Tracing the Origins of AI Images

by

Author: Denis Avetisyan

A new approach links generative AI outputs to their training data, offering insights into stylistic influences and potential copyright concerns.

A multimodal large language model leverages domain-specific ontologies to extract semantic triples from images, constructing per-image knowledge graphs and quantifying semantic similarity through node overlap analysis.
A multimodal large language model leverages domain-specific ontologies to extract semantic triples from images, constructing per-image knowledge graphs and quantifying semantic similarity through node overlap analysis.

Researchers leverage ontology-aligned knowledge graphs constructed from images using multimodal models to enable traceable data attribution in generative AI.

As generative AI models grow in sophistication, so too do concerns regarding transparency, accountability, and potential copyright infringement. This challenge motivates the research presented in ‘Training Data Attribution for Image Generation using Ontology-Aligned Knowledge Graphs’, which introduces a novel framework for interpreting model outputs by automatically constructing knowledge graphs from visual data. Leveraging multimodal large language models, the method extracts structured, ontology-consistent information from images, enabling traceable links between generated content and its training data. Could this approach not only facilitate responsible AI development, but also unlock new avenues for creative collaboration and stylistic analysis?

The Illusion of Intelligence: Unveiling the Generative Paradox

Generative artificial intelligence is fundamentally altering the landscape of content creation, spanning text, images, and even code, at an unprecedented rate. However, this power comes with a significant caveat: these systems often function as ‘black-box’ models. Unlike traditional algorithms where the logic is readily apparent, the internal workings of generative AI – particularly deep learning networks – are largely opaque. The models learn complex patterns from vast datasets, but how they arrive at a specific output remains difficult to discern. This lack of interpretability isn’t merely a technical hurdle; it presents challenges in understanding biases embedded within the AI, verifying the originality of generated content, and ultimately, trusting the results. While the outputs can be remarkably creative and sophisticated, the process remains shrouded, prompting a need for new techniques to illuminate the decision-making processes within these increasingly powerful systems.

The increasing prevalence of generative artificial intelligence introduces novel legal complexities surrounding content ownership and responsibility. Because these systems often produce outputs with limited traceability to their training data or internal reasoning, establishing clear authorship becomes problematic. Current copyright laws, designed for human creators, struggle to address scenarios where content is generated autonomously by an algorithm. This ambiguity creates potential disputes regarding intellectual property rights, particularly when generated content infringes upon existing works. Furthermore, determining accountability for harmful or misleading content-such as defamation or misinformation-presents a significant challenge, as assigning liability to the model itself is currently not feasible under most legal frameworks. The lack of transparency in these systems necessitates a re-evaluation of existing legal doctrines to ensure appropriate protection for creators and recourse for those harmed by AI-generated content.

The burgeoning field of generative artificial intelligence demands a concerted effort to illuminate the internal mechanisms driving content creation. Current models, while proficient at producing text, images, and other media, often function as opaque systems, hindering understanding of how specific outputs are derived. Researchers are actively developing techniques – from attention mechanism visualization to probing internal activations – aimed at dissecting these ‘black boxes’ and tracing the lineage of generated content. Successfully unpacking these decision-making processes isn’t merely an academic pursuit; it’s essential for establishing accountability, mitigating bias, and ensuring responsible innovation in a world increasingly shaped by artificially generated outputs. These methods promise to move beyond simply accepting what a model produces, toward understanding and validating the reasoning behind it.

Knowledge Graphs: Structuring Understanding, Not Just Data

Knowledge Graphs (KGs) represent information as entities, attributes, and relationships, offering a significant advancement over traditional data representations that rely on statistical co-occurrence. Unlike methods that simply identify correlations, KGs explicitly model semantic connections – for example, stating that “a red shirt” is a “type of” “clothing” and has a “color” of “red”. This structured approach enables reasoning and inference; a KG can determine relationships not directly stated in the data. The explicit representation of knowledge also facilitates interpretability, allowing users to understand why a particular conclusion was reached, rather than simply receiving a result based on opaque statistical patterns. Data is typically stored as triples: (subject, predicate, object), which define the relationships between entities and allow for efficient querying and knowledge discovery.

Ontology-aligned Knowledge Graph (KG) extraction automates the process of constructing KGs from image data by leveraging pre-defined, domain-specific ontologies. These techniques move beyond simple object detection to establish relationships between visual entities based on ontological definitions. The process typically involves identifying instances of concepts defined in the ontology within an image, and then creating edges between these instances based on the relationships specified within the ontology. For example, using a fashion ontology, the system can identify a ‘shirt’ and ‘pants’ and establish a ‘worn_with’ relationship between them, even if those items aren’t explicitly labeled as being part of an outfit in the image metadata. This approach enables the creation of richer, more semantically meaningful representations of visual content than can be achieved through purely data-driven methods.

