Note: This will likely be out of date within a year or two of my writing it. This was my take in September of 2026.

Writing manuscripts on paleontological subjects, especially Pennsylvanian-aged invertebrates, is not easy. The field is underfunded, and the age of discovery has long passed for today’s workers. There are, without a doubt, countless new species to find and describe, but the amount of physical and mental work required is much greater than it was 100 years ago.

If you go back nearly 200 years (1835), the first-ever report on North American Pennsylvanian-aged marine invertebrates shows how far we’ve come in terms of writing. T.A. Conrad published a report on five new species of “fossil shells” found in Pennsylvania, using hand-drawn figures with probably exceptional writing quality at the time; yet it leaves much to be desired today. The author had no requirement to label the specimens or deposit them at any institution. They are nowhere to be found or compared today. But with time comes progress, and with progress comes today’s modern rules and safeguards around publishing.

Earlier LLMs, Newer LLMs

Modern large language models in artificial intelligence are less than a decade old and have arguably improved rapidly in just the past couple of years. Using them is tempting for reducing the mental work of producing manuscripts, but they create a lot of new mental work to keep them in check. Most commercial models are trained on Internet text, and many search the web when they answer, and authors like me publish a lot of information there.

This is easy to test. Petalodus is a common fossil tooth found in Paleozoic rocks, but an LLM-assisted Google search or a ChatGPT query will return information that places a lot of weight on text I’ve written on this website. I’ve done my best to research Petalodus and seek peer review when I can, but LLMs trust and use my content because I’ve arguably written the most about the subject online. However, this was early 2026 LLMs.

Late 2026 LLM Capabilities

An LLM is, quite frankly, a tool. The difference between you wielding Microsoft Word vs. someone else is your skill in writing and using the software. When I started this website in 2019, I chose this technology because I was comfortable using it. It made publishing easy. But I also made mistakes. I made many typos and poorly worded sentences. Well, I still do.

But today’s LLMs can run on your machine and connect to the website. They can read everything you’ve ever written on the website and find all the silly mistakes. I wrote Petalous (sic) somewhere? Found it. The reference says 2004 in the text and 2003 in the reference list? Found it. Do you want me to fix all of these? Sure. A few seconds later, it’s done.

The research side is even bigger. Just a couple of years ago, scanning PDFs with an LLM was a challenge, but the latest models can quickly turn most documents into easy-to-scan text and draw useful inferences. You can connect it to your email and have it review the countless emails you sent to experts on the subjects. It can find every fact, opinion, and detail you’ve forgotten, as long as it’s written down somewhere.

Do you want to publish your research in a journal? Tell it which one; it will pull the style guide and guide you through an interactive process for formatting, references, media guidelines, and more. It can compare your work with recently published work in the same space and give you tips and ideas. You still provide all the answers, but you don’t have to remember everything all the time.

Leaning on Technology

Early AI interactions made me sour. They were often wrong, and I kept pointing out every flaw I could find. But arguably, the flaws have started to taper off. They aren’t gone, but they are becoming harder to find in any frontier model out there. If you are so critical of modern AI that you dismiss it, you might as well make fun of Microsoft Word while using a typewriter.

Where It Still Sucks

Back to the tool analogy: AI used by people who either don’t know how to prompt well or are quick to accept poor results leaves many with a bad taste in their mouths. I attended a Labor Day celebration with some of the worst AI imagery I’ve seen. If there was a stereotypical example of AI, they were using it. People who used to settle for nothing are more likely to create something today. It won’t be great, but human thoughts can be fabricated into tangible media faster than ever.

Who Will Read the Research of Tomorrow?

Who will tomorrow’s research papers be for? I often can barely read through an entire manuscript myself unless the subject is directly in my field of interest. Artificial intelligence will read it all, but I hope the abstract can draw me deeper into the paper if I need to. The photos are still fantastic; each figure is really worth a thousand words. But should the research and information be optimized for machines rather than humans? What is research for, besides advancing and clarifying the world’s knowledge?

Did I Write This?

Yes, most of it. Grammar checkers reword a few sentences for me, but this is still me.

In the meantime, check out what I could produce with AI in about an hour of prompting.

  • This is a web app, but I could port the same code to iOS and Android if I wanted to. https://fossil.15656.com/app/
  • My formation pages needed some cohesion. How about an interactive stratigraphic link footer? Look at the bottom of the page. https://fossil.15656.com/local-geology/brush-creek-limestone/
  • I figured, why not. I prompted AI to create an inline embed to showcase one of my favorite pages on the site. It created this while I wrote the article. It is below.