Synthesis: the path to real AI gains in writing
AI didn’t solve writing. It solved drafting.
As you've probably heard by now, writing has been commoditized by AI. At least, that's what we're being told. The message to writers is clear: move up the stack because there's no value in writing anymore.
I believe a more accurate statement is that drafting has been commoditized. AI can now get us from a blank page to a usable first draft in minutes. That's true. And getting there used to take a meaningful amount of time and skill, so there is some value there.
But a workable draft is not a finished piece of content. Someone still needs to push it over the finish line, which is no small task due to the fundamental limits of AI today.
So the cost curve has changed, but unevenly. The cost of producing prose has collapsed. The cost of getting to writing that is genuinely good has not collapsed by nearly the same amount.
If the standard is merely “coherent text that says roughly the right thing,” AI has changed the economics enormously. If the standard is a finished piece with strong ideas, reasoning, structure, expression, and factual content, a large amount of human work remains.
In that sense, making something genuinely good still costs roughly what it always did. The human effort has moved around, and AI can accelerate parts of it dramatically, but the expensive judgment remains.
What remains human?
To understand why AI in its current form cannot carry a piece of writing from draft to finish, we must look at two things. First, a breakdown of the primary components of writing and, second, an analysis of the strengths and weaknesses of AI.
Both were covered in my earlier two-part essay, “Why AI writing is still a mess: we're dividing a task we never broke down.”
In part one, I broke writing down into four components: ideas, structure, expression, and mechanics.
In part two, I looked at the strengths and weaknesses of both humans and AI, and proposed who should handle each component.
The division of labor I reached was roughly this:
Ideas are human-led. AI can generate candidate ideas, but humans decide what is worth saying.
Structure and expression divide according to purpose. AI is very good at arranging ideas and language for clarity. Humans, though, must lead the rhetorical side: decisions about how something should affect a reader, what should be emphasized, what should be withheld, which sequence works best, which wording actually lands.
Mechanics can be delegated heavily to AI.
Further, human-led does not mean AI is prohibited from participating. It means the human retains authority over the judgment. AI can suggest the idea, structure, phrase, analogy, opening, or argument. But the question of whether that candidate deserves to survive is still a human decision.
The important realization coming from my previous work is that human creative judgment remains at the center of high-quality writing. This high-value work usually involves deciding whether an idea is worth including, whether an argument really follows, whether a structural choice creates the intended experience, or whether a sentence is merely fluent versus actually right.
What this means is that the bottleneck today has become judgment. AI can do its work quickly and cheaply. It can generate possibilities cheaply. It can transform material cheaply. It can rewrite cheaply. But humans don’t work the same way. In comparison, we’re slow and must exert real effort to perform our part of the job. This means that all the judgment that is still required, the deciding what is good, correct, useful, coherent, appropriate, or finished remains expensive and effortful.
The current assumption is that one person must supply all of that judgment. I no longer think that’s true.
Could judgment be distributed?
If human judgment remains the scarce input, does it necessarily have to come from one human?
Could several people share that load?
I believe they could.
Think of how writers currently work with AI. They provide some input, get a draft, and then start the cycle of reviewing, editing, providing feedback for new iterations, and repeating until done. But why do we assume this is the only way?
We noticed one obvious thing AI is good at: generation. And we put it to work at that right away. But maybe we missed another valuable way it can contribute. Maybe we ignored its ability to organize, combine and synthesize inputs.
The immediate efficiency argument is straightforward. One person currently has to notice everything: the weak claim, the missing idea, the awkward sentence, the unsupported conclusion, the structural problem, the example that does not quite work. If those judgments can be made by several people in parallel, the total burden on any one person falls.
There is also a quality argument. One person has finite knowledge, experience, attention, taste, subject expertise, and ability to notice problems. Several people potentially bring a larger pool of useful judgment.
This doesn’t require everyone to become a co-author in the traditional sense. A person's useful contribution may be extremely small. They may notice one unsupported assumption. They may dislike a phrase. They may supply one fact that changes the argument. They may say that a section is confusing without knowing how to repair it. Yes, they may propose finished wording, but that is only one possible form of contribution.
This matters because judging something can be much cheaper than independently creating the finished result.
If I ask someone to rewrite a weak section from scratch, I am asking them to reconstruct a lot of context, make several decisions, and perform the expression work. If instead they can simply tell me, “The causal claim in this paragraph is too strong; the evidence only shows correlation,” they have supplied the valuable part of the intelligence at much lower effort. Something else can handle the work of integrating that judgment.
