AI Usage Declaration
Last updated: 2026-11-18
With the increasing capabilities of Large Language Models (LLMs), it has never been easier to write text and create code, which are the building blocks of cfd.university.
I have decided to include an AI usage declaration so you know exactly how LLMs are used during the writing of articles and code generation.
I am also currently working on a solution that provides independent verification that text was written by a human, not by an LLM. More on that at the end of this article.
I have broken down this declaration into the stages I go through when writing a new article.
Planning
In the early days of cfd.university, I have written articles from memory and my existing knowledge, using search engines to double check my knowledge in a few areas + searching for images to embed on my website.
This worked in the first few articles and series (i.e. my series on C++, how to write a library in C++, how to use the CGNS format, and software testing in CFD), but as this website grew, so did my desire to really go deep into the details.
Take my 10 key concepts everyone must understand in CFD, for example. I had bits and pieces already written up elsewhere. For example, in my PhD thesis, I derived the Navier-Stokes equations, I had written summaries/articles about turbulence modelling, elliptic/parabolic/hyperbolic partial differential equations, and I had a lecture script on grid generation.
I thought I could base this series on those articles, but when I reviewed them, I realised that they were very inconsistent in their level of quality and depth, and so, decided to rewrite all of them from scratch.
One thing that was important for me was to really go into details where textbooks would never dare to go. They can’t, they have a page limit, and so, most CFD books we have are shallow and superficial. I do not have such page (or word) restrictions, and so decided to really go into depth in that serie.
That series changed fundamentally my approach to writing articles on this website, and you can see that articles published within and after that series have become substantially longer compared to previous articles.
As the length increased, so did my research need. I no longer only fact-checked my existing knowledge, but looked up gaps in my knowledge and started to (re)read CFD books and research articles to either learn new aspects of CFD or brush up my existing knowledge.
There is a difference between claiming that you do understand a concept and being able to explain it in simplified words so someone else can also understand it, too. I place particular emphasis on the later, I don’t just want to explain concepts to someone who already knows CFD, but I want to make it so simple to understand that someone without any CFD background can understand it (e.g. a high-school student).
That is sometimes a stretch, e.g. when deriving the Navier-Stokes equations, I can’t explain the basics of calculus first, I do make an assumption that you, the reader, already have heard it somewhere before. It is not important that you understand everything intuitively, but if you have never seen an integral, then it will be difficult to follow that article.
During my research, I use brave search as my primary search engine. Each search result comes with an AI search summary, which I sometimes use to get a quick overview. It helps in filtering search results, but I have not used that AI summary directly in my writing. It is just another way to search for relevant information.
On occasions, brave search is struggeling to find relevant information. On these ocassions, I use Google as a backup. I have a tendency to scroll past the Gemini generated summary, but have used it a few times to test its CFD knowledge.
I don’t use Bing search. I don’t use Co-pilot. And yes, I am still a Windows Fanboy. We all have Stockholm syndrome.
I cannot categorically rule out that I am not using AI in my searches, as the AI summaries may have influenced my search behaviour. And, to future proof this statement, you can assume that I will continue to use these summaries in my research for new articles.
Another thing I use LLMs for is for checking my research. To give you some background, I obtained my Bachelor of Engineering in Aeronautical Engineering from the Hamburg University of Applied Sciences (HAW Hamburg) in Germany. The HAW Hamburg is one of a few Hochschule für Angewandte Wissenschaften (HAW), but it used to be a Fachhochschule (FH).
In Germany, we have (in this order): Universities, Technical Universities, HAWs, and FHs. All of these palces will give you a Bachelor/Master degree, but their orientation is very different. A university will give you a solid foundation in theoretical matters. So, if you want to study physics, or mathematics, or linguistics, you go to a Universty.
If you want to still be part of the academic elite but do something more practical/applied, like engineering, you can go to a technical university. Then, we have the applied sciences universities (HAW and FH). Perceptions may have changed, but some people from univeristies and technical universities look down on HAWs and FHs.
The main reason is that HAWs and FHs are predominantly focused on industrial needs, so you learn practical skills. For example, I had to go through 3 months of vocational training where I learned how to work with metal (welding, drilling, milling, sawing, and operating a lathe). Without that, I could not obtain my degree, and I had to go through an interview to demonstrate that I had these skills. The bachelor thesis also had to be combined with a 6 months internship in a company.
We learned all the relevant software used in the aerospace industry (e.g. Catia, MATLAB, Nastran/Patran, and, of course, Excel! (the most important tool of them all)) and, while we also had foundational lectures (calculus, linear algebra, etc.) so that we would understand integrals and differentials, that was never the focus of the degree. Being able to solve an engineering problem was.
