Innovation Is a Raw Draft

Large language models are often accused of being incapable of creating anything new. The argument is familiar: they are trained on existing material, they learn patterns from the past, and they generate plausible continuations. Whatever they produce may appear original, but underneath it is still rearranged history.

There is some truth in this. Ask a language model for ten startup ideas and it will often give you ten polished variations of familiar structures: a marketplace, a personalized assistant, a platform connecting one group with another. The ideas sound complete because the model knows what ideas are supposed to sound like. That is one of its strengths, but it is also one of its limitations.

The criticism, however, rests on a strange assumption: that human innovation begins somewhere outside history. Most human ideas are also made from existing material. New products combine familiar technologies. Scientific theories connect observations that were previously kept apart. New genres borrow from old genres. New interfaces rearrange existing behaviours into a different system.

The smartphone did not invent the camera, the telephone, the touchscreen, the map, or the internet. Its novelty emerged from bringing those things together in a form that reorganized everyday life. Innovation rarely has no ancestry. It usually begins when familiar things enter an unfamiliar relationship.

So the useful question is not whether large language models reuse old patterns. Of course they do. The question is whether those patterns can be rearranged into a possibility that did not previously exist, and whether that possibility can survive contact with reality.

The finished answer is the wrong place to look

We tend to judge AI by its finished outputs. We ask for the final email, the final design, the final strategy, the final code, or the final answer. Then we inspect the result and ask whether it contains genius. Usually it does not. Usually it contains something reasonable.

But innovation rarely begins as a finished answer. It begins as a fragment, an irritation, an unexplained observation, a bad analogy, or a connection that is interesting before it becomes useful. It begins as a raw draft.

A finished AI response often hides this stage. It smooths away uncertainty, resolves contradictions too early, turns an unusual thought into a familiar paragraph, and makes an idea sound more complete than it really is. That is useful when the goal is communication, but it can be destructive when the goal is discovery.

The strange part of an idea is often the most fragile part. It sounds awkward, lacks evidence, and does not yet fit the surrounding logic. It is easy to remove because it cannot defend itself. Polishing may improve the sentence while quietly erasing the possibility inside it.

A possibility machine

Large language models are often described as prediction machines. For creative work, it may be more useful to think of them as possibility machines. Their most interesting capability is not producing one brilliant idea on command, but generating variations around a problem at unusual speed.

Some of those variations are obvious, some are useless, and some are impossible. Others reveal assumptions hidden inside the original question. Occasionally, one opens a direction that a person would not have found alone.

This is less like a machine having a eureka moment and more like accelerated mutation. The model proposes, a person notices, a simulator tests, a piece of code fails, an experiment rejects the idea, another version appears, and something unexpected works.

The innovation emerges across the system. It may belong partly to the model, partly to the person who framed the problem, partly to the tools that tested the result, and partly to the external world that determined whether the idea was real. This is a less dramatic story than either of the popular extremes. AI is neither a passive copy machine nor an independent genius. It is a participant in a process of variation, selection, and revision.

Reality is the editor

The weakest AI ideas remain inside language. They sound good because language models are good at producing things that sound good. The strongest examples move beyond language.

A model proposes an algorithm, and the algorithm runs faster. It proposes a molecule, and the molecule is synthesized and tested. It proposes a mathematical construction, and the construction can be verified. It writes a program, and the program performs better than the existing approach.

At that point, it matters less whether the system understands originality in the human sense. The idea has entered a world where eloquence is insufficient. Reality becomes the editor.

This may be the essential condition for AI-assisted innovation. The model needs an environment capable of saying no. A chatbot can continue producing increasingly convincing nonsense. A chatbot connected to simulation, code execution, proof checking, or physical experiments can have its ideas rejected.

That rejection is part of creativity. Without selection, there are only variations. With selection, some variations become discoveries.

The danger of professional language

Language models are trained to produce plausible language, and plausible language tends to converge towards recognizable forms: the professional proposal, the startup pitch, the research abstract, the strategy document, and the thought-leadership essay. The more polished the output becomes, the more it begins to resemble the average of everything that already succeeded.

This creates a strange effect. AI may increase the number of ideas while reducing the visible differences between them. Everyone becomes more articulate. Every project acquires a value proposition, three target audiences, and a roadmap. Every strange concept is translated into the language of platforms, assistants, ecosystems, and marketplaces.