That starts to point toward a strange possibility: the route to making AI writing require much less work from any one human may actually involve adding more human input overall.
The original promise of AI writing was that one person could offload most of the work. That has not really happened when quality matters, because the person still has to perform so much judgment after generation. Distributed judgment could get much closer to that promise. The burden comes off the individual because the remaining human work is shared.
Why haven’t people been writing like this all along?
Writing has historically been a single-person production task. Editors, researchers, fact-checkers, subject-matter experts, and collaborators have always existed, but they generally supported a writer who remained responsible for turning everything into one coherent artifact.
That arrangement makes sense because writing is globally interconnected.
What I mean by that is if one person suggests changing an argument in a paragraph, the change may affect the opening premise, terminology used earlier, an example later, and the conclusion. Another person might independently suggest deleting the example. A third might add a qualification that changes the strength of the claim. These contributions cannot simply be stacked on top of one another.
Someone has to understand the whole and decide what the resulting piece should become.
That coordination problem gets harder as the number of contributors increases. The cost of receiving ten people's intelligence is not just the cost of collecting ten contributions. Someone then has to interpret them, detect overlap and conflict, understand their consequences, decide what to use, and rewrite the artifact so it still feels like one piece rather than the output of a committee.
A human can do that, but at meaningful scale the coordination work starts eating the efficiency gains created by distributing the judgment in the first place.
This is probably why genuinely many-to-one writing has never been a normal production model. The problem was not merely finding several people who could help. It was combining their help.
AI changes that constraint.
It is already good at taking unorganized material and turning it into organized material. It can compare large amounts of text, identify relationships, maintain context, rewrite globally, and produce a coherent new version quickly. Those capabilities are close to exactly what is needed between distributed human contribution and a unified piece of writing.
This is the new role I’m interested in. I call it synthesis.
Humans provide local judgment. AI performs global synthesis.
The basic model is that humans make relatively local contributions while AI maintains the global piece of content.
“Local” does not necessarily mean sentence-level. A person might challenge the thesis or propose an entirely different direction. The point is that they do not need to solve every consequence of their contribution themselves. They can supply the piece of intelligence they have.
AI then considers that input in the context of the whole.
Suppose somebody says that one paragraph is repetitive. That is not necessarily an instruction to delete the paragraph. The repetition may exist because the earlier explanation is weak and the later paragraph is compensating for it. The right synthesis might be to strengthen the earlier section and remove the later one. Or both passages may be necessary because they perform different logical jobs despite feeling similar on the surface.
The human contribution is therefore evidence or context for the next version, not a command that has to be executed literally.
This is an important part of the theory. If every human contribution were treated as an edit that must be accepted, multiple contributors would probably make the writing worse very quickly. The system would behave like a document accumulating patches.
Synthesis is different. It asks what the artifact should become given all of the available intelligence.
The AI has to make decisions such as:
- Which contributions are actually useful?
- Which overlap?
- Which conflict?
- Can apparently conflicting contributions be reconciled?
- Where does a useful contribution belong?
- Does accepting it require changes somewhere else?
- Does it reveal a deeper problem than the contributor explicitly identified?
- Does a proposed solution damage something else that was already working?
- What should remain unchanged?
That’s why I think of the division as local human judgment and global AI synthesis. Humans do not have to coordinate with one another or manually merge their work. Their judgments enter a common context, and AI produces the next coherent state.
There is a useful relationship here with my earlier writing framework. AI was already assigned a great deal of organizing work because organizing existing material for clarity is one of its strengths. Synthesis expands that role. The main difference is that the material being organized is no longer assumed to come from one person.
The LLM post-training parallel
There is already a major class of AI systems built around a vaguely similar process.
Modern LLMs are generally trained in two broad stages. The first is pre-training, where the model learns from enormous amounts of text and other data, developing its ability to understand and generate language.
The second is post-training, where that general capability is refined to make the model's responses more useful, accurate, and aligned with human expectations. A major part of post-training involves human judgment. People provide feedback by writing examples, choosing between alternative responses, rating outputs, identifying problems, correcting mistakes, or giving expert judgments. These individual contributions can be quite narrow. The system then aggregates that feedback into a changed model, which is evaluated again. The process repeats as needed.
It’s a similar approach to a different end.
