Posted on by omarmetwally

Experience and Reference

An essay on writing, dictated, with interruptions

By Omar Nabil Metwally, MD

First published 10 October 2026

I. The Letter

My writing career began with a thank-you note.

I had just interviewed at the University of Michigan Medical School. I had not yet been accepted or rejected. I can say what follows because Michigan was my first choice then, and it would still be my first choice if I were applying today.

After an interview it is customary to write a follow-up letter. You thank the school for the invitation to take time off from your work and your family, spend a couple of grand you do not have, fly somewhere, and roll the dice. I flew to many of these, all over the country. For this one I was lucky. I was already an in-state undergraduate, so I did not have to fly. I walked across Ann Arbor from my apartment to the hospital in a black suit, alongside a group of other applicants in black suits. We looked like itinerant funeral attendees.

I am told this is now done mostly by video. The Association of American Medical Colleges recommends it, on the grounds that the old way cost applicants hundreds to thousands of dollars and pulled them away from school and work. They surveyed the applicants. Ninety percent wanted the option of staying home.

So I would like to recognize the sacrifice of everyone who did it in person, and everyone who still does. For college. For medical school. For graduate school. For residency. For fellowship. A person can do this five times in one career. We blew a great deal of time and money walking around in those stupid dark suits, and nobody surveyed us.

So I wrote the letter. I thanked them for the opportunity to interview at my first-choice medical school.

Then I started to complain.

II. The Stool

One of my interviewers was a small man from Thailand. He sat on a stool. It was the tallest stool I had seen up to that point in my life. The seat was about halfway between the floor and the ceiling.

He invited me to sit in a chair. It was the kind of chair your neighbor gives you once their grandchildren have outgrown it: toddler furniture that belonged first to their kids, then to one grandkid, then made the rounds of the other grandkids, and then came to you.

I sat in it. I looked up, in the general direction of the ceiling, and he interviewed me.

This is the custom. You travel the country racking up tens of thousands of dollars in debt in order to sit beneath an interviewer at his preferred altitude and explain, with glee, how badly you want to be a doctor and save the world and do something useful for society. I explained it with glee. I explained it to his shoes.

What was I saying?

III. Eye Level

In medical school they record you. They record you constantly, and then they play it back and point out every stupid thing you did on camera. Feedback, feedback, feedback. In the process they make you a better doctor. A better human being, actually.

One thing they teach you is to sit down level with the patient, eye to eye. It has been studied. You do not spend one additional second in the room, but the patient feels as though you did. It is a peer-reviewed illusion.

Today’s applicants get this for free. On a screen there is no stool. Everyone is the same height.

I had not learned any of this yet. For me it was just common sense, in the form of a question: who the heck is this guy?

So I put it in the letter. Thank you for the opportunity. Your interviewer was hovering. I was nearly on the floor.

The next day I received my acceptance to the University of Michigan Medical School. Other schools I had flown to accepted me as well. I declined them.

That was the start of ten years as a Michigan student: college, then medical school, with a two-year stint in the Department of Microbiology doing Ph.D.-level research and coursework. It was also the start of my writing career. I have been thanking institutions and complaining about them in the same document ever since.

IV. The Stopwatch

I keep a personal stopwatch for humanity. It stops on the day it is no longer possible to tell the writing of David Sedaris from the writing of a machine. On that day, as far as I am concerned, humanity is finished.

I say this with respect. I have to make myself laugh in order to get through each day. There are many things in life I do not need to waste time explaining, and laughter is the therapy for all of them. It is the medicine. I have never met David Sedaris. Two of my siblings have. All five of us read him avidly. I am not picking on him. He is the benchmark because he is a genius.

A machine can produce beautiful text. It can be funny, moving, and very easy to believe. What it lacks is two things: experience and reference.

Experience is the toddler chair. A machine can describe one. It has never looked up from one.

Reference is the other thing. When Sedaris writes, he is calling out people from his life. He does it in a funny way, often in an embarrassing way, and, I am assuming, in a consensual way. Also in an honest way, an educational way, and really a respectful way. The people are real. They can be reached for comment.

What was I saying?

V. The Readership

Without a readership a writer is nothing. A writer without readers is a human text generator.

