ArtBulb: AI Art Copyright Framework

• ArtBulb is an interpretable, quantifiable framework that assesses AI-generated artworks for a legally protectable, distinctive artistic style based on consistent, unique, and prompt-accurate criteria.
• It employs Description-Guided Clustering and multimodal large language models to measure stylistic consistency, creative uniqueness, and expressive accuracy through integrated visual and textual embeddings.
• The framework bridges legal and technical domains by providing structured, expert-guided reports that quantify style proximity to human art, supporting copyright judgment decisions.

Searching arXiv for the primary source on ArtBulb and closely related work to ground the article. ArtBulb is an interpretable, quantifiable framework for AI art copyright judgment that was introduced to determine, in a legally and technically grounded way, whether a set of AI-generated artworks exhibits a protectable, distinctive artistic style and whether it infringes on existing human artists’ styles. The framework is presented in “From Imitation to Innovation: The Emergence of AI Unique Artistic Styles and the Challenge of Copyright Protection” (Jia et al., 7 Jul 2025). It combines a style description–based multimodal clustering method, termed Description-Guided Clustering (DGC), with multimodal LLMs (MLLMs) to operationalize three criteria for distinctive artistic style: stylistic consistency, creative uniqueness, and expressive accuracy. The same work also introduces AICD, described as the first benchmark dataset for AI art copyright annotated by artists and legal experts (Jia et al., 7 Jul 2025).

1. Definition and problem setting

ArtBulb addresses a problem that arises from the increasing legal recognition of AI-assisted works when they meet originality requirements and involve substantial human intellectual input. The central difficulty identified in the underlying work is that systematic legal standards and reliable evaluation methods for AI art copyrights are lacking. ArtBulb is proposed specifically to fill that gap by assessing whether AI-generated images have a consistent style across multiple works, are distinct from existing human artists’ styles, and accurately express the human creator’s prompts (Jia et al., 7 Jul 2025).

The framework is positioned as a bridge between legal and technological communities. In the formulation associated with ArtBulb, copyrightability is not treated as a property of an isolated image alone, but of a set of works claimed to instantiate a style attributable to human creative direction. This suggests a shift from one-off image appraisal toward corpus-level style assessment, where evidence is accumulated across multiple outputs rather than inferred from a single artifact.

ArtBulb also addresses infringement risk. It is not limited to determining whether an AI creator has a distinctive artistic style; it is also designed to test whether that style is too similar to the styles of known human artists or existing AI styles with recognized copyrights. In that sense, ArtBulb functions both as a copyrightability framework and as a style-proximity assessment system (Jia et al., 7 Jul 2025).

2. Legal and conceptual basis

The paper grounding ArtBulb surveys legal practice in China, the United States, the European Union, and other jurisdictions. It reports that courts increasingly accept AI-assisted works as copyrightable if they belong to a recognized category, have tangible form, exhibit originality, and reflect intellectual labor of a human author. The discussion specifically references a Beijing court decision of November 27, 2023, the Chinese “Tencent Dreamwriter” case, the U.S. “Zarya of the Dawn” registration, and Thaler v. Perlmutter as a counterexample for fully machine-authored works (Jia et al., 7 Jul 2025).

From these precedents, the framework distills a working legal notion: for an AI artist, defined as a human using AI, to claim copyright over AI-generated images, the images must embody a distinctive artistic style attributable to that human’s creative input rather than merely reproducing or trivially varying existing artists’ styles (Jia et al., 7 Jul 2025). ArtBulb operationalizes this notion through three criteria.

The first criterion is stylistic consistency. For a set $C_A$ of AI artworks by one creator, with visual embedding function $f_I$, the intra-cluster variance must satisfy

$$\frac{1}{|C_A|^2}\sum_{A_i,A_j\in C_A}\|f_I(A_i)-f_I(A_j)\|_2^2\leq\epsilon_c^2$$

with $\epsilon_c = 0.60$ (Jia et al., 7 Jul 2025). This criterion formalizes the idea that a creator’s style should persist across changing subjects or contents.

The second criterion is creative uniqueness, expressed as differentiation from existing artists. For AI artwork $A_i \in C_A$ and any human artwork $H_j$ in human clusters $C_H$,

$$\min_{A_i\in C_A,H_j\in C_H}\|f_I(A_i)-f_I(H_j)\|_2\geq\epsilon_d$$

with $\epsilon_d = 0.25$ (Jia et al., 7 Jul 2025). This criterion corresponds to a measurable originality condition grounded in style-space separation.