In post-training, the target of convergence is model behavior. Human judgments ultimately influence a new model state.
In our writing theory, the target is one piece of content. Humans inspect a particular evolving piece, contribute intelligence about it, and AI synthesizes that intelligence into a new state of that piece.
Although the targets differ, the underlying process follows the same pattern:
- human judgment is distributed into relatively small contributions
- those contributions are treated as evidence rather than simple commands
- machinery combines them into a new global state
- humans inspect that state again
- the cycle repeats
The analogy does not prove that synthesis will work for writing. There are some meaningful differences between the two processes. Contributors to writing may supply original substance, rather than merely evaluate existing output. Different humans may also have different kinds of authority. Someone's judgment about what they personally mean is not interchangeable with another person's preference, and a subject-matter expert's correction may deserve different treatment from a stylistic opinion.
Still, the existence of large-scale human-feedback systems shows that useful human intelligence does not have to arrive in the form of one person performing the whole task. Small judgments can have value when there is machinery capable of combining them.
That is the part I want to carry across.
There are, however, two important questions that remain unresolved.
Can AI actually synthesize well enough?
At first glance, we have a contradiction. I already argued that important writing decisions require human judgment. Synthesis itself clearly requires decisions. If AI cannot be trusted to make writing judgments, why should it be trusted to synthesize human input?
I think the answer depends on separating two kinds of judgment.
The higher-level creative judgment is about potential value: Is this idea worth considering? Is this criticism meaningful? Does this sentence feel wrong? Is this example stronger? Is there a problem here? Would this direction improve the piece?
That is the judgment I still think humans need to supply.
Synthesis then requires what I would call application-and-fit judgment. Given the current piece and the available human intelligence, what should be done with the contribution? Where does it belong? Is it already represented? Does it conflict with something? What else needs to change if it is incorporated?
These are not trivial decisions, and AI can certainly make them badly. But they are more closely related to the organizing, comparing, transforming, and coherence-maintaining work that AI already does well.
Humans are supplying evidence that something has value. AI is deciding how that evidence should affect the current whole.
The separation will not always be clean. At some point the synthesizer still has to choose between competing human suggestions, and that can involve exactly the kind of creative judgment AI is weak at. So even with this division of responsibilities, AI will sometimes have to make decisions that humans are better equipped to make.
What makes the model plausible to me is iteration.
The AI's synthesis decision does not have to be perfect or final. It produces a new version. Humans can then inspect that version and react to what happened.
If a useful criticism was misunderstood, someone can say so.
If two ideas were merged badly, someone can point that out.
If a synthesis choice weakened the argument somewhere else, that becomes new evidence.
The AI synthesizes again.
This changes the standard I care about. I don’t need AI to be a perfect editor that correctly resolves every contribution on its first attempt. I need the combined human–AI process to move toward a strong result over repeated cycles.
This is consistent with what I’ve already observed in one-human-plus-AI writing. AI frequently gets individual reasoning or wording decisions wrong. The human user notices those failures, supplies more judgment, and the piece improves through iteration.
The unproven step is whether that still works when the useful intelligence is spread among several humans. AI is no longer synthesizing mainly one person's intent plus its own proposals. It has to construct a coherent whole from judgments that may come from different perspectives, contain different assumptions, and occasionally point in incompatible directions.
It’s still unclear whether current AI can do that reliably enough.
Can many people improve writing without averaging it into blandness?
The second major question is whether distributed judgment destroys uniqueness.
There is an obvious failure mode where every contribution pushes the piece slightly toward the middle. One person softens a strong claim. Another removes an unusual phrase. Another prefers the conventional structure. Another dislikes an opinion that feels too sharp. The aggregate result becomes reasonable, polished, and dead.
That would be especially bad because genericness is one of the weaknesses we are trying to solve with AI-generated writing. A many-human system that creates another route to the average would not solve much.
I don’t think it follows automatically that synthesis produces averaging, though. There is an important difference between collecting intelligence and distributing authority.
Many people can contribute without every person having an equal vote over what the final piece means.
One person may notice that a claim is factually unsupported. Another may dislike the claim because it conflicts with their taste. Those inputs should not necessarily carry the same weight.
More importantly, a piece can still have an owner whose intent acts as a governing constraint. Other people can improve the work without acquiring equal authority over its identity.
This may be one of the essential conditions for synthesis to work in creative domains: contribution can be highly distributed while authorship or final authority remains concentrated.
