with visualization and visual analytics beyond ordinary scientific and engineering obligations. Drawing on historical examples and current research dilemmas, it synthesizes ethical concerns and proposes additional obligations for visualization designers, builders, and researchers. It concludes that visualization work requires caution, attention to hidden impacts, and cultivation of appropriate values and virtues.
Problem
Visualization can strongly influence decisions, yet its ethical duties beyond ordinary scientific and engineering obligations remain unclear.
Method
The paper draws on historical and contemporary examples, ethical perspectives, and current visualization research dilemmas to develop additional obligations.
Results
The paper identifies moral dimensions and ethical tensions in visualization, including privacy, analytical guidance, explainability, provenance, uncertainty, and hidden impacts.
Takeaways & Limitations
Visualization researchers should examine the broader impacts of their work, make hidden values and harms visible, and cultivate ethical values and virtues.
Takeaways & Limitations
Visualization can create distance from affected people, and making human components more prominent may still fail to significantly increase empathy with suffering.
Abstract
from arXiv · showhide
Visualizations have a potentially enormous influence on how data are used to make decisions across all areas of human endeavor. However, it is not clear how this power connects to ethical duties: what obligations do we have when it comes to visualizations and visual analytics systems, beyond our duties as scientists and engineers? Drawing on historical and contemporary examples, I address the moral components of the design and use of visualizations, identify some ongoing areas of visualization research with ethical dilemmas, and propose a set of additional moral obligations that we have as designers, builders, and researchers of visualizations.
1 INTRODUCTION
The paper argues that visualization research has an inescapable moral character because it shifts power and reflects social, political, and ethical values. It examines this character historically and in current research, then proposes additional obligations for visualization researchers.
- Visualization research has moral character because technical work can shift power among social groups and affect society politically.
- Historical and contemporary examples illustrate that even superficially apolitical or trivial visualization research has moral implications.
- The paper challenges the view that data and visualization are apolitical, ethically neutral, or merely objective reporting.
- The paper examines ethical implications in current visualization research and proposes obligations beyond researchers’ existing duties as scientists, teachers, and citizens.
2 AGAINST THE NEUTRALITY OF DATA
The paper rejects the neutrality of data: collecting, quantifying, structuring, and omitting data are political acts with ethical consequences. Data choices can redistribute visibility, constrain applicability, and shape political action.
- Thomas Paine repurposed accounting data about naval forces to argue for English weakness and American strength in support of independence and revolution.
- Mass data collection can frame people as standing reserves of ad revenue, content creators, soldiers, or bodies.
- Collecting and distilling people into data has political power, as historical statistics supported taxation, war-making, and Nazi bureaucratic violence.
- Refraining from collecting data can exclude populations, constrain generalizability, and contribute to unequal outcomes such as biased face-recognition performance.
- Data are not raw, neutral facts because someone collects or processes them for a purpose, making observation and quantification political acts.
3 AGAINST THE NEUTRALITY OF VISUALIZATION
The paper argues that visualizations are rhetorical artifacts rather than neutral depictions: design choices shape messages, visibility, persuasion, and the treatment of human suffering. Ethical visualization therefore requires confronting rhetoric, power, omission, and the limits of empathy.
- Nazi maps made colonization visible while omitting or minimizing displaced populations and the original borders of annexed Poland.
- Visualizations can turn human suffering into abstract figures, creating distance between affected people and those consuming the data.
- Visualizations are rhetorical and can persuade through minor design choices, including titles, genre conventions, and presentation of uncertainty.
- Even statistical graphics can support political arguments, while Du Bois intended apparently factual charts to advance a moral and political argument about African-Americans.
- Critical infovis and data feminism respond by making the values, politics, and power imbalances of visualization explicit.
4 CONCERNING TRENDS IN VISUALIZATION RESEARCH
Visualization research contains emerging ethical dilemmas because visualizations allocate power and responsibility, while researchers must balance conflicting values rather than follow a single clear rule.
- Emerging visualization topics can allocate power and responsibility in ways that produce unethical or irresponsible outcomes.
- Charts can lend arguments authority and an appearance of objectivity, so designers must consider their persuasive use.
- The paper selects ongoing research areas where ethical values and virtues conflict, rather than where one rule determines the correct path.
- The discussion concludes each topic with open-ended design dilemmas about the ethical implications of visualization research.
Automated Analysis
Automated visual analytics can broaden access to data insights, but may encourage statistically unsupported conclusions and create a tension between user empowerment and sound decision-making.
- Automatic recommendations aim to help analysts discover important data relationships without manually exploring trivial patterns.
- Automated analytics can amplify unsupported conclusions because users may lack statistical tools to validate visual insights.
- “P-hacking machines” are attractive because finding a pattern provides a better user experience than finding nothing, despite users’ limited statistical expertise.
- These systems may enable unjustified and damaging decisions without sufficient attention to the people harmed or users’ capabilities.
- Automatic insights can empower users lacking time or expertise, while excessive guidance can reduce user agency.
- The central design dilemma is how prescriptive analytics systems should be when users risk statistically spurious conclusions.
Machine Learning
Visualizing machine-learning decisions raises an ethical trade-off between explaining consequential models to affected people and preserving model complexity, accuracy, privacy, and resistance to manipulation.
