The dataset was constructed using illustrations from the English version of the PTCG. A total of approximately 20 000 cards released between the Base Set in January 1999 and the Pitch Black expansion in July 2026 were surveyed. To ensure comprehensive coverage and account for database inconsistencies, the manual screening process was conducted twice across all cards using two independent public online databases: Pikawiz (https://www.pikawiz.com/cards, last access: 17 July 2026) and Pokellector (https://www.pokellector.com/sets, last access: 17 July 2026). Every card, including identical artworks reprinted across different sets, promotional campaigns, or mini-sets like McDonald's and Trick & Trade, was individually inspected by hand. To further validate dataset completeness, targeted searches were also conducted using Pokémon names associated with mineral imagery (e.g., Sableye). However, to prevent statistical duplication, any identical illustration was counted only once as a single unique entry in the final dataset.

All cards were screened against predefined mineralogical criteria. Cards were included if they clearly showed features such as crystal faces, repeating geometric clusters, identifiable gemstones (including cut and organic gemstones), or distinct geological formations like natural ice crystals. However, official game mechanics and lore items with crystalline appearances, such as the “Terastal” phenomenon or “Evolution Stones,” were systematically excluded because this study focuses on environmental geological backgrounds created by illustrators, rather than pre-defined, franchise-specific items.

Following this dual-database screening, this study finalized a mineral illustration dataset of 223 cards, organized by release year and set number (see Supplement excel table and Figs. S2 to S4 in the Supplement). Representative mineral illustrations from this dataset are shown in Fig. 1, reproduced strictly for non-commercial academic research and educational critique under fair use guidelines.

Illustrated background environments were treated as natural outcrops for geological analysis. For each card, 13 categorical variables were recorded by the author, a trained geoscientist, to describe the mineral setting, growth habit, and physical form (Sect. S2.1). These variables were grouped into environmental context (e.g., growth setting, host rock type, and crystal exposure types) and intrinsic mineral traits (e.g., crystal size, dominant color, transparency, luster, optical effects, crystal form, crystal system, surface features, aggregation patterns, and the presence of paragenetic minerals). These parameters represent the fundamental diagnostic data used by geoscientists to identify inferred mineral species and deduce formational environments (Nesse, 2017; Okrusch and Frimmel, 2020).

Multiple Correspondence Analysis (MCA) was applied to the categorical variables to identify patterns in visual feature associations (Benzécri, 1979; Abdi and Valentin, 2007). This technique converts categorical responses into geometric coordinates, creating a multidimensional spatial map where frequently co-occurring variables are plotted closer together. By mapping these associations without imposing numerical assumptions, MCA facilitates the quantification of recurring visual combinations and identifies dominant representation patterns in the dataset (see Sect. S2.2 in the Supplement).

Inferred mineral species were initially evaluated by the author independently of artificial intelligence (AI) results. Since assigning natural mineral names to fictional illustrations is inherently interpretive, ambiguous images were assigned two plausible candidates to minimize bias, as detailed in the Supplement excel file. All mineral names assigned in this study are documented in Okrusch and Frimmel (2020).

These human interpretations were compared against three automated tools: two large language models (ChatGPT-5.3 and Gemini 3.0; standard publicly available versions) and a specialized rock-scanning app (Rock Identifier v2.19.2). ChatGPT and Gemini were selected as the most widely used AI platforms (Moore, 2026), reflecting how the general public might apply unspecialized AI to scientific tasks. Similarly, Rock Identifier was chosen because, as of June 2026, it was the most downloaded geological application across mobile platforms with highly positive user reviews. This selection effectively represents the primary tools an average collector would likely use to identify a mineral.

The AI models and the rock-scanning app were prompted to identify the mineral type from full card images and explain their reasoning (Sect. S2.3). Because generative outputs can vary, each image was queried three times, recording either the majority result or the inference most visually consistent with the crystal.

Historical statistical results indicate that PTCG publication rates accelerated from approximately 1.4 cards to 4.0 cards per day by 2023 (Fig. S1). The first clear mineral illustration appeared in the 2000 Dark Dragonair card (Fig. 1a). Following this initial appearance, mineral artwork steadily increased to account for over one percent of all new releases since 2007 (Fig. S1). This creates a steadily growing mineral dataset that reflects visual trends over nearly three decades, ensuring that players entering the game during different periods all have continuous opportunities to encounter mineral-themed card art.