I am not sure exactly how much of the owner's own intelligence has to remain in the artifact. But I do think the amount may be greater than zero.
For writing to genuinely represent someone, there has to be some minimum amount of that person's intent, ideas, experience, taste, judgment, or selection in the process. Otherwise the work may be useful and high quality, but it becomes difficult to say that it expresses that person's thinking.
The amount could become surprisingly small. Perhaps the owner supplies the original intent, a few important ideas, and judgment at a handful of high-leverage points while many other people contribute supporting intelligence and AI handles almost all synthesis and expression.
That would still be very different from today's one-human-plus-AI workflow.
What matters is preserving an asymmetric relationship between contributions. The system should be able to gather broadly without assuming that broad input defines the final voice or meaning.
This also means disagreement is potentially useful rather than something to eliminate mechanically. Several humans seeing the same passage differently may expose a real ambiguity or reveal a choice that needs to be made. Synthesis should not necessarily average the positions. Sometimes the correct global response is to preserve one and reject another.
So the open question is more specific than “Does collaboration make writing generic?” It is whether a synthesis process can take advantage of additional judgment while preserving coherent intent and allowing unique choices to survive.
What becomes possible if the two questions resolve well
If AI can synthesize distributed human judgment well enough, and if the process can preserve uniqueness rather than average it away, then the economics of high-quality writing change much more substantially than they did when AI first became good at drafting.
The first gain is speed. Human judgment can happen in parallel instead of one person serially reviewing every part of the piece.
The second is burden. The person who owns the work no longer has to be the source of every useful observation, correction, example, factual check, or rhetorical decision. Much more of the effort can leave their plate.
The third is potentially quality. A single person has natural limits. Several people can collectively notice more, know more, and bring a broader base of experience and judgment.
If these gains materialize, the cost of going from an initial idea to genuinely strong finished writing could fall substantially. That would be different from the first AI writing breakthrough, where only the cost of getting from an initial idea to a draft fell substantially.
The future of writing
Throughout this series, I've pushed back against the popular narrative that writing has been commoditized. That there's little value left in the profession. But if we’re being completely honest with ourselves, there's a real reason that narrative exists. AI has taken over a large part of what freelance writers, agencies, and writing companies have traditionally been paid to do.
The old model was straightforward. A client provided a brief, a writer produced the article, and the client paid for the finished work. Much of the value was in the execution: turning ideas and direction into finished prose.
AI has made a large part of that execution cheap and nearly instantaneous. So it's understandable why people are questioning the future of writing services.
But as we've established, AI hasn't actually solved writing. It's solved drafting. The difficult work of getting from draft to done remains, and because AI doesn’t hold the judgment required to finish a piece properly, humans still have to do that part of the job.
The difference is that this work is no longer being outsourced in the same way. Instead, much of it has fallen back onto the client, who now generates a draft with AI and then spends their own time evaluating, correcting, refining, and trying to get it to a finished state.
So perhaps writing execution hasn't lost its value after all. Perhaps it's simply changed form.
And if that's true, maybe there's a new version of the writing services industry waiting to emerge.
What if, instead of paying one writer to produce an entire article, clients could pay for smaller contributions of human intelligence throughout the writing process?
And what if, instead of engaging only a single writer, several people could contribute ideas, criticism, corrections, and judgment as a piece evolves, while AI handles the work of synthesizing those contributions into successive versions?
Maybe this becomes the new form of outsourced writing execution.
The problem then becomes figuring out how to organize and manage this new form of writing service. And the industry isn't currently set up to do that.
The current model is built around assigning complete writing jobs and paying for finished deliverables. We don't yet have mature systems for organizing smaller contributions of human intelligence, tracking the value each person creates, and compensating them accordingly.
But if synthesis can deliver the gains I've described, I think that's a problem worth solving.
Since AI arrived, we've been so captivated by its ability to generate that we've focused all of our attention on using it to do the writing for us. The promise was that we'd be able to produce good content faster, cheaper, and with far less human effort.
But here we are, years later, still generating generic content that nobody wants to read, or spending hours trying to turn AI drafts into something worth publishing. We've made enormous gains in drafting, but the gains in the overall writing process haven't been nearly what we expected.
Maybe we've been looking in the wrong place.
Maybe the next major gains in AI writing won't come from generation alone. Maybe they'll come from its ability to organize, combine, and synthesize the intelligence of many humans in ways that simply weren't possible before.