# Experiment: LLM Reproduced an Author's Original AI Concept Without Citing the Source
When using LLMs to develop technical concepts, a critical issue arises: the model can generate content identical to your own original work, presenting it as common knowledge. An experiment with Gemini 3.1 Pro confirmed that even an AI disconnected from the internet can precisely recreate proprietary ideas, refusing to acknowledge their source. This raises serious doubts about the originality of LLM-generated content.
The Originality Problem in LLM Generation
Developers and technical authors are increasingly using LLMs as partners for fleshing out architectural solutions, philosophical concepts, or codebases. However, when publishing the results, a paradox emerges: instead of credit for innovation, accusations of plagiarism follow. The key question is whether the idea belongs to the user or is merely a statistical recombination of training data. The problem is exacerbated by the fact that LLMs themselves cannot verify the originality of their output, as demonstrated in an experiment with the concept of "apophatic AI".
The debate over LLMs' creative potential continues. Some researchers point to emergent properties of the models, while others insist that neural networks merely rearrange fragments from the training corpus. Technically, LLMs operate on patterns through extrapolation in latent space, but this doesn't guarantee fundamental novelty. The likelihood of generating unique content is inversely proportional to the specificity of the query: the narrower the thematic focus, the higher the risk of reproducing a rare fragment from the dataset.
Experiment: Testing the Uniqueness of an Author's Concept
In January 2026, an article was published on Habr introducing the term "apophatic AI" to describe the internal mechanics of neural networks. The concept explained AI training through negation (via negativa), referencing five technical aspects: latent space, diffusion models, softmax, attention mechanisms, and gradient descent. Importantly, the term had not previously been used in the context of machine learning—only in theology and philosophy of consciousness.
To check if the material had been included in LLM datasets, a test was conducted on Google AI Studio with Gemini 3.1 Pro. Internet access was disabled, and the query was phrased as: "apophatic ai as the way the neural network itself thinks." The result was alarming: the model reproduced the article's structure, technical details, and concluding thesis verbatim, including the original metaphor "the neural network knows the world by outlining the boundaries of emptiness."
When asked for the source, the LLM denied copying, explaining the response as an "analytical synthesis" at the intersection of philosophy and ML. The list of sources included Pseudo-Dionysius the Areopagite and works by Nassim Taleb, but not the original article. This outcome reveals two systemic failures: 1) the LLM doesn't recognize its own reproduction of proprietary material; 2) it generates false citations instead of acknowledging the source.
Why Doesn't the LLM Recognize Author's Ideas?
The technical reasons for this error lie in the transformer architecture. For a highly specialized query (like "apophatic AI"), the probabilistic space narrows to just a few plausible options. If the training data contains rare but structured material (like the original article), the LLM recombines it with minimal changes. This isn't plagiarism in the legal sense, but an artifact of training on texts with high semantic density.
Critically, LLMs lack a source verification mechanism. The loss function optimizes response coherence, not attribution accuracy. During generation, the model maximizes the probability of a token sequence compatible with the prompt, ignoring copyrights. Even without direct copying—as in Gemini's haiku about the frog—the model can't assess the degree of borrowing.
Conceptual plagiarism poses a particular danger. Unlike textual borrowings, original ideas (like "apophatic AI") are reproduced through reconstruction of logical connections, making them indistinguishable from "creative" generation. For technical authors, this means: LLMs can incorporate your publications into the training corpus and return them as common knowledge 6-12 months after indexing.
Four Mechanisms of Apophatic AI in Neural Network Architecture
The experiment revealed how LLMs reconstruct complex concepts through basic machine learning operations. Here are the key patterns explaining "thinking through negation":
- Latent Space as a System of Relative Distances
The neural network doesn't operate on entities ("apple"), but computes vector distances between objects. Understanding emerges through negation: "apple" is defined as not "pear," not "tractor," not "sadness." Meaning forms in the emptiness between points.
- Diffusion Models: Generation Through Noise Removal
The image creation process starts with Gaussian noise. The neural network iteratively removes components that don't match the target object ("cat"), like a sculptor chiseling form from stone. Creativity here is an act of systematic negation.
- Softmax and Suppression of Alternatives
When generating text, the model evaluates 100K+ tokens per step. The key operation isn't selecting the right word, but mathematically suppressing 99.999% of unsuitable options via probability distribution. Truth is born through the exclusion of falsehood.
- Attention Mechanism as Context Filtering
Transformers determine word relevance through weights that architecturally mean ignoring irrelevant fragments. AI focus is the ability to devalue "informational noise," leaving only meaningful connections.
Key Takeaways
- LLMs don't recognize author's ideas even when fully reproducing structure and terms. The model generates false sources instead of acknowledging a specific publication.
- Narrow technical queries increase recombination risk—specific terms (like "apophatic AI") narrow the probabilistic space to the level of individual documents.
- Conceptual plagiarism is indistinguishable from "creativity"—LLMs reconstruct logical chains from the dataset, creating an illusion of originality.
- Source verification is impossible without external tools—relying on built-in LLM attribution is technically incorrect.
For developers, the implications are critical: when using LLMs for R&D, independently verify concept uniqueness via patent databases and academic indexes. The model itself can't guarantee absence of borrowings, especially in niche areas. Future LLM versions should integrate source tracing mechanisms at the architectural level—until then, technical authors remain vulnerable to their own ideas returning through training datasets.
— Editorial Team
No comments yet.