Anthropographics: Human-Centered Data Visualization
- Anthropographics is a term defining human-centered data representations across urban emotional mapping, human-shaped visualizations in journalism, and anthropomorphic language profiling in AI.
- It employs methods such as geo-located social media analysis, kernel density estimation, and multi-dimensional frequency profiling to reveal human presence in data.
- The approach enhances practical applications in urban planning, ethical data visualization, and AI evaluation while addressing concerns about privacy, representation, and bias.
Anthropographics is a polysemous term used in several research traditions to denote human-centered forms of data representation. In urban computing, it refers to graphical mappings of geo-located emotional expressions in cities, derived from social-media streams and organized across time, space, emotion, theme, and culture (Iaconesi et al., 2014). In information visualization and data journalism, it denotes visualizations in which the fundamental mark takes the shape of a human figure, typically to “show the people behind the numbers” in demographic or humanitarian data (Dhawka et al., 2022). In research on LLMs, it has been redefined as the multi-dimensional frequency profile of anthropomorphic linguistic behaviours emitted during dialogue, measured over categories such as empathy, first-person pronoun use, agency, and sentience claims (Ibrahim et al., 10 Feb 2025). A plausible implication is that anthropographics functions less as a single stabilized technical concept than as a family of methods for making human presence, human feeling, or perceived human-likeness legible in data.
1. Terminological scope and major senses
The term has at least three explicit technical definitions in the recent literature.
| Usage domain | Definition | Representative paper |
|---|---|---|
| Urban emotional mapping | Graphical mappings of how people feel in urban spaces using geo-located emotional expressions | (Iaconesi et al., 2014) |
| Information visualization | A class of visualizations in which the fundamental mark takes the shape of a human figure | (Dhawka et al., 2022) |
| LLM evaluation | The multi-dimensional frequency profile of anthropomorphic linguistic behaviours in chat | (Ibrahim et al., 10 Feb 2025) |
In the urban-emotion sense, anthropographics is tied to the harvesting and visualization of emotional traces left on major social networks, with the goal of understanding recurring emotional patterns around work, study, leisure, consumption, waiting, and social interaction in cities (Iaconesi et al., 2014). In the information-visualization sense, anthropographics foregrounds demographic or humanitarian subjects through stylized human icons or silhouettes, often in journalism and social-good communication (Dhawka et al., 2022). In the LLM-evaluation sense, the basic object is not a figure or a map but a vector of behavioural frequencies, , where each measures the proportion of model messages exhibiting a given anthropomorphic behaviour (Ibrahim et al., 10 Feb 2025).
A further, adjacent usage appears in anthropometry-centered computer vision and graphics, where estimated body measurements can drive “Anthropographic visualization” such as percentile plots, radar charts, or avatar reconstruction (Škorvánková et al., 2021). This usage is not identical to the visualization and LLM senses above, but it shows that the term has also circulated near anthropometric modeling pipelines.
2. Urban emotional anthropographics
The 2014 work on emotional landmarks in cities formulates anthropographics as an end-to-end framework for constructing graphical mappings from geo-located emotional data harvested from social networks (Iaconesi et al., 2014). The harvesting stage includes Twitter, Facebook public posts, Flickr, Foursquare check-ins, and Instagram captions. Geolocation can be explicit through latitude/longitude tags or implicit through place names and check-in metadata using gazetteers such as OpenStreetMap; harvesting can be done with bounding-box queries, streaming APIs for real-time “consciousness of streams,” or batch dumps for historical analysis. Pre-filtering includes language detection across 29 languages via WordNet translations and keyword or emoticon filters.
Each harvested post is stored with a metadata schema that includes a unique post identifier, anonymized user identifier, UTC timestamp, geo-coordinates and accuracy estimate, ISO 639-1 language code, raw text, preprocessed tokens, emoticons, theme tags, emotion labels with confidence scores, and processing flags. Annotation proceeds by tokenization and normalization, emoticon detection, Latent Semantic Analysis embedding through a term-document matrix and SVD reduction, and similarity scoring against WordNet-Affect synset vectors for basic emotions including joy, anger, fear, sadness, surprise, and boredom. Emotions are assigned when the similarity exceeds a threshold (Iaconesi et al., 2014).
