In 2022, AI-assisted writing still felt like a specialised toolchain. A writer might use a text generator to continue a paragraph, a trend tool to inspect demand, an editor to correct grammar and a publishing system to move a finished draft onto the web. The tools were already useful, but they were visibly separate. You assembled them around a process that was recognisably older than the software itself: find an idea, research it, make an argument, draft, revise and publish.
Since then, almost everything around that process has changed.
Chat interfaces made language models broadly accessible at the end of 2022. Larger models became more capable at following instructions; multimodal inputs made images and documents part of the conversation; tool use and function calling made it possible for a model to participate in workflows rather than only continue text. A writer can now move from a rough question to an outline, source list, counterargument, structured brief, image direction and revision pass without leaving one working surface.
That is a real change. It is not, however, the same as replacing the work of writing.
The Toolchain Collapsed Into a Conversation
The earlier article on the AI-assisted writing process described a collection of specialised tools: research systems, note-taking environments, grammar assistants, text generators, analytics and publishing software. That landscape still exists, but the interfaces have changed.
Today, a strong model can act as a flexible layer between those tools. It can turn research notes into questions, identify gaps in a brief, compare two outlines, create a metadata draft or suggest several ways to frame a difficult paragraph. Connected systems can also retrieve documents, call structured functions or work with code and files. The difference is less “one better writing tool” than a more continuous workspace.
This continuity is valuable. It reduces the friction of moving between thought, research and revision. It also creates a new temptation: because the system can keep going, the writer may start to confuse momentum with progress.
The Process Did Not Disappear
A useful writing process still begins before the prompt.
It begins with an observation, tension, experience or question that deserves attention. “Write an article about AI writing” is not yet an idea. “The tools changed after 2022, but the conditions for making a useful argument did not” is closer. It contains a claim that can be tested, limited and made interesting.
Research follows for the same reason it always did: a promising claim is not evidence. Language models can propose terms, locate possible sources and explain why a source might matter. They cannot fully automate the judgment required to decide whether a source is authoritative, current, representative or being used in the right context.
The model’s fluent answer can even make this harder. It often arrives with the shape of understanding before the underlying evidence has been checked. The writer has to return to the source, inspect the date, distinguish a primary document from a summary, and ask what would count as a counterexample. Research is not a stage that can be completed once and forgotten; it is a series of decisions about what the article is allowed to claim.
This is where the broader shift toward navigation as a form of knowledge matters. As access to possible explanations becomes cheaper, selecting a trustworthy path through them becomes more valuable.
LLMs Have an Expansion Problem
One practical difference between writing with an LLM and writing alone is what might be called an expansion problem.
Give a model an open instruction and it tends to help by adding: more sections, more examples, more caveats, more options, more adjacent topics, more polished transitions. This is not always a defect. Expansion is useful in discovery, when a writer needs to see a wider possibility space. It is useful when a brief is thin, when a topic needs multiple framings or when an argument is still searching for its pressure point.
But expansion becomes harmful when the task is convergence.
A blog post needs a centre. It needs to know what it will leave out. A paragraph needs to stop before it explains every implication. An example needs to serve the argument rather than becoming a second article. A model optimised to remain helpful can treat every open door as an invitation to build another room.
That is the expansion problem: the system produces additional language and additional possibility faster than it produces a reason to choose among them.
The solution is not to demand shorter answers at random. It is to make boundaries explicit:
- State the article’s one-sentence claim before asking for an outline.
- Define the reader, the decision or question the piece should help with, and the limits of the topic.
- Ask for a fixed number of alternatives, examples or counterarguments.
- Give each section a job and remove material that does not perform it.
- Separate exploration prompts from production prompts.
Exploration can be deliberately expansive: “Show me three competing interpretations and the evidence each would need.” Production should be deliberately bounded: “Revise this 180-word section without adding claims, examples or headings.” The writer changes the model’s role by changing the kind of space it is allowed to explore.
Research Cannot Be Fully Automated
Automation can accelerate research operations. It can help collect documents, extract entities, compare recurring claims, summarise a long report or maintain a source table. These are meaningful gains.
Yet research is not simply the collection of information. It is the construction of a relationship between a question and evidence. That relationship depends on context the model does not own: why this question matters now, which source has standing, what the reader needs to understand, which uncertainty is acceptable and who bears the cost of being wrong.
Consider a simple claim such as “AI writing tools became more capable after 2022.” It is easy to generate. A responsible version asks: capable in what way? Compared with which baseline? For which task? At what cost? With what remaining limitation? OpenAI’s 2022 introduction of ChatGPT, the 2023 GPT-4 release and the move toward more steerable, tool-connected interfaces document genuine changes in the product landscape. They do not prove that a generated article is accurate, original or worth publishing.
The writer still has to perform the last mile of interpretation. In many cases, that is the work that gives the article its value.
A Better Division of Labour
The most durable use of LLMs is neither automation theatre nor nostalgic refusal. It is a division of labour.
Use the model to expand the visible space when you need options. Use it to compress material when you need a map. Use it to challenge a weak connection, test an outline, expose a missing assumption or translate a technical source into questions for a general reader.
Keep the human role where stakes and meaning concentrate:
- choose the question;
- decide what evidence counts;
- recognise the difference between a fresh insight and a smooth paraphrase;
- set the boundaries of the piece;
- accept responsibility for the published claim.
This is not a small remainder after automation. It is the process.
The Same Work, With New Friction
Since 2022, the speed and surface area of AI-assisted writing have changed dramatically. The tools are more conversational, more capable, more integrated and more willing to take the next step. That makes them powerful collaborators.
It also means writers need better constraints. The challenge is no longer only overcoming the blank page. It is resisting a page that never stops filling itself.
An idea still has to be found. Research still has to be judged. A good article still needs a point of view and an ending. The tools can make the path faster, wider and more surprising. They cannot decide where it should lead.