The third criterion is expressive accuracy, which links the output to the human creator’s prompts. Let $C_A^I$ denote clusters from images and $f_I$0 clusters from prompts or descriptions. ArtBulb requires

$f_I$1

with $f_I$2 (Jia et al., 7 Jul 2025). This criterion is intended to capture evidence that human textual direction guided the creative process rather than the output being only loosely related to the claimed input.

These thresholds are described as empirically tuned via ablation to balance precision and recall in classifying works as copyrightable (Jia et al., 7 Jul 2025). A plausible implication is that the framework treats legal originality not as a purely doctrinal abstraction but as a calibrated decision rule over representations, distances, and cluster agreement.

3. System architecture and Description-Guided Clustering

ArtBulb takes as input a set of AI-generated images, the corresponding prompts, and information on the model used to generate them. The pipeline then performs prompt augmentation through entity substitution, regenerating images with the same model to produce a richer set $f_I$3 of AI works for style assessment (Jia et al., 7 Jul 2025). This step is intended to improve statistical reliability by testing style persistence under content variation.

Feature extraction is multimodal. Each image $f_I$4 is embedded through CLIP’s visual encoder $f_I$5,

$f_I$6

and each image is assigned a textual description $f_I$7 generated by GPT-4o, which is embedded through CLIP’s text encoder $f_I$8,

$f_I$9

(Jia et al., 7 Jul 2025). The use of GPT-4o-generated style descriptions rather than only original prompts is significant because the descriptions are said to capture color palettes, strokes, composition, and other style-relevant semantics.

The core technical mechanism is Description-Guided Clustering. In DGC, both the visual and textual embeddings are normalized to unit norm in $$\frac{1}{|C_A|^2}\sum_{A_i,A_j\in C_A}\|f_I(A_i)-f_I(A_j)\|_2^2\leq\epsilon_c^2$$0, and two clustering heads produce soft assignments over $$\frac{1}{|C_A|^2}\sum_{A_i,A_j\in C_A}\|f_I(A_i)-f_I(A_j)\|_2^2\leq\epsilon_c^2$$1 clusters:

$$\frac{1}{|C_A|^2}\sum_{A_i,A_j\in C_A}\|f_I(A_i)-f_I(A_j)\|_2^2\leq\epsilon_c^2$$2

(Jia et al., 7 Jul 2025). These assignment vectors encode style membership probabilities.

DGC aligns image and text views through cross-modal mutual distillation. For each sample $$\frac{1}{|C_A|^2}\sum_{A_i,A_j\in C_A}\|f_I(A_i)-f_I(A_j)\|_2^2\leq\epsilon_c^2$$3, nearest neighbors are computed in both visual and text spaces, and the framework uses a distillation loss

$$\frac{1}{|C_A|^2}\sum_{A_i,A_j\in C_A}\|f_I(A_i)-f_I(A_j)\|_2^2\leq\epsilon_c^2$$4

where $$\frac{1}{|C_A|^2}\sum_{A_i,A_j\in C_A}\|f_I(A_i)-f_I(A_j)\|_2^2\leq\epsilon_c^2$$5 (Jia et al., 7 Jul 2025). Two auxiliary objectives are added: a confidence loss,

$$\frac{1}{|C_A|^2}\sum_{A_i,A_j\in C_A}\|f_I(A_i)-f_I(A_j)\|_2^2\leq\epsilon_c^2$$6

and a cluster entropy loss,

$$\frac{1}{|C_A|^2}\sum_{A_i,A_j\in C_A}\|f_I(A_i)-f_I(A_j)\|_2^2\leq\epsilon_c^2$$7

with $$\frac{1}{|C_A|^2}\sum_{A_i,A_j\in C_A}\|f_I(A_i)-f_I(A_j)\|_2^2\leq\epsilon_c^2$$8 and $$\frac{1}{|C_A|^2}\sum_{A_i,A_j\in C_A}\|f_I(A_i)-f_I(A_j)\|_2^2\leq\epsilon_c^2$$9 (Jia et al., 7 Jul 2025). The total loss is

$\epsilon_c = 0.60$0

with $\epsilon_c = 0.60$1, $\epsilon_c = 0.60$2, and $\epsilon_c = 0.60$3 in experiments (Jia et al., 7 Jul 2025).