Ontologies, such as Fashionpedia and the Lightweight Ontology for Describing Images (LIO), function as formalized vocabularies that define the concepts and relationships relevant to visual understanding. These ontologies provide a structured framework by specifying entities – like clothing items, colors, or materials – and the semantic relationships between them – such as ‘is a type of’, ‘is made of’, or ‘has a part of’. By grounding image analysis in these predefined knowledge structures, systems can move beyond simply identifying objects to understanding their attributes and how they relate to each other, enabling more accurate and interpretable visual reasoning. The use of established ontologies ensures consistency and facilitates knowledge sharing across different visual understanding tasks.

Neo4j is a graph database management system designed for efficient storage and retrieval of highly connected data, making it well-suited for knowledge graphs. Unlike relational databases which require complex joins to navigate relationships, Neo4j utilizes a property graph model where data is stored as nodes and relationships, enabling traversals and pattern matching with significantly improved performance. Data is accessed using Cypher, a declarative graph query language, which allows users to express queries focusing on the structure of the graph rather than implementation details. This architecture supports complex reasoning and inference tasks, as well as scalability to handle large and evolving knowledge graphs, and features ACID compliance for data integrity.

Our system effectively extracts semantic triples from fashion images, constructing a knowledge graph-visualized here as a star due to the single-garment focus of the dataset-based on a provided prompt.
Our system effectively extracts semantic triples from fashion images, constructing a knowledge graph-visualized here as a star due to the single-garment focus of the dataset-based on a provided prompt.

Deconstructing the Black Box: Methods for Model Introspection

Model interpretability refers to the degree to which a human can understand the causes of a model’s decisions. Several techniques are employed to assess this, broadly categorized as either intrinsic or post-hoc methods. Intrinsic methods involve designing inherently interpretable models, such as linear regression or decision trees, while post-hoc methods attempt to explain the behavior of complex, black-box models after they have been trained. Post-hoc techniques include methods like feature attribution, which quantify the importance of input features to the model’s output, and example-based explanations, which identify training samples that most influenced a specific prediction. The choice of method depends on the model’s complexity, the desired level of explanation, and the computational resources available.

Influence Functions and Retraining-Based Attribution methods quantify the contribution of specific training data points to a model’s prediction on a given input. Influence Functions operate by approximating the change in model parameters resulting from removing a training sample, then tracing that change back to its effect on the target prediction. Retraining-Based Attribution, conversely, assesses impact by retraining the model without the specific training instance and observing the resulting difference in prediction. Both techniques provide a numerical score reflecting each training sample’s ‘influence’ – higher scores indicate a greater contribution to the model’s output, allowing identification of potentially problematic or crucial training data.

Embedding-Based Similarity assesses the relationship between a language model’s generated output and its training data by representing both as vectors in a high-dimensional space. Cosine similarity or other distance metrics are then used to quantify the closeness of these vectors; high similarity scores suggest the generated content is closely aligned with examples seen during training. This technique is valuable for identifying potential issues such as memorization of training data, the replication of biases present in the training set, or the unintended generation of content mirroring sensitive or problematic examples. Analysis focuses on identifying nearest neighbors in the training data to the generated output, allowing for direct comparison and attribution of content origins.

The knowledge graphs (KGs) generated by our methodology incorporate approximately 15 distinct relationship types, indicating a complex and detailed representation of semantic connections. This relational diversity surpasses many existing approaches that rely on fewer, more generalized relationships. The presence of these multiple relationship types allows for a more nuanced understanding of the data and the model’s internal reasoning, facilitating detailed analysis of how different concepts are interconnected and contribute to the model’s output. The specific relationships captured enable a granular assessment of the model’s knowledge and potential biases, contributing to improved interpretability and trustworthiness.

Unlearning experiments were conducted to assess the efficacy of the proposed method in removing the influence of specific training data points. Results demonstrate performance levels comparable to those achieved by established latent-space unlearning techniques. Specifically, the model maintained comparable accuracy on held-out test sets after ‘unlearning’ targeted data, indicating that the method effectively reduces the model’s reliance on those specific samples without significantly degrading overall performance. This parity in performance, when contrasted with latent-space approaches, validates the method’s ability to achieve effective data removal and suggests its suitability for applications requiring data privacy or model correction.

LatentExplainer is a post-hoc interpretability technique designed to generate natural language explanations for the predictions of sequence-to-sequence models. It operates by learning a latent representation of the input sequence and then decoding this representation using a separate language model, conditioned on the model’s internal state at the time of prediction. This process produces a human-readable explanation that highlights the features of the input deemed most relevant by the model. Unlike attention mechanisms which only indicate where the model looked, LatentExplainer aims to articulate why a particular decision was made, offering a more comprehensive and understandable account of the model’s reasoning process. The generated explanations are typically expressed as complete sentences, making them accessible to users without specialized knowledge of machine learning.