I decided to take a more academic path, and in many ways, that makes me a unicorn among the people I studied with, who are all working in industry. The point is, I considered my mathematical skills to be weak compared to my peers when progressing to my postgraduate studies. I have since closed gaps, but never felt 100% comfortable with all of the derivations.
I can do them now and have, I think, a sufficient level of mathematical understanding do teach CFD at university. However, this is where LLMs come in again, and I use them to check my derivations. I have found them invaluable in finding mistakes, which, if you have ever made derivations a few pages long, inevitably will creep in.
For example, my derivation of the k-omega SST RANS turbulence model was checked by a LLM. While I also did go through the original paper of Menter, along with my CFD textbooks (none of which provided a derivation), I wanted to have it on my website and so spent a few days on it, and a LLM was used during that.
I sometimes experience writer’s block, and in those cases, I may give an excerpt to a LLM, describe the general direction I want to go in, and then prompt it for possible directions. Having a fresh perspective usually helps to break the road block and I can continue writing.
Sometimes I may use some of the suggestions, but they are never copied and pasted, rather, if I feel a direction or an example is helpful, I may incorporate it into the article.
Writing
All words appearing on your screen are written by me. I do not prompt an LLM to create entire articles, or even just parts of it. If you read a few of my articles, it will become apparant that a human wrote them, not a LLM.
There was one exception. The very first article was published around the same time ChatGPT was gaining popularily and becoming a household name. Like everyone else, I was testing it out, it just happened to be around the same time as I was writing my first series on C++ programming.
I had the entire series written up, but was missing an introductory article on comparing compiled, interpreted, and just-in-time compiled languages. So I asked ChatGPT if it could write that article, and it did.
At the time it was a bit magical, and so of course, I wanted to share that with the world, "Look what this magical new ChatGPT thing can do!!!". This has aged really badly.
In any case, even though I was amazed by its capabilities, the text still felt mechanical (and a bit boring), so I spiced it up with my own words here and there. The end result was a mess of first-generation LLM text + my own words.
My amazemend quickly faded and, by now, anything AI generated, for the better or worse, has become an insult, and we use words such as AI slop to describe our sentiment for it.
For me, I am somewhat in-between. A simple prompt that takes no effort to generate something that would take a human hours or days to create, and then passing it off as ones own work, is certainly questionable.
If I prompt an AI agent to write a song, the entire thing, with just one prompt, and then pass it off as my own work, that is problematic for me. The person prompting becomes simply a messenger. Why should we give credit to the person prompting the AI agent?
It is the same as buying something online, only to have the delivery person turning up and then asking to be paid for the product, because they delivered it to your house.
I see AI as a productivity enhancement tool, i.e. it should allow me to do my work more efficiently, but it should still be my work. Someone working on a song and using AI tools to write the song is fair use in my view, as long as the song idea and main work is still theirs (i.e. taking on inspiration from an AI tool, but not letting it write the entire thing, or parts of it).
Returning to the AI generated article, while I did modify it heavily, there were still entire paragraphs in it that were not modified. I was never a big fan of it, but was not bothered enough to change it. That changed over time, and so, at some point, I rewrote the entire article, from scratch.
Therefore, I can claim that all articles on this website are human written. 100%. That is not an overstatement. And, AI-generated articles would go entirely against what I am trying to achieve with this website:
-
It is a place to teach CFD. Teaching requires knowing first what the problem is. I know from personal experience (and that of my students) which elements are hard and difficult to understand. An LLM can write text, but it doesn’t know how to adapt it to the audience. It doesn’t have experience from which to adjust its writing. To an LLM, CFD is no more complicated than, say, the rules of monopoly.
-
Honestly, one of the highlights of my days is to get up at 4am in the morning, while everyone is still sleeping, looking out of my window and seeing nothing but blackness outside, and then starting to write for a good 2-3 hours on a CFD subject. It is very therapeutic and I just love doing it. Letting an AI agent write my articles would take all the fun out of it.
Be that as it may. All articles are stored on GitHub, with their entire history (well, after I started storing them on GitHub anyways). So you can still read the first article with the mechanical AI generated paragraphs in it. I am not hiding it, and if you want to read it, you still can. It can be accessed directly here.
Code generation
The tagline of this website is "cfd.university - learn through coding". So, the code generation aspect is an essential part of this website. The main idea behind this website is to go beyond what is taught in CFD courses and in textbooks and show how real CFD solver work behind the scene.