The system compares learned AI style clusters against a large reference corpus of human artists’ styles and against existing AI styles already recognized as unique in ArtBulb’s database. It then checks the three legal-technical criteria and supplies the resulting statistics to an MLLM—primarily GPT-4o, with Qwen2-VL and DeepSeek-VL also evaluated—to produce a chain-of-thought explanation and a human-readable copyright report (Jia et al., 7 Jul 2025). The paper states that MLLMs are not used to guess copyright status directly; rather, they interpret the quantitative outputs of DGC.

4. AICD dataset and reference corpus

ArtBulb is evaluated using AICD, described as the first benchmark dataset specifically designed for AI art copyright analysis (Jia et al., 7 Jul 2025). AICD serves two roles: it provides a reference space of human artistic styles and a labeled evaluation set of AI-generated works.

The human reference corpus begins with 567 WikiArt artists with at least 30 works each, then adds 672 film and game illustrators, 593 children’s picture-book artists, 781 contemporary artists, and 172 Chinese painters in traditional styles. After clustering and filtering for clusters where more than 50% of works come from a single artist with clear copyright, the dataset contains 364,297 artworks from 2,785 artists, each representing a distinct personal style cluster (Jia et al., 7 Jul 2025). This corpus is the style reference space against which AI styles are measured for uniqueness and potential infringement.

The AI-generated evaluation subset includes three categories. First are AI works with potential copyrights: 1,786 AI images with known creators and prompts, collected from social media, AI art demos at conferences, and online gallery exhibitions. Five independent artists and legal experts scored each image on a 1–5 scale for confidence of independent copyright, and images with average score greater than 4.0 were retained, producing 542 positive AI artworks (Jia et al., 7 Jul 2025).

Second are AI works without independent copyright: 2,000 images generated from generic prompts such as “generate an image of a dog,” representing cases with no significant creative effort (Jia et al., 7 Jul 2025). Third are potential infringement cases: 5,000 AI images generated using CopyCat’s synthetic method plus prompts referring to specific human artworks, from which experts labeled 1,278 images as clear infringements (Jia et al., 7 Jul 2025).

The annotations are expert-based. Five artists and legal experts rated independent copyright confidence, while separate panels of artists and copyright lawyers selected infringement cases. The paper does not report a numeric inter-annotator agreement such as $\epsilon_c = 0.60$4, but it states that only strongly agreed cases were included through consensus thresholds such as average score greater than 4.0 and “clear infringement” labels (Jia et al., 7 Jul 2025). This suggests a benchmark designed for high-confidence supervision rather than broad coverage of ambiguous edge cases.

5. Evaluation and empirical performance

ArtBulb is evaluated along two tracks: clustering of human artist styles and binary classification of AI copyright status (Jia et al., 7 Jul 2025). In the clustering track, DGC is compared against DEC, DAC, SCAN, SIC, and k-means on MOCOv3, DINOv2, CLIP, and BLIP2 features, using ACC, NMI, and ARI as metrics.

On WikiArt, the best baseline reportedly achieves accuracy around 36%, whereas DGC reaches 41.45% ACC, 61.20% NMI, and 15.72% ARI, outperforming all baselines (Jia et al., 7 Jul 2025). Similar improvements are reported across video game art, contemporary art, and children’s books. The paper interprets this as evidence that style clustering is materially harder than category clustering and that DGC provides a stronger backbone for style-based analysis.

For binary copyright judgment, the evaluation compares pure MLLMs in a naive question-answering mode, classification-based models CSD and ARTSAVANT, DGC alone, and the full DGC with ArtBulb framework (Jia et al., 7 Jul 2025). The reported averages are summarized below.

Method	Average ACC	Average F1
GPT-4o	~0.42	~0.41
CSD	0.77	0.79
ARTSAVANT	0.80	0.80
DGC	0.81	0.79
DGC w/ ArtBulb	0.86	0.88

The full framework achieves average ACC 0.86 and F1 0.88 across style domains (Jia et al., 7 Jul 2025). Domain-specific examples reported include Western art with ACC 0.85 and F1 0.85, comic or line-drawing with ACC 0.88 and F1 0.89, video game art with ACC 0.87 and F1 0.88, and graphic design with ACC 0.83 and F1 0.83 (Jia et al., 7 Jul 2025).