A total of 223 PTCG cards containing mineral illustrations, produced by 85 different artists, were included in the analysis (Sect. S1.3). Although six artists each contributed more than ten mineral-themed illustrations (accounting for 29 % of the dataset), most artists produced only one or two cards. Despite this diversity of illustrators, the statistical results reveal remarkably consistent patterns in both the geological settings and crystal features depicted in the artwork. Quantification of these features showed that nearly half of the illustrated minerals occur in caves (99 cards, 44 %, Figs. 1d, 2a). Most are shown fully exposed (202 cards, e.g., Figs. 1f, 2a) rather than embedded in host rock (e.g., Fig. 1k), and when host rocks are present, they lack sufficient petrological textures for lithological identification.

Using the official species heights of the depicted Pokémon as scale references (see Sect. S2.1), over half the illustrated crystals exceed 30 cm along the c axis (Fig. 2c). These crystals are typically well formed (euhedral), pristine, unweathered, highly symmetrical (e.g., Fig. 1f), and terminate in elongate prismatic habits. Most occur as crystal clusters (e.g., Fig. 1a, n), whereas twinned crystals are uncommon (Figs. 1o, 2d). At the assemblage scale, the vast majority of illustrations depict only a single inferred mineral species (Figs. 1c, d, l, u, w, and 2b), while coexisting mineral assemblages are rare (e.g., Fig. 1i).

Because microscopic features such as inclusions, cleavage, and growth striations are absent, crystal system classification is necessarily based on two-dimensional geometric outlines. The dataset is dominated by tetragonal and hexagonal systems (the latter inclusive of trigonal forms) (e.g., Fig. 1, u, v), which together account for approximately 84 % of the identifiable crystal systems (Fig. 2h). These are typically illustrated as columnar crystals terminating in a single pyramid. The remaining crystal systems are notably underrepresented. The isometric system appears as the third most common type, mainly as octahedra (Fig. 1e, p, q, y), whereas orthorhombic (Fig. 1x), monoclinic, and triclinic systems (Fig. 1m) are rare. A final category of unidentifiable forms consists primarily of massive crystals lacking distinct faces, as well as faceted gemstones (Fig. 1b, k, t, w).

These highly repetitive environmental and crystallographic characteristics provided the basis for inferring the most likely mineral species represented in each illustration. Consequently, 22 mineral groups were manually inferred across the dataset. Varieties were consolidated into broad species groups; for example, amethyst (Fig. 1r) was classified under quartz. The inferred mineral diversity was highly uneven: the six most frequently inferred groups – ice, quartz, calcite, zircon, corundum, and fluorite – account for 83.4 % of the entire dataset (Fig. 3). Within this subset, ice and quartz overwhelmingly dominated, accounting for approximately 32 % and 28 % of all inferred minerals, respectively.

To statistically validate this simplified visual template and examine the relationships among these characteristics, MCA was performed. A Benzécri correction was applied to avoid underestimating the variance of categorical variables (see Sect. S2.2). The resulting ordination reveals three distinct patterns (Fig. 3). Calcite forms a clearly isolated cluster along the primary axis, whereas the confidence ellipses of ice, quartz, corundum, and zircon overlap extensively, indicating highly similar visual characteristics. Fluorite occupies a broader area of the feature space, suggesting greater morphological variability.

To compare human and machine interpretation of these stylized illustrations, three AI systems were applied to the complete dataset of 223 cards. Before analyzing the artwork, a control experiment using 43 photographs of real mineral specimens was conducted to verify that the AI systems could reliably identify realistic mineral images. The control photographs yielded accuracies of 95 % for Gemini, 86 % for ChatGPT, and 74 % for Rock Identifier (see Sect. S3), providing a baseline for comparison with the stylized card illustrations.

Across the 223 card illustrations, agreement between manual interpretation and AI ranged from 45 % to 56 %, whereas the specialized rock-scanning application achieved only 12 % agreement (Fig. 4a). Agreement between ChatGPT and Gemini was likewise limited (42 %), indicating considerable variation among AI interpretations of stylized artwork.

Agreement also varied substantially among mineral groups (Fig. 4b). Approximately half of the illustrations contain fantasy elements. Human–AI agreement was consistently lower for these fantasy-style illustrations than for more geologically realistic depictions (Fig. 4c). Compared with the manual baseline, AI systems systematically overidentified quartz while underrecognizing tetragonal minerals (e.g., zircon, Fig. 4d).

The quantified morphological patterns reveal a substantial divergence between PTCG representations and natural mineral occurrences. In the PTCG, minerals are typically depicted as giant (> 30 cm), pristine, and perfectly symmetrical crystals isolated in caves (e.g., Fig. 1d). In contrast, most real-world minerals exist as interlocking grains, veins in igneous or metamorphic rocks, or weathered fragments in placer deposits (e.g., Fig. 1s; Nesse, 2017; Okrusch and Frimmel, 2020). While giant euhedral crystals do occur in nature, such crystals are uncommon and require exceptionally stable geochemical conditions, such as those in the Naica Mine, Mexico (Fig. 1l; García-Ruiz et al., 2007). Likewise, minerals in natural systems commonly occur as paragenetic assemblages rather than isolated crystals. For example, quartz and feldspar are ubiquitous companions in many rock types, yet only 6 % of the dataset depicts multiple mineral species. PTCG illustrations clearly favor visually simplified geological scenes over realistic mineral occurrences.

