Header: AI is making analytical reports cheap enough that they become tools for thinking, not just finished products.
(Note - this essay generated by Chat GPT 5.5; fifth draft; substantial prompting.)
(This essay began when Stephane Budel/Decibio showed me a 10-page Claude Opus report in April.)
When Reports Become Part of Thinking
From Index Medicus to ChatGPT Reveals the Falling Friction of Knowledge Work
I have lived through two revolutions in knowledge work. The first unfolded over several decades. The second seems to be unfolding in just a few years. Although separated by nearly half a century, I think they are fundamentally the same story.
As an MD-PhD student in the early 1980s, performing a comprehensive medical literature search was a significant undertaking. Younger physicians and scientists may find it difficult to imagine how much effort it required.
In 1980, the standard tool for a student was Index Medicus: rows of encyclopedia-sized volumes that occupied many linear feet of library shelves, printed in extraordinarily fine type. A serious literature review meant you sat for hours in front of them. You searched these indexes by hand, copying references ink and paper, and then locating the bound journals themselves, up and down the levels of the library.
By the mid-1980s, technology had advanced, but the process remained cumbersome. The National Library of Medicine maintained the MEDLINE database, but few investigators searched it directly. Instead, one met with a medical librarian, described the project, and the librarian formulated and executed the search. Several days later, after a charge of perhaps $50 to a research grant, a stack of paper arrived containing titles and abstracts (probably from a dot-matrix printer.) It was a wonderful service, but it was also slow, expensive, and infrequent.
Then came the Internet.
In 1997, the National Library of Medicine made MEDLINE freely searchable through PubMed. That event transformed biomedical scholarship.
Not long afterward, around 2000, journals increasingly provided downloadable PDF articles. Interlibrary loans, photocopiers, and mailed Xerox copies gradually faded from scientific life.
Then, during the following two decades, open-access publishing expanded dramatically, placing millions of scientific papers within immediate reach of anyone.
From Burden to Reflex
Looking back, the remarkable thing was not simply that literature searching became cheaper. It became ordinary. Scientists no longer asked whether a literature search justified the time and expense. They simply searched whenever another question occurred to them. The search itself had become part of the thinking process.
LLM: The Speed of Change
I believe something similar is happening today with large language models. Three years ago, asking an AI to write about a topic might produce several competent paragraphs. Today, with one or two sentences of prompting, the best systems routinely generate five- or ten-page reports, complete with headings, tables, comparisons, summaries, and coherent organization.
Recently I asked ChatGPT to analyze approximately 650 AMA Proprietary Laboratory Analyses codes, or PLA codes. The assignment was intentionally difficult. The codes represent a heterogeneous collection of laboratory tests developed over many years.
I did not suggest categories or organizational principles. The model responded by constructing a useful ten-part taxonomy entirely on its own. More interestingly, it recognized that certain features—such as methylation testing, RNA-based assays, and algorithm-driven analyses—were not categories at all but cross-cutting characteristics that appeared throughout the taxonomy. It had distinguished between a classification system and a new second layer of attributes that spanned the classification. It wrote about those, too, all on its own.
Thirty years ago, producing that first draft would have required a day of concentrated work. Today, it required only a minute.
 |
| click to enlarge |
This is where I think we often misunderstand what AI is changing. The important product is not the final report. It is the intermediate intellectual artifact: a background memo, a literature synthesis, a taxonomy, a comparison table, a first draft. These are rarely the final products of professional work. They are the tools experts use while they are thinking.
As the effort required to produce these artifacts approaches zero, they cease to be special projects. They become routine. Some people worry that if reports become inexpensive, knowledge workers will simply generate thousands of reports. History suggests otherwise. Scientists did not begin performing thousands of pointless PubMed searches after 1997. Instead, for the first time, literature searching rapidly became integrated into ordinary scientific reasoning. Researchers explored more hypotheses, verified more assumptions, checked more references, and became more intellectually ambitious because one important source of friction had largely disappeared.
I suspect reports will evolve in exactly the same way. A consultant will no longer wonder whether it is worth spending two days producing a background memorandum. Instead, a report will be generated almost automatically, read critically, refined, challenged, discarded if necessary, and replaced by a better one. The report becomes conversational rather than ceremonial. It becomes part of thinking.
This does not diminish the value of expertise. Quite the opposite. When first drafts become inexpensive, judgment becomes more valuable. The experienced physician (or strategy consultant) recognizes which studies matter. The experienced scientist notices the missing control experiment. The experienced consultant recognizes that one organizational framework reveals more than another. Artificial intelligence increasingly supplies organization, synthesis, and speed, but human expertise supplies judgment, priorities, skepticism, and context.
Seen this way, today’s AI revolution resembles the arrival of free PubMed more than it heralds the arrival of an autonomous scientist. PubMed did not replace scientists. It removed one of the major sources of friction in scientific inquiry. Large language models may be doing something analogous for many forms of intellectual work.
The revolution is not that AI can write a report. The revolution is that reports themselves are becoming inexpensive enough to become part of the thinking process. That is a subtle distinction, but I believe it is the one historians will eventually remember.
###
Below: Illustration Ver1. Nice, but harder to read on a phone.