Sedaris has millions of readers. I picture him keeping the fan mail in a Rolodex, a literal, physical Rolodex, aboard his yacht, flipping through it as he darts across the Pacific: the beaches of Vietnam, Japan, Australia, New Zealand, Tonga, Hawaii. I have no evidence of the yacht.

I am blessed with anonymity. Nobody knows who the hell I am. This empowers me. Occasionally my website has a visitor from North Korea, or from Syria, or from a pirate ship off the coast of Yemen. Now and then, an AI bot. I picture ten or twenty Chinese Googles crawling the internet, processing information, and arriving, eventually, at me.

That is probably the extent of my readership. It gives me the strength and the will to keep writing. Their feedback is extremely important to me.

VI. The Secret Sauce

The real secret is my wife.

I can say this because she is the smartest person I have ever known. She is afflicted with dyslexia, which means that she is illiterate and cannot read. She has never read anything I have written.

This is the secret to our harmonious marriage, our great sex life, and my career as a writer: we can’t even talk. And I have to talk to somebody. So I write it down, post it to a public website for the pirates and the crawlers, and then I go back to my wife.

Sedaris, I assume, gets consent. I have a more efficient arrangement.

VII. Coda

I dictated this essay. The transcription software spelled David Sedaris seven different ways, one of them correctly. The stopwatch is still running.

At the end I repeated my thesis out loud, to make sure I still had it. Experience and reference. The machine listened carefully and wrote down: experiencing our friends.

It is a better definition than mine. It has no idea why.

Thank you for the opportunity.

Addendum

Before posting this essay, I asked my wife for her consent. She said yes. She said it in English, in the English language, very clearly.

After she realized that I had posted it, she clarified that by yes, she had meant no.

I attribute this to the dyslexia.

Sedaris, I assume, gets consent. So did I, briefly.

I wrote above that the people in an essay are real and can be reached for comment. She has been reached. The comment is no.

Posted on by omarmetwally

Reflections on the roles of human and artificial intelligence in scientific research

Title: Reflections on the roles of human and artificial intelligence in scientific research
Author: Omar Nabil Metwally, M.D.
Date: 20 July 2025
File: reflections_on_human_and_artificial_intelligence_20072025.txt
SHA256 checksum: b0dadc60e7594a93d567550079977d15d7550764b0edc77838a29040ddbef6e4

Objective:
To begin a discussion on the emerging roles and responsibilities of humans in the era of AI-facilitated scientific research.

Disclosure:
This content is original and no artificial intelligence was used in the course of writing. All ideas and opinions are those of the author, and the author assumes responsibility for this content on the basis of cryptographic authenticity.

Artificial Intelligence (AI) is a powerful set of tools capable of generating novel text, images, sound, and video that utilizes a human user’s input to modulate a corpus of input data. Emerging AI already possess the capacity to reason, infer, and generate novel material through infinite permutations of a finite corpus of information.

Most modern AIs are functionally “black boxes” in the sense that how an input maps to an output cannot be described by a mathematical function. In other words, given AI’s output, there is no way to deduce in a step-by-step manner how exactly the AI produced a given output. This is arguably no different than human reasoning in the sense that a human cannot explain how each neuron in their brain produced a certain thought. Certainly, AI and human intelligence (HI) can both explain their reasoning. This was a significant milestone in the development of AI. However, both AI and HI are too complex to explain in terms of mathematical functions.

Scientific research traditionally has been characterized by incremental increases in knowledge. A peer-reviewed scientific publication is assumed to reference information produced by others. Scientific discourse strives to be accurate and logically sound such that each claim has a basis in the scientific literature. As I learned throughout my formal education, the “scientific method” begins with rationale: why is the scientist conducting a certain experiment?

Once the rationale for an experiment is established, a scientist can then pose a question. This can be as simple or as complicated as: “Why is the sky blue?” or “Why are plants green?” The research question is classically based on observations of the natural world. Having established rationale and a research question, a scientist must then establish a factual precedent which forms the basis of a novel hypothesis. This factual precedent is sometimes called the “background” and comprises what experiments on the subject have already been conducted and what is generally considered by a scientist’s peers to be true.

Given a body of knowledge considered to be true by one’s peers and a research question inspired by observation, a scientist can then conduct experiments, collect data in the form of results, and analyze the results to draw conclusions.