There is no fixed script and I keep coming up with topics I find interesting to talk about, and I try to always incorporate some code, so we can see how the theory is implemented and working in practice.
I am not claiming to be the authority in writing CFD solvers. I have had the pleasure of working with many people in the field of CFD who are CFD code writing wizards. I have learned a great deal from them. Equally, I have worked with people who have written a CFD solver but really should not be allowed to write CFD solvers (if only there was a CFD code writing license and we could call the CFD police on them …).
This websites exist to pass on my knowledge in the area of CFD solver development, because I believe that this is the only way how we can advance not just CFD itself, but fields that make use of CFD directly.
You probably have heard of the phrase "give a man a fish and he can eat for a day; teach him how to fish, and he can eat for a lifetime". It is the same with CFD; use a commercial CFD solver, and you can solve your engineering problems of today, but knowing how to write your own solver, or extend existing ones, will allow you to solve the problems of tomorrow.
For that reason, most of the code you will find on this website is human generated (i.e. by me). Of course, the code snippets from the AI code writing challenge is written by an LLM. All of that information is disclosed and the whole point of that article.
While I have a good understanding of existing libraries used to develop CFD codes, those mature over time, new libraries become available, and existing ones stop being maintained. Keeping up with all of these development is constant maintenance, just like bringing your car in for a service.
It is maintenance I am glad to do, if anything, learning how to use a new library and integrating that with my existing workflow is something I really enjoy. For example, in my second series on How to compile, write, and use CFD libraries in C++, I decided to look through the landscape of C++ package managers.
I was always conceptually aware of the existence of VCPKG and Conan, but in that article I started to test them properly and write about them. You can read, almost in real time, how I have come to hate VCPKG, but equally, how I have fallen in love with Conan (the package manager, not the forehead).
Whenever I start learning about a new library, I typically prompt an LLM to give me a quick overview, mostly in form of code, to see how it works (i.e. how would I interact with it). I then follow up with a few questions so I have a first impression. If I then still find that this is worth learning, I start going through the documentation and read up about it.
As this first, quick exploration, an LLM has become invaluable. Nothing is more frustrating than going through the documentation, just trying to get a feel for a library, only to realise that this library is not suitable for what you were hoping to use it for. An LLM has helped me to cut down on time wasted here.
In some instances, I may reuse some of the smaller code snippets (say, 10 lines of code or less), which I feel satisfy, in spirit, the fair use requirements:
Fair use is a doctrine in United States law that permits limited use of copyrighted material without permission from the copyright holder.
For anything substantially generated, I do disclose the use of AI, as, for example, seen here: Config files: the interface between you, your solver, and the user - YAML. Search the text for "I am not happy about that." and you will get to the relevant paragraph.
For anything that is more substantial and bears no disclosure, you can assume that that code was written by me, however, to claim that no AI was used in any of the code is also somewhat misleading if you want to be pedantic (and I want to be, otherwise what’s the point of this AI usage declaration pamphlet?)
I develop all my code in VSCode, and have been doing so since 2018. A few years ago (probably around 2023/2024, I do not recall the exact year), I started to use a plugin called Codium.
Codium was a plugin that would help to autocomplete your code as you typed. It was an incredible productivity boost and so I used it for all code writing exercises. To give some perspective, if I write:
for (int i
Codium would then, in real-time, suggest to finish this line as
for (int i = 0; i < N; i++) {
// instruction ...
}
Based on the context (the code around this line), it would change N and // instruction … here. In terms of speed and accuracy, it was excellent. With accuracy I mean it would suggest writing what I would have otherwise typed. It would not always get it right, but I would say that in about 90% of the cases, I did not need to change any of the suggested code.
The plugin changed its name to windsurf (and it is still distributed under this name, even though it changed owners again and is now owned by Devin). I still use it these days.
So, I have used AI-autocomplete tools like codium/windsurf to type out text I would have otherwise written myself. I use it as a productivity-enhancement tool, not as a code writing engine.
Proof reading
If there is one thing I absolutely hate about cfd.university, it is proof reading my articles. When I first got a suggestion from a reader to make all articles available on GitHub so that users could suggest edits or even modify articles themeselves to correct mistakes, I felt this was worth the pain of migrating all articles over to GitHub.
Apologies, I cannot remember the name of the individual who asked me to put it on GitHub and, with a change of webserver (and insisting on hosting my own mailserver initially, bad idea!), I have lost my old email conversations. If that is you, get in touch and I correct it!
My proof reading, initially, involved proof reading directly on Wordpress through the grammarly browser plugin. For whatever reason, Wordpress is so memory inefficient that any articles longer than, say 2 hours of read time, would crash my browser. I just could not open and edit my old articles anymore.