Knowledge graph analysis enhances image similarity assessments by not only confirming high similarity (0.82 baseline score) but also explaining it through shared attribute relationships highlighted by green edges.
Knowledge graph analysis enhances image similarity assessments by not only confirming high similarity (0.82 baseline score) but also explaining it through shared attribute relationships highlighted by green edges.

Beyond Imitation: Towards Responsible and Controllable Generation

The capacity to distill and replicate nuanced artistic styles, such as that of Studio Ghibli, hinges on the ability to move beyond simple keyword association and delve into the underlying semantic structure of an aesthetic. Ontology-aligned Knowledge Graph (KG) extraction offers a method for achieving this, by identifying and representing the core elements – recurring themes, visual motifs, character archetypes, and emotional tones – that define a particular style. This process doesn’t merely catalogue what defines ‘Ghibli Style,’ but establishes how these elements relate to one another, creating a network of interconnected concepts. By mapping these relationships within a KG, the system can then generate content not simply containing Ghibli-esque elements, but exhibiting the coherent stylistic fingerprint characteristic of the studio’s work, demonstrating a pathway towards computational creativity grounded in artistic understanding.

The capacity to refine pre-trained models through ‘Local Training’ represents a significant advancement in generative AI, enabling nuanced stylistic control without necessitating exhaustive retraining. This technique allows for focused adaptation using comparatively small, specialized datasets – effectively teaching a model to emulate a particular aesthetic or creative voice. Rather than overwriting the extensive general knowledge embedded within the foundational model, local training subtly adjusts its parameters, prioritizing the characteristics present in the limited dataset. The result is a system capable of generating content that adheres to specific stylistic guidelines, such as the distinctive visual language of Studio Ghibli, while retaining its broader understanding of the world and avoiding the pitfalls of overly specialized or incoherent outputs. This approach offers a pragmatic pathway towards responsible generative AI, allowing creators to guide the process without sacrificing the model’s inherent capabilities.

A crucial step towards controllable generative AI lies in verifying the connection between generated content and intended stylistic choices, and recent advancements demonstrate a measurable ability to do just that. Through style-induced triple matching, the system achieved 2.63% accuracy in linking generated outputs to specific stylistic elements, representing a statistically significant result. This process involves identifying relationships – “triples” of subject, predicate, and object – within both the generated content and a defined stylistic ontology, then assessing the degree of alignment. While seemingly modest, this level of accuracy provides a quantifiable metric for stylistic fidelity and opens avenues for refining models to more consistently embody desired aesthetic qualities, moving beyond purely visual appeal towards semantically grounded stylistic control.

Analysis of the knowledge graphs generated by the system reveals a robust level of semantic coherence, as demonstrated by the consistent presence of 5 to 7.1 shared attributes within retrieved groupings. This suggests that the model doesn’t simply associate keywords, but rather constructs relationships based on underlying conceptual connections; for example, concepts relating to ‘forest spirits’ consistently share attributes such as ‘mythological creature’, ‘nature dwelling’, and ‘benevolent entity’. This degree of shared characteristic indicates the model effectively captures nuanced meanings and establishes a network of related ideas, moving beyond superficial associations to build a more comprehensive and logically structured understanding of the desired aesthetic or thematic elements.

The convergence of ontology-aligned knowledge graph extraction and localized model training unlocks a powerful synergy in generative AI, allowing for the production of content that is not only aesthetically pleasing but also deliberately shaped by specified artistic parameters. This approach moves beyond simply generating plausible outputs; it enables the creation of works demonstrably linked to desired styles – in this case, the distinct visual language of Studio Ghibli films. By grounding the generative process in a structured understanding of artistic elements and then refining the model’s focus through targeted training, the system consistently produces content exhibiting coherent semantic groupings – an average of 5-7 shared attributes within retrieved knowledge graphs confirms this – and a measurable alignment with the intended creative vision. The result is a framework for responsible generation, offering a pathway toward AI that respects and reproduces artistic intent with increasing fidelity.

The development of generative AI necessitates a shift towards accountability and ethical practice, and progress in stylistic control represents a crucial step in that direction. By gaining insight into the model’s internal reasoning – how it associates visual features with aesthetic qualities – developers can move beyond simply generating content to understanding and directing its creative choices. This level of control isn’t merely about achieving a desired look; it addresses fundamental concerns about authorship, originality, and the potential for unintentional biases embedded within the generated output. Ultimately, the ability to reliably align stylistic output with defined parameters fosters a more responsible approach to AI, enabling the creation of content that is not only innovative but also demonstrably aligned with creative intent and ethical considerations.