The formal representation of emotional distribution is given by a normalized density function over space and time:
Here, is the number of posts labeled with emotion , are the location and timestamp of post , and the kernels may be Gaussian or Epanechnikov, with example scales of in space and in time. A spatio-temporal autoregressive model is further proposed to represent emotional evolution across discrete time steps through local persistence, spatial interaction, and exogenous inputs such as special events or weather (Iaconesi et al., 2014).
Visualization modalities include heatmaps, isocontours, and 3D surfaces for spatial emotion density; timeline sliders and linked small multiples for temporal comparison; thematic glyphs at points of interest; conversation graphs in which nodes are users or posts and edges represent reply or retweet links; and an “Emotional Compass” in which radial thickness integrates emotion density along directions from the user’s current position (Iaconesi et al., 2014). The framework is presented as useful for anthropology, sociology, urban planning, public policy, art and design, and location-based services. The same work also identifies data biases, privacy risks, cross-cultural validity problems, and scalability constraints in real-time kernel density estimation as open issues.
3. Human-shaped anthropographics in visualization and journalism
In visualization research, anthropographics are defined as visualizations whose fundamental mark is a human figure (Dhawka et al., 2022). The standard rationale is to humanize statistics, especially for demographic and humanitarian data, by replacing abstract marks such as bars or lines with silhouettes, pictographs, or other human-shaped glyphs. One common design pattern is that each icon stands for one individual, thereby increasing granularity and making counts legible as collections of people rather than as aggregate magnitudes (Dhawka et al., 2022).
Early empirical work discussed in this literature used simple, stylized human-shaped icons in donation-allocation experiments to test whether these marks elicited more empathy than bar charts or other abstract encodings (Dhawka et al., 2022). Morais and collaborators later formalized a seven-dimensional design space consisting of granularity, specificity, coverage, authenticity, realism, physicality, and situatedness. This formalism remains the most explicit design framework associated with anthropographics in the visualization literature (Dhawka et al., 2022).
The same body of work documents broad uptake in journalism. The New York Times used rows of identical human silhouettes to memorialize COVID-19 fatalities; The Washington Post used repeated icons for mass-shooting casualties; the Associated Press filled state outlines with shaded human figures to map Afghan evacuees; and The Guardian paired anthropographics with graphic imagery in coverage of famine (Dhawka et al., 2022). In direct experimental work on pictographs of mass-shooting victims, anthropographics are described as depictions of people that foreground the human dimension of data rather than abstract quantity marks, and as a form that may amplify personal connection, emotional resonance, and perceived authenticity (Sukumar et al., 12 Sep 2025).
This sense of anthropographics is therefore inseparable from two linked ambitions: representational humanization and affective engagement. At the same time, the empirical record summarized in the literature does not support a simple claim that human-shaped marks always increase empathy or prosocial action; later studies repeatedly report modest, mixed, or context-dependent effects (Sukumar et al., 12 Sep 2025).
4. Affective response, prosociality, and identity effects
Direct empirical testing of anthropographics has focused on affective change, charitable behaviour, and identity-linked response. In a crowdsourced study with 0, participants viewed a pictograph of approximately 1,400 mass-shooting victims arranged in a grid, with all victims from one racial group highlighted and a dotted enclosure indicating the expected population-proportionate count for context (Sukumar et al., 12 Sep 2025). Affect was measured before and after exposure with the 9-point Self-Assessment Manikin valence scale, and the change score was defined as 1. Across all conditions, racial concordance had a modest but significant effect: in-group viewers showed greater negative affect change than out-group viewers, with means of 2 versus 3, and the ANOVA main effect of concordance was 4, 5, 6 (Sukumar et al., 12 Sep 2025). The highlighted racial category itself did not produce a significant main effect, and neither mediation through in-group identification nor moderation by in-group favoritism was supported.