The same idealization extends to optical and physical properties (Fig. 2). The dataset shows a severe scarcity of dark-colored minerals, earthy luster, opaque crystals, and fractured textures (Fig. 1g, m). Instead, minerals are consistently portrayed with bright color, glassy translucency, vitreous luster, and high geometric symmetry. These idealized characteristics closely match the commercial image of crystal in the gemstone market rather than raw ores found in the field. Furthermore, while common rock-forming minerals like quartz, feldspar, and calcite typically appear white, colorless, or gray in nature (Okrusch and Frimmel, 2020), the dataset shows a dominant preference for vibrant blue (Fig. 1f, j), further reinforcing the stylized visual identity of fictional minerals.

These environmental and optical biases stem from a fundamental reliance on a standardized visual archetype. Driven by aesthetic impact rather than scientific documentation, illustrators utilize a highly restricted visual vocabulary inspired by commercial specimens and curated museum displays. Consequently, crystal symmetry is treated merely as a flexible design element rather than a rigid geological constraint.

This reliance on familiar visual templates is reflected in the distribution of crystal forms, which are highly concentrated in tetragonal or hexagonal prisms terminating in pyramids along the c axis (e.g., Fig. 1f). Unexpectedly, tetragonal prisms dominate as the most common shape, sharply contrasting with the global abundance of hexagonal forms (e.g., quartz, calcite) in both natural environments and commercial markets. This discrepancy suggests that media representation is driven by visual impact rather than geological sampling.

These morphological biases extend beyond geometry to influence inferred mineral identity. Ice provides a robust test of this hypothesis because, unlike most other minerals, its identification relies primarily on contextual evidence (Pokémon types, move names, cold environments; Fig. 1z) rather than crystal morphology. If artists faithfully represented diverse minerals, ice should exhibit the hexagonal form expected for terrestrial ice Ih. Instead, MCA shows ice forms the most tightly clustered group (Fig. 3), indicating remarkably consistent visual representation. Crucially, more than half of the inferred ice crystals are drawn as tetragonal prisms, despite terrestrial ice crystallizing almost exclusively in the hexagonal system. This firmly demonstrates that illustrations prioritize a culturally familiar “crystal” archetype over actual crystallography.

Even excluding ice, the inferred mineral distribution remains highly concentrated. Among the remaining 153 cards depicting minerals stable under ambient conditions, nearly 40 % were inferred as quartz, while most of the remainder were assigned to stalactitic calcite, zircon, corundum, or fluorite. Collectively, these six dominant mineral groups account for 83.4 % of the dataset, showing that mineral identity is concentrated in a limited set of recurring groups. In contrast, common rock-forming silicates (feldspar, pyroxene, amphibole, mica) are almost absent. This pattern likely reflects their limited visual familiarity and their lack of traditional crystalline aesthetics. Likewise, gemstones such as diamond, beryl, and chrysoberyl are rare, probably because the public most commonly encounters them as polished gems rather than natural crystals.

The MCA results provide quantitative evidence for this visual simplification (Fig. 3). While natural quartz displays immense diversity from microscopic grains to banded agate, the cards restrict it to a single template of clear geometric prisms. The overlapping confidence ellipses for ice, quartz, corundum, and zircon strongly suggest that a single visual template dominates multiple mineral categories. The distinct calcite cluster primarily reflects its consistent depiction as stalactitic calcite in cave settings rather than common rhombohedral cleavage faces. In contrast, fluorite remains visually distinct because its unique cubic or octahedral habits are strong enough to resist the universal template. However, it remains the only clear exception, highlighting how rarely illustrators depart from the dominant crystal archetype.

Beyond mineral identity, some illustrations also depict physically or geologically implausible features. These include incompatible crystal systems within a single illustration (Fig. 1n), unusually large crystals growing from artificial environments or floating without visible support (Fig. 1c, f), faceted gemstones embedded directly in rock formations (Fig. 1d, f), and self-luminous crystals without an identifiable excitation source (Fig. 1j). Together, these examples show how artistic conventions override the physical and geological constraints that normally govern mineral occurrence and growth.