The experimental pipeline leverages visual characteristics of the “Ghibli World” (left) to derive a Δstyle metric (right) represented by blue nodes indicating semantic features.
The experimental pipeline leverages visual characteristics of the “Ghibli World” (left) to derive a Δstyle metric (right) represented by blue nodes indicating semantic features.

The European Union’s proposed AI Act signals a pivotal shift towards regulating artificial intelligence, particularly concerning transparency and accountability. This landmark legislation categorizes AI systems based on risk, with high-risk applications – such as those impacting critical infrastructure or fundamental rights – facing stringent requirements. Developers will be obligated to provide detailed documentation, conduct rigorous testing, and ensure ongoing monitoring to demonstrate compliance. A core tenet of the Act is fostering public trust by mandating clear information regarding the data used to train AI models and the reasoning behind their outputs. By establishing a legal framework that prioritizes responsible AI development, the EU aims to mitigate potential harms and unlock the technology’s benefits while safeguarding citizen rights and promoting innovation within defined ethical boundaries.

The proliferation of generative AI models has instigated complex questions surrounding copyrights and intellectual property, demanding a robust legal framework to balance innovation with creator rights. These systems, trained on vast datasets often containing copyrighted material, can produce outputs that closely resemble existing works, blurring the lines of originality and authorship. Establishing clear guidelines is therefore paramount; these guidelines must define the extent to which AI-generated content infringes upon existing copyrights, and clarify ownership of these novel creations. A thoughtfully constructed legal landscape will not only safeguard the livelihoods of artists, writers, and other creators, but also incentivize further development of generative AI by providing a predictable and stable environment for investment and creative exploration. Without such clarity, the potential for legal disputes and stifled innovation remains significant, hindering the responsible and beneficial integration of this transformative technology.

Addressing the legal complexities surrounding generative AI demands a proactive shift towards systems designed for interpretability and responsible output. Rather than reacting to legal challenges as they arise, developers are increasingly focused on building models where the reasoning behind generated content can be understood and audited. This includes techniques for tracing the origins of training data and identifying potential biases embedded within the algorithms. Furthermore, prioritizing responsible generation-through methods like reinforcement learning from human feedback and the implementation of safety filters-mitigates the risk of producing harmful or infringing content. This approach not only reduces legal exposure but also fosters greater public trust in these powerful technologies, paving the way for wider adoption and innovation.

The trajectory of artificial intelligence development increasingly emphasizes a holistic approach, moving beyond sheer computational power to prioritize systems that are inherently understandable and demonstrably accountable. This shift acknowledges that the true potential of AI is unlocked not simply through complex algorithms, but through the ability to trace its reasoning and assign responsibility for its outputs. Researchers are actively exploring techniques – such as explainable AI (XAI) and robust auditing frameworks – to illuminate the “black box” of neural networks and ensure alignment with fundamental human values. Ultimately, the sustained advancement and societal integration of AI depend on building confidence through transparency, fairness, and a commitment to ethical considerations, fostering a future where these powerful technologies serve humanity’s best interests.

The pursuit of understanding generative AI necessitates a ruthless paring away of obfuscation. This paper champions that principle. It builds traceable connections – data attribution – via knowledge graphs, revealing stylistic influences within generated images. This echoes G.H. Hardy’s sentiment: “Mathematics may be compared to a box of tools.” The tools – here, multimodal LLMs and knowledge graphs – are only useful if applied with precision to dismantle complexity. The work focuses on revealing underlying structures, much like reducing a problem to its essential components. Abstractions age; principles don’t. Every complexity needs an alibi, and this research offers a compelling one for the ‘black box’ of image generation.

The Road Ahead

The construction of traceable provenance for generated imagery, as demonstrated, merely clarifies the starting point, not the destination. The method reveals what influenced a generation, but offers little on how that influence was enacted, or, crucially, why. The knowledge graph, however meticulously built, remains a descriptive artifact, not a generative principle. Future work must move beyond identifying stylistic echoes and address the mechanisms of stylistic translation within the generative model itself.

The problem, predictably, will not be technical. Establishing a definitive link between data point and generated artifact will be simple compared to navigating the inevitable legal and ethical thickets. The notion of ‘attribution’ itself requires careful dissection; is it sufficient to identify a contributing image, or must a demonstrable causal link – and, consequently, responsibility – be established? Simplicity, in this instance, is not an option.

Ultimately, the value lies not in perfecting the cataloging of influences, but in stripping away the illusion of creation ex nihilo. The model does not invent; it remixes. The focus should shift from celebrating novelty to understanding constraint. The fewer elements required to achieve a recognizable style, the more profound the insight. The art, it seems, lies not in what is added, but in what is left out.

Original article: https://arxiv.org/pdf/2512.02713.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-12-03 20:44