Related work on affective objectives in humanitarian visualization, while not using specialized icon-based anthropographics, is important for interpreting the broader design space (Lee-Robbins et al., 3 Jul 2026). In data videos about Somalia’s drought, displacement, and food insecurity, three narrative forms were compared: data-driven, human-driven, and mixed. Human-driven narratives based on photographs and stories of individuals produced the highest mean donation, 71.028$f_i$9 and the mixed narrative produced $\theta_{\text{emotion}}$00.75$\theta_{\text{emotion}}$1F(2,457)=5.53$\theta_{\text{emotion}}$2p=0.004$\theta_{\text{emotion}}$3p=0.003$\theta_{\text{emotion}}$4b=+0.31$\theta_{\text{emotion}}$5p=0.036$.
Taken together, these findings indicate that human-centered visualization strategies can increase emotional engagement, but the pathway is not monotonic. Anthropographics can reveal social-identity effects, as in racial concordance, yet sequentially mixing data and human stories can reduce donations relative to human-focused narratives alone (Sukumar et al., 12 Sep 2025, Lee-Robbins et al., 3 Jul 2026). The literature therefore treats affective outcome as an empirical property of specific encodings, framings, and contexts rather than as an inherent consequence of using human-shaped marks.
5. Critiques, risks, and ethical design
A critical strand of anthropographics research argues that human-shaped visualization can reproduce problematic social framings when applied to marginalized populations (Dhawka et al., 2022). Three critiques are central.
Homogeneous depictions of marginalized populations. Because many anthropographics rely on a small set of generic templates, they can flatten diversity within the groups they represent. The critique cites examples such as identical silhouettes used for migrants from Southeast Asia and the Middle East, and the New York Times COVID-19 memorial, in which every decedent appears as the same neutral-gray profile. The concern is that homogeneity induces a generalized “other” rather than recognition of distinct individuals and subgroups (Dhawka et al., 2022).
Treating marginalization as an inclusion criterion. Anthropographics often appear when datasets emphasize crisis, persecution, or suffering. The Guardian’s famine story is discussed as a case in which human icons were paired with graphic photographs of starvation, reinforcing a narrative of helplessness. The critique is that repeatedly selecting datasets because they depict trauma may normalize the view of marginalized groups solely as victims, obscuring agency, resilience, or ordinary life (Dhawka et al., 2022).
Insufficient contextualization of datasets about marginalization. Design choices can acquire unintended symbolic force when they are made without sufficient historical or cultural context. The “Slave Voyages” data sculpture used colored glass beads to encode numbers of enslaved people, even though glass beads had historically functioned as currency in the slave trade. The Associated Press map of Afghan evacuees used a brown-shaded ramp that some viewers could misread as a skin-tone metaphor. These examples are used to argue that materials, textures, and color ramps in anthropographics cannot be treated as neutral (Dhawka et al., 2022).
The literature links these problems to broader equity-aware visualization frameworks, including Schwabish and Feng’s “Do No Harm” guide, Dörk et al.’s emphasis on contingency and power dynamics, and Data Feminism principles (Dhawka et al., 2022). Proposed responses include techniques for representing demographic differences more explicitly, participatory design with affected communities, broader evaluation criteria beyond empathy, testing for unintended interpretations, and transparent documentation of design decisions. The general position is not that anthropographics should be abandoned, but that their humanizing intent does not eliminate the need for historical, political, and representational care.
6. Anthropographics in AI systems and adjacent computational uses
In LLM evaluation, anthropographics has been reformulated as a quantitative profile of anthropomorphic model behaviour over dialogue (Ibrahim et al., 10 Feb 2025). The profile is defined over 14 behaviour types grouped into personhood claims, internal states, physical embodiment claims, and relationship-building. The evaluation framework uses automated 5-turn conversations generated from 120 base prompts expanded into 960 scenario-specific seeds across friendship, life coaching, career development, and general planning, yielding 4,800 target-model messages per model. Three judge models label 13 of the 14 behaviours in every message, for 561,600 judgments in total, while first-person pronouns are counted directly. The key methodological claim is that many anthropomorphic behaviours emerge only after the first turn; nine of the fourteen first appeared in turn 2 or later for at least half of the dialogues (Ibrahim et al., 10 Feb 2025).
