Manifold Filling: Does AI Really Create, or Does It Just Recycle the Past?

Over the past two years, artificial intelligence (AI) has made enormous progress in creativity, generating striking works of art, composing music and even writing poetry.

Let’s explore this concept in more depth, analyzing why AI is not fundamentally inventive and putting into perspective the differences between human intelligence and AI when it comes to producing genuine innovation.

By deconstructing AI’s technical feats, we will highlight the intrinsic limits that prevent it from achieving true creativity comparable to that of humans.

With impressive performance in image generation, music composition and text creation, one might think AI is on its way to surpassing humans as a creative force. But is this perception well founded?

A concept often mentioned in AI research is “manifold filling” (filling the data space), which allows AI models to learn and generate content based on pre-existing patterns. At first glance, this can give the impression that AI is capable of innovating and creating something new. However, on closer analysis, science shows us a very different reality.

What does “manifold filling” mean in AI?

In simple terms, manifold filling refers to the way AI models generate outputs by interpolating and extrapolating from existing data. Unlike humans, AI does not pull its ideas out of thin air; it analyzes and combines known patterns to produce plausible content.

The manifold hypothesis in AI

Real-world data (images, text, music, etc.) often lies on a lower-dimensional space (a manifold) within a high-dimensional environment. AI models learn these structures and can generate new outputs that seem original but are in fact interpolations of existing data.

For example, if you ask an AI to generate a “new kind of cat,” it will combine existing features of felines, but it will not invent a fundamentally different creature with an entirely new biology.

This concept is central to certain methods such as semi-supervised learning and Laplacian regularization (Cabannes et al., 2021), where models spread knowledge from labelled points to neighbouring unlabelled points within the underlying manifold. In the study by Fefferman, Mitter and Narayanan (2013), the manifold hypothesis is tested and mathematically validated, showing that high-dimensional data concentrates around low-dimensional substructures (a property called concentration of measure), which implies that AI always works within a pre-existing framework.

In practical terms, this means that AI may seem to produce something new, but in reality it remains a prisoner of statistically observable correlations and regularities.

Imagine you are learning to recognize geometric shapes: if you have a few clearly identified examples of circles and squares, you can guess that other similar shapes are also circles or squares, even if no one has explicitly told you so. AI works in a similar way: it uses labelled (known) data to help classify or interpret unlabelled data, assuming they follow the same underlying logic.

In other words, it “propagates” information from known points to those that are still unknown, based on relationships in the data.

So, does AI really create something new?

The short answer: not really. AI “creativity” is fundamentally different from human creativity. AI is excellent at recognizing patterns, making connections across huge databases and generating content from existing information. It can interpolate styles, optimize solutions and improve pre-existing concepts.

But, unlike humans, it never questions the frameworks within which it operates.

Its reasoning is probabilistic and dependent on the data it has seen.

• ✅ Recognizing patterns — relationships in huge datasets.
• ✅ Interpolating — smooth transitions between known data points.
• ✅ Optimizing — improving ideas based on pre-existing models.

• ❌ Questioning fundamental assumptions.
• ❌ Making radical conceptual leaps.
• ❌ Having intention or intrinsic meaning.

Human creativity vs AI: examples

Art & creativity

Tools like MidJourney or DALL-E can generate impressive images by analyzing thousands of works of art and producing results consistent with existing styles.

However, they do not invent a fundamentally new art movement.

By contrast, a human artist like Picasso was able to break with the conventions of his time to create Cubism, a revolutionary and unprecedented approach.

Science and discovery

In science, AI is a powerful tool for analyzing billions of molecular structures and proposing new drugs by optimizing existing properties.

But it remains unable to rethink scientific paradigms.

For example, while Einstein revolutionized our understanding of the universe with relativity, an AI trained only on Newtonian physics would simply have sought to refine that model without ever conceiving an entirely new theory.

Transformers and NLP

Models such as GPT-4 or BERT (Wolf et al., 2020) generate fluent text by extrapolating learned linguistic patterns, without any real semantic innovation equivalent to inventing a new conceptual framework.

Unlike a human who could invent a new philosophical concept, a new language or a new grammar, AI assembles existing pieces into configurations that seem new but remain entirely dependent on training data.

Creativity by “lucky accident”

Anantrasirichai & Bull (2022) show that what may seem innovative in AI is often a statistical illusion: chance combinations rather than creative reasoning. Perceived novelty frequently results from a random exploration of the manifold, not from a conceptual breakthrough.

What we perceive as novelty produced by AI is often an illusion arising from the probabilistic nature of the model, not a genuine conceptual breakthrough.

Why is this difference fundamental?

AI learns from the past, while humans imagine what does not yet exist. No database led directly to scientific revolutions like quantum mechanics or to cultural innovations like Surrealism: these were human breakthroughs. AI remains a formidable tool to support creativity, not a substitute for human inventiveness.

An AI, however powerful, would never have invented human rights — it would merely optimize existing laws according to the models it has analyzed. Humans, on the other hand, can redefine the questions themselves.

TL;DR

AI explores and optimizes existing knowledge without redefining the foundations. As Colton & Wiggins (2012) point out, a truly creative system should not only generate original works, but also justify and contextualize its output within a new conceptual framework — something current models do not do.

Yet current AI models have no intrinsic understanding of the meaning or cultural impact of what they produce. They do not break conventions; they follow them.

While AI can generate stunning works of art and help accelerate scientific research, disruption and radical innovation still belong to the human realm. For now, AI is an excellent assistant, but not yet a true creator in the human sense of the word. It can remix, improve and combine, but it cannot redefine the rules of the game.

Sources

• Anantrasirichai, N., & Bull, D. (2022). Artificial intelligence in the creative industries: A review. Artificial Intelligence Review, 55(5), 589–656. https://doi.org/10.1007/s10462-021-10039-7
• Cabannes, V., Pillaud-Vivien, L., Bach, F., & Rudi, A. (2021). Overcoming the curse of dimensionality with Laplacian regularization in semi-supervised learning. Neural Information Processing Systems (NeurIPS), 35.
• Colton, S., & Wiggins, G. A. (2012). Computational creativity: The final frontier? Proceedings of the European Conference on Artificial Intelligence (ECAI). https://doi.org/10.3233/978-1-61499-098-7-21
• Fefferman, C., Mitter, S., & Narayanan, H. (2013). Testing the manifold hypothesis. Journal of Machine Learning Research, 14, 3703–3740.
• Lamb, C., Brown, D. G., & Clarke, C. L. A. (2018). Evaluating computational creativity: An interdisciplinary tutorial. ACM Computing Surveys, 51(2), 28. https://doi.org/10.1145/3167476
• Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., ... & Rush, A. M. (2020). Transformers: State-of-the-art natural language processing. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, 38–45. https://doi.org/10.18653/v1/2020.emnlp-demos.6
• Yang, Q., Steinfeld, A., & Zimmerman, J. (2019). Unremarkable AI: Fitting intelligent decision support into critical, clinical decision-making processes. Proceedings of the CHI Conference on Human Factors in Computing Systems, 1–11. https://doi.org/10.1145/3290605.3300468