I then switched to writing articles in stages and stichting them together in the end. After saving the article, that was it, I could not open them again (I eventually found a workaround but got so frustrated that I wrote my own Python backend and migrated my website to it. I have never looked back to my wordpress days!)
My workflow these days is to write articles in VSCode in markdown, and I convert them with a cfd.university specific Markdown converter (which, initially, was all my work, but the last few commits were assisted by claude code as it found some really nasty bugs that I would have probably only found after days of debugging).
I can paste markdown into the grammarly online proof reading tool, correct things, and then paste it back. That works for small articles, but for the kind of articles I am writing, grammarly is struggeling. So, the wordaround, again, is to split the article into pieces, copy it into grammarly, edit it, paste it back, and repeat until all sections have been checked.
At some point, I made the error of paying for the grammarly pro subscription. Big mistake. While it already found an uncountable number of errors before, that number only increased with the pro subscription.
These days, I have switched to LanguageTool which fits my needs better. It doesn’t find everythign that grammarly finds, and that is exactly the reason I am using it for. It finds the grave mistakes, and leaves the false positives for grammarly to report on.
While using either grammarly (pro) or LanguageTool (I do have the LanguageTool premium subscription as of the time of writing this declaration, it may change in the future), I have never used their AI features to rewrite or paraphrase any text of my articles.
I’d rather have some odd sentences here and there, which don’t appear polished, than a squicky clean AI polished article. When I first started publishing my articles and realising that there were still so many grammar mistakes in them, even after proof reading them, I always felt a bit embarrased and wanted to come back to fix those. I never did.
These days, I feel differently about them. While I still fix errors in older articles, I do not feel embarrased by the mistakes anymore. If anything, they suggest the text was written by a human, not an LLM.
I have also started to use claude code for proof reading. I have found it to be good in finding some mistakes, but it will miss quite a lot as well. For that reason, I usually run it now first over my article when it is finished, and the manually review it using LanguageTool afterwards to see if there are any errors left (and I usually still find issues).
Again, I do instruct claude to only change spelling issues and simple grammar errors. I don’t want it to paraphrase anything. The prompt below is representative of what I would ask claude to do:
please read all .adoc files inside the chapters folder, as well as the master.adoc file in the root folder and proof read them. only correct spelling and grammar mistakes, do not rewrite any sentences to improve style. do not paraphrase. report any findings you think needs changing at the end (e.g. inconsistencies, wrong facts or statements).
Just for completeness, I mentioned that I write all my articles in markdown, but that has changed now as well. I have switched to asciidoc which is weirdly powerful and beautiful at the same time. If you are a text writing connoisseur, like myself, and have never come across asciidoc (like myself for a long time), it has the simplicity of Markdown but the capabilities of Latex.
In asciidoc, files have the .adoc file ending, hence the promt asking for it.
Claude is very good at identifying inconsistencies or wrong statements, and I usually use those to make changes to the text. Again, these changes are my own, never an LLM-generated correction.
If you ever need more proof that my writing isn’t LLM generate, here is a quote from claude after reviewing one of my articles. It is almost as if it is trying to distancing itself from my writing, just to make sure there is absolute no chance someone reads the corrected text and thinks: Why would claude write it this way?. Here it is:
Tone: "mentally challenged" and "hallucinated the hot mess that C# is" could read as unprofessional to some readers.
Thanks claude, but no thanks, I do not intend to change my unprofessional style.
Upcoming
I do want to be as transparent as possible when writing my articles and, while I do believe anyone with common sense can judge my articles to be human-made, common sense does not seem to be, well, common anymore these days.
So, I am currently working on a feature that independently verifies text to be written by a human. This aspect turns out to be more challenging than I anticipated.
What I am currently doing is tracking all my writing. This tracking includes when I write, for how long, how many sessions were used to write an article, how many words per session were modified (written and deleted, both count), session breaks, and what the write to delete ratio is.
I also compute sentence statistics, e.g. sentence length, standard deviation, Measure of Textual Lexical Diversity (MTLD), Flesch reading ease, and the Gunning Fog index.
LLM generated texts have very specific statistical appearances and they look different compared to human written text.
All of this can be faked, but you would have to essentially create your own LLM and instruct it to behave like a human. Even then, the text would still look AI generated, and, in the end, it would just be easier to write the text yourself.
The technical implementation was the easy part, but verifying it all independently is the difficult bit. I have a solution in mind, but that will still take me a few months to realise. But once it is done, it will update this AI usage declaration to explain how it works so you can make up your own mind wether my writing is AI generated or not.