The AI analysis serves as a comparative test to examine how this visual simplification affects mineral interpretation. When processing photographs of real minerals (Sect. S3), the AI tools and manual evaluations yielded consistent and scientifically accurate answers. This demonstrates that modern AI performs well on natural mineral photographs when they follow geological principles (74 %–95 %). As shown in Fig. 4c, large language models achieved moderate agreement with human experts (62 %–66 %) on geologically plausible illustrations, but agreement was lower (26 %–45 %) for highly stylized, fantasy-based artwork. Consequently, this sharp decline in AI performance serves as quantitative evidence of the media-driven stereotype itself. It demonstrates how artistic stylization systematically strips away authentic diagnostic features, replacing geological reality with a simplified, symbolic crystal archetype that defies scientific identification.

A geologically trained observer can integrate environmental context with crystal features. For example, when a transparent crystal emerges from snow, grass, or water, a geologically trained observer recognizes that the environment is inconsistent with quartz formation and identifies it as ice (e.g., Fig. 1h, z). In contrast, current AI models lack this contextual reasoning and suffer from anchoring bias, focusing solely on the isolated, stylized crystal shape (e.g., O'Leary, 2025; Takenami et al., 2025). Because these generalized shapes heavily mimic the universal template, quartz becomes the most commonly identified mineral across all AI platforms (Fig. 4d). By over-relying on this visual anchor instead of analyzing the entire scene, the models produce predictions that fluctuate wildly across different iterations. These observations suggest that current AI should be treated as a supporting tool rather than a definitive identification system when interpreting stylized media.

Another critical limitation is that AI models struggle to infer three-dimensional symmetry from two-dimensional artwork. Consequently, they rarely utilize crystal form and symmetry as diagnostic criteria. This causes a fundamental disconnect between the inferred mineral species and the depicted morphology, leading models to repeatedly misclassify clear tetragonal prisms as hexagonal quartz (Fig. 4d). Dedicated rock-scanning applications perform worst, as their training on photorealistic specimens renders them unable to interpret artistic abstractions. Ultimately, these illustrations preserve the cultural image of a “crystal” while actively omitting many of the geological characteristics needed for reliable identification.

The analytical results show that Pokémon card backgrounds often portray minerals as isolated, large, translucent, and perfectly formed crystals, reinforcing a standardized visual template. Similar limitations in mineral representation have been reported in other informal educational resources, including virtual mineralogical museums (Calvo and Lucha, 2024). This template likely reflects the influence of commercial gemstone markets, museum displays favoring well-formed crystals (Francek, 2013), and the artistic backgrounds of illustrators. Recognizing this simplification provides an opportunity for geoscience communication, as these familiar images can be used to discuss how popular culture shapes perceptions of minerals and geological processes.

PTCG illustrations are particularly suitable for mineral education because they are freely accessible online and familiar to many children worldwide. Selected cards function as visual hooks in school science workshops or museum tours, where comparisons with real specimens introduce natural mineral forms and geological environments. This use of the Pokémon franchise for science communication also has a concrete precedent in the Pokémon Fossil Museum, including its 2026 North American exhibition at the Field Museum in Chicago, where visitors compare Fossil Pokémon with real fossils to explore paleontology (Field Museum, 2026).

Although six dominant mineral groups account for most of the 223 mineral-related cards, the dataset still spans 22 inferred mineral groups. Several cards nevertheless depart from the dominant visual template by depicting features such as cleavage, fractures, or coexisting minerals (e.g., Fig. 1e, g, i). These examples can serve as educational “Easter eggs,” allowing educators to use familiar illustrations alongside real specimens to introduce mineral properties and geological relationships. The dataset therefore provides both a recognizable visual pattern and a range of mineralogical exceptions for educational use.

PTCG remains a creative medium, and scientific realism need not constrain artistic expression. Future releases could build on this existing collection without sacrificing artistic appeal. Occasional mineralogical “Easter eggs” could draw on geological references or physical specimens to depict a wider variety of crystal habits and mineral associations. Collaboration with mineralogists could introduce diverse, scientifically recognizable properties, such as calcite birefringence or fluorite fluorescence (e.g., Fig. 1j). More realistic mineral assemblages and deposit settings, such as characteristic mineral associations in pegmatites, could broaden the geological contexts represented. Such details could give players another way to explore the real-world ideas behind the artwork and add a deeper layer of engagement.

Ultimately, analyzing these media-driven stereotypes moves beyond simply providing teaching tools. It empowers the audience to consume popular culture with a more critical eye. During routine gameplay, players may notice the Earth science imagery in the card backgrounds and reflect on the natural world that inspired these designs. By bridging the gap between familiar entertainment and authentic geological knowledge, this approach can broaden public engagement with the Earth sciences.