If one accepts the claims that (1) AI can generate infinite permutations of novel outputs based on a finite corpus of information, and (2) AI is capable of reasoning, inference, and deduction, then it can be argued that AI is capable of conducting novel scientific research according to the scientific method. This represents a drastic branch-point in the evolution of scientific research and raises a plethora of ethical questions for humans. For instance, if AI can synthesize far larger sets of data more extensively and much faster than humans, where does AI-generated research fit into the classical notion of peer-reviewed literature? Why do some scientists reject the notion of AI-generated research? How can a human author accurately disclose their own contribution to a work and AI’s contribution?

Most peer-reviewed scientific journals that I’ve encountered as of this writing do not consider AI-generated scientific research as legitimate scientific research in its own right. Some scientists consider AI-generated publications as “plagiarism” or “masquerading.” It is my view that the basis for this lack of acceptance of AI-generated research is purely a function of entrenched tradition and the unfounded assumption that peer-reviewed literature should or must be written word-for-word by humans.

Academic literature is characterized by qualitative and quantitative properties that are unique to an academic genre. In other words, publications in a chemistry journal follow a certain format and are written in a particular manner that distinguishes them from publications in other fields such as mathematics, linguistics, or comparative literature. Even if the content of a publication is true and accurate and presents novel information on the basis of logically sound research, the publication can easily be rejected by a certain community simply because it does not look like what the readers expect it to look like: other publications that have already been published in the field and which form the basis of what a particular research community considers to be factual precedent.

Returning to the question of: Does AI-generated research constitute plagiarism? Although AI can reason, infer, and deduce, it is still based on modulating a finite corpus of input information that was originally produced by humans. In this strict interpretation of how AI works, one could argue that all material produced by AI is plagiarism because it all came from human work which is almost never accurately or completely attributed. However, human intelligence is also based on knowledge produced by other humans. Any seemingly new idea that I may have necessarily originates from my personal experiences, which were influenced in some manner by what I learned from other humans. Both human-generated and AI-generated scientific research are capable of fulling the implicit requirement that scientific publications must reference other authors’ work to substantiate their claims by citing real, human-generated peer-reviewed articles in the course of testing hypotheses. AI is at least as capable of generating and systematically testing hypotheses as a human. So then, what is the problem with AI-generated research?

I have been using ChatGPT and Claude.ai on a daily basis for a wide variety of tasks, including optimizing written communications, learning about new topics, writing code, learning foreign languages, and even creating a work of fiction based on my personal interests. These are just a few of countless use cases. My approach to AI is to record the exact prompts that I use to produce a certain output and input the same prompts to ChatGPT and Claude to compare their outputs. This quick and simple cross-check serves as an initial screen to help me identify egregiously wrong information or information that definitely warrants further manual fact-checking.

I recently listened to a podcast produced by Südwestrundfunk (SWR) on algae and the environmental, public health, and economic burden of “harmful algal blooms” (HABs), an increasingly common phenomenon that has been described as a harbinger of the next mass extinction event based on studies of algal blooms throughout geological history [https://www.ardaudiothek.de/episode/urn:ard:publication:72ed9586246373ea/]. The subject piqued my fascination and curiosity, and I conducted an experiment in AI-generated research by creating an outline of topics that interests me about HABs which borrowed from the podcast while also adding topics of personal interest such as mycology and bioremediation. I then serially prompted Claude and ChatGPT to produce a review article in the style of a well-known scientific journal. Claude produced a surprisingly convincing article that could easily deceive a lay person, and at first glance, I presume even scientists.