- Opaque machine-learning models can affect critical decisions, creating duties to communicate their reasoning to impacted populations.
- Much explainability research serves expert users rather than people affected by decisions such as loan or parole eligibility.
- Public-facing communication lacks standard methods and established success stories for visually explaining algorithmic decisions.
- Simpler models may be more explainable but often less accurate, placing transparency and utility in conflict.
- Model explanations can let bad actors game systems, while full transparency can compromise the privacy of people whose data were collected.
- The design challenge is deciding how much abstraction to use and which parts of algorithmic decision-making to display.
Provenance
Visualization systems increasingly need to expose how analyses were produced, but documentation remains harder in accessible GUI tools than in coding-based environments.
- Growing system complexity and many possible analytical actions make it difficult to reconstruct how a chart or conclusion was produced.
- Visualizing data and decision provenance can support criticism and transparency, yet most systems prioritize exploration over documenting analytical processes.
- Coding environments support transparent analysis communication but require scripting expertise, unlike GUI visual analytics systems.
- The resulting gap between ease of analysis and ease of documentation blurs exploratory and confirmatory analytics.
- Surfacing provenance and analytical choices can help viewers resist unreliable information and navigate the tension between agency and correct, useful communication.
- Open design questions include which alternate analytical decisions to show and whether provenance should be structured to reveal irregularities.
5 WHAT ARE OUR OBLIGATIONS AS VISUALIZATION RESEARCHERS?
Visualization researchers have ethical duties beyond those associated with science and engineering because they shape how people encounter data, patterns, and conclusions. The paper frames these duties as principles for addressing ethical challenges, while recognizing that they can conflict and require caveats.
- Visualization work combines the ethical obligations of scientists and engineers, including avoiding breaches of consent, harm, and work dangerous to the public good.
- Visualization researchers have special responsibilities because they influence how people perceive patterns and draw conclusions from data.
- The paper examines ethical challenges involving visibility, privacy, and power, and proposes principles intended to reduce visualization’s negative impacts.
- These principles are not absolute: applying them can create unwanted ethical effects or conflict with other virtues.
Make the Invisible Visible
The paper argues that visualization should expose labor, uncertainty, and potential impacts that are often omitted from finished visual artifacts. Greater visibility can support fairness, reproducibility, accountability, and better-informed decisions, but may reduce comprehensibility.
- Hidden labor: Making data, analysis, design, and user-research labor visible supports fairness, reproducibility, openness, and standards development.
- Hidden uncertainty: Uncertainty should be shown because its presentation can measurably affect decisions, including hurricane-risk judgments and perceived electoral closeness.
- Hidden uncertainty: Public audiences can use uncertainty information effectively when it is presented through appropriate visual designs.
- Hidden impacts: Researchers should examine potential harms, misuse, excluded groups, predatory applications, discontinuation effects, and opportunity costs alongside positive system outcomes.
- Caveats: Adding transparency, uncertainty, and counter-narratives can make visualizations harder to interpret and limit their audience.
Collect Data With Empathy
The paper argues for empathetic data practices that limit unnecessary collection, protect privacy, and reconnect abstractions with people. These goals conflict with analytical breadth, context, and the risks of biased or emotionally manipulated judgments.
- Small data: Researchers should resist pressure to collect ever more personal data because collection can erode privacy and agency.
- Anthropomorphize data: Visualizations should recenter people affected by data, especially in contexts involving human suffering, because abstraction can weaken empathy.
- Obfuscate data: Privacy-preserving designs can aggregate, fuzz, or restructure sensitive data while communicating limits on accuracy or detail.
- Caveats: Restricting collection can reduce analytical quality and scope, introduce selection bias, and produce unjust or contradictory conclusions.
- Caveats: Empathic reactions can be shaped by moral luck and misused to support violent or discriminatory ends.
Challenge Structures of Power
The paper proposes challenging power structures by supporting due process, advocating for underrepresented causes, and resisting unethical analytical work. These obligations are constrained by unequal resources, retaliation risks, and conflicts between truth-seeking and existing power structures.
- Power and resources: Data collection and academic research resources favor governments and corporations, limiting visualization’s ability to challenge powerful actors.
- Data due process: Data-driven decisions can turn imperfect datasets into unappealable assertions, so affected people deserve agency, representation, and understandable procedures.
- Data advocacy: Visualization expertise can amplify marginalized causes and increase attention to ongoing injustices through more relevant datasets.
- Unethical behavior: Researchers may need to pressure or slow unethical analytical work through dissent, resignation, or other forms of resistance.
- Caveats: Promoting truth may require amplifying expertise and power while suppressing conflicts, creating a tension for data advocates.
6 CONCLUSION
The paper synthesizes ethical perspectives on visualization and calls for greater mindfulness about visualization’s ethical implications. It also acknowledges that ethical design cannot be guaranteed by a complete, prescriptive decision criterion.
- It is not a complete survey or an exhaustive prescriptive criterion for guaranteeing that visualizations are designed or deployed ethically.The paper notes that disagreements about moral courses of action make such a criterion unlikely.
- Future work should develop pedagogy for visualization researchers and designers and study visualization’s ethical impacts after deployment.
- The work synthesizes existing critical and ethical views of data and visualization while calling researchers to cultivate appropriate values and virtues.