Qualitative case analysis is also central. The paper describes four archetypal outcomes: copyright affirmed when a consistent new style is distinct from human styles and accurate to prompts; no protection due to inconsistency; no protection due to similarity to a specific human artist’s style; and no protection due to prompt mismatch (Jia et al., 7 Jul 2025). ArtBulb is said to focus on style rather than mere content, aligning its decisions with legal reasoning about substantial similarity.

Interpretability is evaluated through Average Human Ranking (AHR), where 10 legal experts and artists rate explanations on a 1–5 scale. GPT-4o alone receives about 2.05, whereas GPT-4o guided by DGC outputs reaches 3.70; Qwen2-VL rises from 1.80 to 3.55 under ArtBulb guidance (Jia et al., 7 Jul 2025). On 19 real infringing images, five nontechnical legal professionals assign an average score of 4.50/5 to ArtBulb reports for ease of use, clarity of language, depth of analysis, and actionable recommendations (Jia et al., 7 Jul 2025). This indicates that the framework’s contribution is not only predictive performance but also report usability.

6. Practical role, interpretability, and institutional use

ArtBulb is described as an expert-support tool rather than a replacement for judicial discretion (Jia et al., 7 Jul 2025). For legal professionals, it offers quantified measures of style similarity and distance, visual examples of AI and nearest human clusters, and a structured report that maps technical outputs to legal concepts: consistency to coherent style, uniqueness to originality and lack of substantial similarity, and accuracy to evidence of human intellectual direction.

For artists and creators, the framework can test whether several works actually form a recognizable style cluster and whether AI outputs are close to an existing human style cluster (Jia et al., 7 Jul 2025). Creators whose works satisfy the criteria can use the generated report as documentation when asserting rights or seeking registration. This suggests a documentary function in addition to a classificatory one.

For platforms and regulators, the framework is proposed as a backend service for proactive detection of style-mimicking content, flagging possible infringements, and logging style clusters associated with specific users or models (Jia et al., 7 Jul 2025). Regulators could also use ArtBulb or AICD for policy testing, such as estimating how often a model produces works that cluster too close to known artists.

Interpretability is supported through cluster visualizations, textual style-feature breakdowns, and explanation reports (Jia et al., 7 Jul 2025). Because the MLLM receives structured outputs such as intra-cluster variance, distance to nearest human clusters, and image-text AMI, the system’s explanations are anchored in explicit measurements rather than unrestricted narrative generation. A plausible implication is that ArtBulb aims to reduce the evidentiary gap between embedding-space metrics and legal argumentation by explicitly narrating how one maps into the other.

7. Limitations, risks, and broader significance

The paper identifies several limitations. Copyright doctrine is jurisdiction-dependent, and ArtBulb is described as being inspired heavily by Chinese and U.S. practice rather than formally adapted to every jurisdiction’s standards (Jia et al., 7 Jul 2025). The thresholds $\epsilon_c = 0.60$5, $\epsilon_c = 0.60$6, and $\epsilon_c = 0.60$7 are empirical and dataset-dependent, so different artistic domains may require retuning. The reference corpus, although large, is not exhaustive and emphasizes certain art forms, which may bias results against under-represented styles or traditions (Jia et al., 7 Jul 2025).

There are also sociotechnical risks. Over-reliance on an automated framework could lead platforms or agencies to treat ArtBulb scores as definitive rather than advisory. The dynamic addition of AI styles into the reference repository may make future AI artists less likely to be judged unique, raising questions about saturation of style space and about who gets to “own” a stylistic neighborhood (Jia et al., 7 Jul 2025). The paper further notes concerns about bias against less-represented communities and about the possible normalization of AI styles that sit near human styles.

The broader significance of ArtBulb lies in its attempt to concretize legal notions such as originality, substantial similarity, and human intellectual contribution into measurable criteria backed by a benchmark and an explanation layer (Jia et al., 7 Jul 2025). This does not settle the normative question of how copyright law ought to treat AI-generated art, but it supplies a reproducible framework for analyzing that question in operational terms. In that respect, ArtBulb belongs to a growing class of systems that transform open-ended cultural and legal disputes into multimodal inference problems with explicit thresholds, datasets, and reporting structures.