The same study validates the automatic scores with a human subject experiment involving 6 participants (Ibrahim et al., 10 Feb 2025). Participants chatted with a model configured to be either “high-anthropographics” or “low-anthropographics.” The high condition scored higher on the Godspeed Anthropomorphism Survey, with a mean of 7 versus 8, and higher on AnthroScore. Across Gemini 1.5 Pro, Claude 3.5 Sonnet, GPT-4o, and Mistral Large, anthropographic profiles were reported as very similar: validation and first-person pronoun use appeared in more than half of messages, while explicit human-AI relationship claims and physical movement occurred in fewer than 10% (Ibrahim et al., 10 Feb 2025).
A complementary qualitative method is provided by a prompt-based walkthrough study of ChatGPT, Gemini, Claude, and Copilot (Maeda, 22 Feb 2025). That work catalogs anthropomorphic features under four lenses—cognition, agency, relation, and biological metaphors—and reports that anthropomorphism appears both as subjective language, such as “I think” or “I feel,” and as a sympathetic conversational tone, such as “I’m here to help” (Maeda, 22 Feb 2025). Socio-emotional cues in prompts increased the incidence of such features, and role assignments such as “You are my best friend” or “You are my career coach” elicited more relational and emotional markers.
In adjacent anthropometry-centered research, estimated body measurements can support “Anthropographic visualization,” including stature-versus-chest/waist percentile plots, spider-web radar charts, and SMPL-like avatars (Škorvánková et al., 2021). Related systems reshape or generate bodies from limited measurements, such as 19-parameter body reshaping with MICE imputation and local mapping, 36-measurement conditional generation in AnthroNet trained on 100,000 synthetic humans, and A2B conversion of 36 tailor-style measurements into consistent SMPL-X body-shape parameters (Zeng et al., 2021, Picetti et al., 2023, Ludwig et al., 2024). Behavior-aware scene generation extends anthropometric profiles further into layout optimization by translating object-behavior relations into anthropometric constraints and reporting improvements in task completion time, trajectory efficiency, and manipulation space in real-scale studies (Jin et al., 3 Mar 2026). These works are centered on anthropometry rather than the classic human-shaped-visualization sense of anthropographics, but they show the term’s expansion into broader computational pipelines for human-centered modeling.
7. Research directions and conceptual synthesis
Across its different meanings, anthropographics is consistently concerned with translating human-related phenomena into structured, inspectable representations. In the city-emotion tradition, the target is collective feeling distributed over urban space and time (Iaconesi et al., 2014). In visualization and journalism, the target is the human legibility of demographic or humanitarian counts through human-shaped marks (Dhawka et al., 2022). In LLM research, the target is the measurable surface of anthropomorphic behaviour in multi-turn dialogue (Ibrahim et al., 10 Feb 2025).
Several research directions recur across these domains. One is richer representation of heterogeneity: urban emotional anthropographics raises cross-cultural validity and multilingual expression; critical visualization research calls for representing demographic differences rather than relying on homogeneous silhouettes; and LLM anthropographics distinguishes among fourteen behaviour types rather than treating anthropomorphism as a single scalar (Iaconesi et al., 2014, Dhawka et al., 2022, Ibrahim et al., 10 Feb 2025). A second is stronger evaluation: recent work moves beyond descriptive visualization toward donation behaviour, affect change, civic engagement, and human-likeness judgments in large-sample studies (Lee-Robbins et al., 3 Jul 2026, Sukumar et al., 12 Sep 2025, Ibrahim et al., 10 Feb 2025). A third is governance and ethics: privacy risks in geo-located emotion mapping, representational harms in marginalized-population visualizations, and over-trust or undue influence in anthropomorphic AI systems all motivate domain-specific safeguards (Iaconesi et al., 2014, Dhawka et al., 2022, Ibrahim et al., 10 Feb 2025).
A plausible implication is that the most stable core of anthropographics is methodological rather than ontological. The term does not identify one fixed visual form, one dataset type, or one application domain. Instead, it denotes a family of representational practices that attempt to render people, feelings, identities, or human-likeness visible in ways that conventional aggregate graphics or single-turn benchmarks do not.
