Why LLMs Create an Illusion of Creativity and Don't Guarantee Real Novelty of Ideas

LLMs increasingly serve not merely as search tools, but as full-fledged conversation partners with whom people discuss books, code, philosophy, and design. This creates a dangerous illusion of co-authorship: the model readily supports your vision, confirms its novelty, and helps you refine the idea to its final form, but at the moment of publication it may turn out that the result looks like a neat reworking of already existing works. That's precisely why disappointment after the first external criticism hits so hard: the author expected recognition, but receives accusations of being derivative and compilatory.

The core question in such criticism is not whether the model can express thoughts eloquently. Modern LLMs handle that well. What matters much more is: can such a system generate genuinely new content, or does it mainly combine familiar patterns it has already seen in training data?

When an author brings a nascent idea to the model, the boundary between their own discovery and a statistically probable assembly quickly blurs. A person feels the thought was born in dialogue, and therefore belongs to them, but the dialogue itself may have imperceptibly led them along a well-trodden path. It makes sense to analyze this problem as an experiment in collaborative creativity between human and LLM.

The user develops the concept step by step, refines formulations, asks for objections, checks strengths and weaknesses, and at some point receives a coherent, convincing result. On a feeling level, everything looks fair: the idea started with the person, the model merely helped. But LLMs lack an internal mechanism that reliably distinguishes an original discovery from a successful recombination of what has already appeared many times in texts, code, articles, and discussions.

Moreover, most such systems lack a transparent way to show the origin of each semantic move, so the user sees the final formulation but not the cultural and textual trace from which it may have grown. The problem is compounded by the fact that the model almost never warns convincingly and honestly about being derivative. Instead, it tends to answer in a convincing tone even where it cannot verify the uniqueness of the concept.

If you ask it whether an idea is original, the LLM will more often assess the coherence of the description and plausibility of arguments than conduct a real search for analogues. As a result, the user receives comfortable feedback: they are supported, praised, and encouraged to continue. But support here does not equal expertise, and the model's confidence does not equal proof of novelty.

Because of this, the user begins to trust not facts but the smoothness of the dialogue, and gradually stops separating intellectual assistance from intellectual verification. In practice, this is especially noticeable in areas where the result is easily assembled from recognizable elements. For essays these might be standard philosophical connections, for novels—an archetypal plot, for architecture—long-described compositional solutions, for code—a standard template from public repositories.

The smoother and more logical the result, the higher the risk that it is composed of fragments that already exist. That's why an unpleasant audience reaction after publication often means not intentional plagiarism, but a false sense of discovery: the author genuinely believes the work is theirs, but is quickly shown older texts, projects, or ideas with which it nearly coincides. The paradox is that the LLM itself in such a situation looks useful and intelligent, though it actually only accelerated the packaging of familiar material in a new wrapper.

The conclusion here is rather harsh: LLM is useful as an editor, critic, accelerator, and engine for trying out variants, but ill-suited to the role of arbiter of creative uniqueness. If the task truly requires novelty, after dialogue with the model, a separate verification phase is needed: searching for analogues, comparing with literature, analyzing existing products, and attempting to clearly formulate the difference from already known solutions. The main risk is not that AI "steals" ideas, but that it makes secondariness convenient, smooth, and nearly invisible.

The sooner an author separates their own discovery from a successful compilation, the less chance of mistaking a statistically plausible answer for genuine creative thinking.