Something interesting happened to the scientific paper this week.
Researchers published a system called Paper2Agent that can take a research paper, together with its code, data and methods, and turn it into an AI agent capable of answering questions, running analyses and even working with agents created from other papers.
That is impressive in its own right.
But I think there is a bigger idea hiding inside it.
What if the paper was never really the knowledge?
What if it was simply the presentation layer?
We have been flattening knowledge into documents
A scientific project contains much more than prose.
It contains evidence, data, methodology, code, claims, relationships, uncertainty, interpretation and provenance.
Then we compress all of that into something like:
paper.pdf
The same thing happens everywhere.
Product knowledge becomes a manual.
Business performance becomes a board paper.
Financial activity becomes an annual report.
Company knowledge becomes a website.
Analysis becomes a dashboard.
Documents became dominant because humans needed information organised and presented in a form they could consume.
That distinction did not matter very much when humans were always the final user.
AI makes it matter.
The document is not necessarily the knowledge. It is often a representation of the knowledge designed for a human.
AI exposes the awkwardness of the current model
Look at how we commonly build AI systems today.
We take structured knowledge, flatten it into documents, extract the text, break it into chunks, create embeddings, store those embeddings, retrieve a few fragments and then ask a language model to reconstruct what the original material meant.
In simplified form:
knowledge → document → extract → chunk → embed → retrieve → reconstruct meaning
That works surprisingly well.
But viewed from a distance, it is also slightly absurd.
We are taking knowledge that may already have structure, relationships and provenance, turning it into a presentation format for people, and then spending enormous effort teaching machines how to reverse-engineer the structure back out.
A more natural architecture might be:
knowledge → human presentation
and at the same time:
knowledge → machine interface
The PDF, page, report or dashboard then becomes one possible view of the underlying knowledge rather than the primary computational home of that knowledge.
What does that look like in practice?
Consider a product manual.
Today we might have an 86-page PDF covering products, components, symptoms, faults, warnings, repair procedures and parts.
An agent would probably rather have direct access to those objects and their relationships.
It might need to understand that a particular symptom relates to three likely faults, that one of those faults applies only to a specific product revision, and that a repair procedure requires a particular part and carries a particular safety warning.
The manual can still exist.
But it can increasingly be generated from the underlying knowledge rather than being the only place that knowledge lives.
The same applies to financial reporting.
A report might present revenue, margin, geography, period comparisons and management commentary in tables and charts.
An AI agent would benefit more from direct access to the metrics, periods, entities, calculations, source data and provenance behind those charts.
And a research paper can be thought of in the same way: evidence, methods, datasets, code, claims, citations and interpretation, with the publication being one formal representation of that work.
But a document does more than display information
This is where the argument needs an important correction.
A document is not merely formatting.
An author decides what to include, what to exclude, what order the reader should encounter the material in, what evidence matters, what deserves emphasis and what conclusion should follow.
A good paper, report, article or presentation therefore contains something beyond data.
It contains an argument.
Presentation is itself information.
Two people can start with the same underlying evidence and produce very different documents without either of them being dishonest.
Imagine a business where revenue is down, margin is down, customer retention is up and a new product is growing quickly.
One executive might write:
Short-term profitability is deteriorating and requires intervention.
Another might write:
The business is successfully trading short-term margin for growth and retention.
The underlying numbers are the same.
The interpretation is not.
That interpretation is part of the intellectual work, and it should not disappear just because AI can access the raw information directly.
Humans have always packaged evidence with interpretation
Historically, much of that synthesis had to happen before information was distributed.
A government publishes a 200-page report.
A journalist reads it and writes a 1,000-word article.
I read the article.
The information has moved through several layers:
evidence → report author's interpretation → journalist's interpretation → reader
Each layer selects, compresses and frames what came before.
That has been necessary because human attention is scarce.
Most readers cannot inspect every source, every dataset, every methodological choice and every competing interpretation before forming a view.
So someone else does part of that work for us.
AI changes where interpretation can happen
This may be one of the more interesting consequences of separating knowledge from presentation.
If an AI can access the underlying evidence, methodology, provenance and the author's own interpretation, some synthesis can happen when the information is consumed rather than only when it is published.
Instead of receiving only:
evidence → author's interpretation → me
I can increasingly have:
evidence + provenance + author's interpretation → my AI → me
That allows a different type of interaction.
I can ask:
What evidence most strongly supports this conclusion?
Which assumptions does it depend on?
What evidence weakens it?
What did the author choose to emphasise?
What plausible alternative interpretation fits the same evidence?