The roughness disappears, and so does some of the personality. More importantly, some of the conceptual accidents that might have led somewhere else disappear too.

A strange project becomes “a platform.” An unusual social relationship becomes “a marketplace.” A speculative tool becomes “an AI-powered assistant.” Once named this way, the idea begins to behave like the category. The draft closes too early.

Using AI backwards

Most AI products are designed around completion. They promise to remove blank pages, uncertainty, and unfinished work. The user provides a prompt, and the system provides an answer. This makes sense for productivity, but it makes less sense for invention.

A system designed for innovation would not only ask for the best answer. It would help keep several answers alive. It might ask what assumptions are hidden inside the problem, what would make the problem disappear, which part of the idea sounds wrong but remains interesting, what someone from another field would notice, which directions cannot all be true, what could be tested today, and what should remain unresolved.

Instead of collapsing a field of possibilities into one polished response, it would preserve tension between different drafts. It would expose disagreements, remember discarded paths, and allow an unfinished idea to remain unfinished long enough to develop its own shape.

This is where the human role becomes especially important. Humans choose what matters. They recognize emotional weight. They notice that an apparently useless idea connects to an old frustration. They continue with a direction despite weak early evidence. They decide which problem deserves months or years of attention.

A model can generate possibilities. It does not automatically know which possibility should become a life.

Raw drafts as an innovation practice

A raw draft is more than a bad version of a finished piece. It is a different kind of object. It preserves options, contains paths that have not yet been resolved, allows contradictions to sit next to each other, and records what someone was trying to understand before they learned how to explain it.

This makes raw drafts more important in the age of AI. When polished text becomes nearly free, polish becomes less informative. It no longer proves that someone spent a long time thinking. It may only prove that someone knew how to request a coherent result.

The valuable material may therefore move earlier in the process: the fragments, the prompts, the rejected connections, the moments where the person disagreed with the model, and the idea before it acquired the vocabulary used to sell it.

Creative work may increasingly be judged through the quality of its search process. What possibilities were explored? What was rejected? What changed the direction? What remained unresolved? What surprised the person making it? The draft becomes evidence of thought.

Innovation is not a purity test

There is a temptation to treat originality as a purity test. If an idea was influenced by existing work, it is considered less original. If AI participated, it is considered less human. If a system generated many candidates, the winning candidate is treated as less meaningful.

But innovation has always depended on tools, collaborators, institutions, and accumulated knowledge. A microscope changes what can be discovered. A spreadsheet changes what can be calculated. A search engine changes what can be remembered. A programming language changes what one person can build.

Large language models change the number of conceptual drafts a person can generate, compare, and test. This does not guarantee better ideas. It creates a larger field of possibilities.

That field can become repetitive noise. It can also contain a path that would otherwise never have appeared. The result depends on how the tool is used, what evaluates it, and whether someone is willing to remain with unfinished material.

The unfinished advantage

The most important AI-generated idea may not arrive looking important. It may appear as a bad answer with one good sentence, a misunderstanding that reveals a better question, a ridiculous combination that becomes practical after one assumption is removed, a hallucination that is useless as a fact but valuable as a design direction, or a fragment that causes someone to remember something unrelated.

These moments are difficult to measure because the model did not produce the final invention. It produced the draft that changed the direction of the person who eventually did.

Perhaps this is the wrong moment to obsess over whether machines are creative in isolation. Humans are rarely creative in isolation either. The more useful question is what kinds of human-machine environments produce work that neither side would have produced alone.

The answer will probably not be found in the final response box. It will be found in the unstable space before the answer, where old patterns are rearranged, weak signals are preserved, and possibilities remain open.

Innovation is not the absence of history. It is what happens when history is drafted differently.

AI authorship note

This essay grew from Kris Haamer's questions about whether large language models can produce genuine innovation or only repeat old patterns. AI supported research, drafting, editing, and implementation. Kris directed the argument toward Raw Drafts, identified unfinishedness as the central frame, and repeatedly pushed the piece away from fragmented statements and toward sustained paragraphs before publication.

Connected reading

• Software for One Afternoon
• The Infinite Canvas Is No Longer the Hard Part
• Open Models and the Rest of the Work
• The Internet Is a Studio Wall