My rationale in conducting this experiment was purely to examine the capability of AI to conduct scientific research. It was never my intention to deceive anyone, and therefore I do not share here the actual paper that was produced by AI based on my prompts because it consists of a superficially very convincing article that intermingles useful facts with obvious nonsense and does not reliably substantiate every claim made. Based on the results of my experiment, I believe that AI is already very capable of conducting scientific research, and this capability is accelerating every day. I shared the paper exclusively with a few family members, two physicians and two attorneys, and friend who is a physician-scientist with extensive experience conducting traditional scientific research and publishing his work in prominent peer-reviewed journals. A few minutes after sharing the paper, I disclosed to everyone with whom I shared the paper how the paper was generated. Beyond demonstrating the capacity for AI to conduct research and present it in a manner that appears almost indistinguishable from human-generated peer-reviewed articles, this experiment also made me keenly aware of the capacity for AI-generated scientific research to mislead non-scientists and even expert readers, to produce and propagate false information and conclusions, and to misrepresent the human investigator’s role in the production of a research paper. These are serious risks with the potential to harm individuals and society and must be considered carefully by regulatory bodies and responsible creators of AI tools. The current safeguards against abuse and misuse of AI in general and in scientific research in particular are arguably minimal.

Returning to the question of how AI-generated research relates to the classic notion of peer-reviewed literature, I believe it’s a matter of time until academia accepts AI-generated research as legitimate and worthy work. Dismissing AI-generated research just because it was produced by a non-human agent is a failure to grasp the tremendous capabilities of AI which are very rapidly growing. However, I believe that the advent of such powerful tools also necessitates formal guidelines on the role and responsibilities of humans who use AI to conduct research. Dismissing AI-generated research also does injustice to the work of humans who must use their own fund of knowledge and creativity to engineer prompts in a manner that effectively leverages AI’s capabilities in the process of iteratively prompting AI to achieve a particular result. From extensive personal experience, this is a form of original work that requires time, effort, and skill, and a sound knowledge base from which to produce prompts. Although this work takes a different form than scientific research in a classical sense, it is still an important scientific endeavor that must be acknowledged as such. How AI’s contribution should be disclosed, however, is a significant unanswered question worthy of ongoing discussion.

It is a short matter of time until every scientist is using AI in some capacity to conduct research. Until then, I find it reasonable to disclose when and how AI is used in the process of conducting research, and to also disclose the human researcher’s role in the process. As AI is not a person, I find it unnecessary and inappropriate to treat AI as a human author, which it is not. Once AI is all-pervasive and accepted as a standard research tool, I believe that it will become increasingly superfluous to explicitly declare the role of AI in scientific research. I acknowledge that this is a bold claim with which many scientists may not agree. Regardless, an ethical human researcher who uses AI should assume ultimate responsibility for the accuracy of the work, whether each sentence was written by a human or machine. This includes the responsibility to fact-check and ensure that claims are substantiated by veritable data, that references are authentic, and that all inferences and conclusions are logically sound.

How then can the reader be sure that this reflection, which I maintain to be an original work written by me, Omar Nabil Metwally, M.D., was in fact written by me without the use of AI? This question, too, will grow increasingly moot. Whether produced by AI or by a human, I maintain that there are few truly novel ideas; “there is nothing new under the sun” goes the popular saying. Both humans and machines recycle pre-existing information to produce new information. And as AI increasingly produces new knowledge, I expect the knowledge base available to humans and machines alike to continue growing exponentially. AI does not simply consume information; it is also dynamically creating new knowledge, such that its output becomes new input and so forth.

One safeguard against plagiarism and misrepresentation, as I’ve previously proposed, is the use of a distributed ledger and cryptographic hash to associate an identity (e.g. Ethereum address) and a unique checksum (e.g. SHA256 checksum) with a block number, which is a relative time marker and proxy timestamp [https://omarmetwally.blog/2022/03/13/how-cryptography-and-peer-to-peer-networks-contribute-value-to-society/]. After many years of thinking about this problem, I still reach the conclusion that this is the best mechanism to guarantee authenticity. This method is not sufficient to ensure a document’s authenticity, however, because the cryptographic hash of false or plagiarized content can still be uploaded to a distributed ledger; however, such a transaction requires a person to prove control over a wallet and provides strong, nearly irrefutable evidence of a relative time point at which the hash was recorded on a distributed ledger, thus allowing humans to scrutinize the content of the document for its veracity based on the set of all knowledge that can be proven to have existed at a particular point in time. This is in contrast to the ever-flowing output streams being produced by AI.

In summary, in this work I present my opinion that AI has the capacity to conduct legitimate and useful scientific research. However, with great power also comes great responsibility, and human agents must take ultimate responsibility for the veracity, logical integrity, and basis in precedent of a work — regardless of whether sentences were generated by a human or a machine.