The author's view remains important.
But it is no longer necessarily the only lens through which I encounter the underlying material.
We may be moving toward a world where more understanding is formed locally, at the point of consumption, from evidence plus competing interpretations rather than being fully packaged in advance.
Raw data is not truth
It would be easy to take this argument too far.
Separating knowledge from presentation does not mean that raw data somehow becomes objective truth.
Data itself contains choices.
Someone decided what to measure, what not to measure, how categories were defined, how information was collected, which records were excluded, how missing values were handled and what level of uncertainty was acceptable.
Many important ideas are not reducible to rows in a database at all.
So the destination is not:
biased documents → objective raw data
It is something more useful.
We can preserve the layers separately while keeping them connected:
evidence → provenance and methodology → claims and relationships → interpretations → presentation
Today we often collapse all of those into a single artefact.
Tomorrow we may be able to preserve the structure underneath it.
The immutable record still matters
There is another reason documents are not going away.
They provide a record.
A signed contract must preserve exactly what the parties agreed to.
A scientific paper must preserve exactly what the authors published.
A regulatory filing must preserve exactly what was disclosed.
A board paper may need to show exactly what information and recommendation directors received at a particular point in time.
If AI dynamically constructs a new explanation every time I ask a question, we still need something authoritative underneath it.
This suggests an important distinction:
The record should be immutable. The view does not have to be.
A published paper can remain the immutable record of what the authors claimed.
An AI-generated explanation of that paper for a geneticist can be one view.
An explanation for a student can be another.
An agent comparing the paper with ten other studies can create another.
All of those views should remain traceable back to the original record and, where possible, to the underlying evidence.
In that sense, documents may become more important as records even as they become less important as the primary computational representation of knowledge.
AI may also change who controls the presentation
There is another consequence that I think is easy to underestimate.
A great deal of human communication depends on controlling the sequence in which someone encounters information.
A presentation starts with the problem, establishes its scale, introduces evidence, presents a solution and ends with a recommendation.
A report decides which findings appear in the executive summary and which are buried on page 87.
An article decides which fact appears in the headline and which receives a single sentence near the end.
This is not inherently manipulative. It is how humans build arguments and persuade one another.
But an AI intermediary can disrupt that sequence.
If someone sends me a 50-slide proposal, my agent does not necessarily need to experience slides one through fifty in the order the author intended.
It can tell me:
Here is the recommendation. These are the three main pieces of evidence supporting it. These assumptions are necessary for the argument to work. This evidence weakens the conclusion. These alternatives were not discussed.
That subtly shifts some control over information from the person presenting it toward the person consuming it.
Persuasion will not disappear.
Stories, speeches, essays, presentations and carefully constructed arguments will continue to matter because humans respond to far more than raw evidence.
But for analytical information, AI may make it easier for the recipient to interrogate the argument rather than simply receive it in the form chosen by the author.
The document becomes a view
This suggests a different way of thinking about content systems.
The canonical object does not have to be a PDF, page or presentation.
It can increasingly be a richer knowledge object containing evidence, provenance, relationships, claims and interpretations.
From that, we can produce different outputs for different purposes.
A formal publication.
A signed record.
A website.
A report.
A chart.
An AI explanation.
A conversation.
The original author's document remains important because it preserves authorship, argument and history.
But it becomes one authoritative view over something richer.
Paper2Agent may be showing us the transition
This is why I find Paper2Agent interesting beyond scientific publishing.
It starts with the world we already have:
paper → reconstruct knowledge and capabilities → agent
That is valuable because there are billions of documents already containing important knowledge.
But if we were designing information systems from scratch for an AI-native world, perhaps we would increasingly start from the other direction:
knowledge → agent
and:
knowledge → paper
The same underlying research could be available to machines as structured evidence, code, methods and tools while still producing a traditional paper for publication and historical record.
That model applies well beyond science.
It applies to manuals, policies, reports, dashboards, websites, contracts and much of the content organisations create.
We may have been storing the presentation rather than the knowledge
For most of computing history, we have stored information in the form we wanted people to consume it.
That made sense because people were the users.
AI introduces another consumer that can interact much closer to the underlying structure, evidence and capabilities.
That does not mean the document disappears.
We will still need publications, signed records, reports, presentations and carefully constructed human arguments.
But their role may change.
The document can become a view of the knowledge rather than the place where the knowledge itself lives.
And once we make that distinction, a great deal of the information architecture we have taken for granted starts to look very different.
Sources
Stanford Medicine: AI agents that can collaborate across research papers