My AI Policy for this Blog
How I use and do not use AI/LLMs, and how I make this transparent.
tl;dr
- I write the articles myself.
- I use AI for proofreading and for testing ideas on article structure and storyline.
- I do not use AI generated images in this blog.
- I may use AI for research, but will check the results.
- The technical stack for this blog was built with heavy AI assistance.
- I commit to labeling each article with more detailed information about how AI was used for this specific article.
AI/LLMs1 and their use have become a hot-button issue for many people. No matter your personal opinion on whether AI will usher in an age of abundance, eradicate humanity, or simply drown us in terrible slop, I believe transparency about how AI was used in creating content for human consumption is important, to allow people to determine whose words they are reading - those of a human, or those of Claude. This article serves as my personal rulebook for this blog.
The Central Tension of AI Use🔗
AI can be extremely helpful in writing long-form text - it can provide feedback and proofreading, can illuminate where you are making logical jumps or not explaining a step well enough, and it can be helpful in brainstorming different storylines or article structures. Especially if you are writing for a personal blog, on your own time, and do not have someone else you can bounce ideas off of, I have found this to be quite helpful for quickly improving my drafts.
At the same time, AI writing has a certain… feel to it. I have found that I started to notice this more and more, and I generally don’t enjoy it, especially if it is not clearly labelled as AI writing. If I read someones personal blog, I want to read what they have to say, not what their Claude subscription thinks.
With that in mind, the following are my personal rules for how I use AI for this blog.
Rule 1: The Words are Mine, Always🔗
Every single word you read in these blog posts is written by me, with the exception of quotes (obviously), and the occasional code fragment. While I will use AI tools to proofread and brainstorm, I will never let Claude rewrite the article by itself, and I use the suggested changes as inspiration for my own twist on them. While I cannot rule out that sometimes these edits end up verbatim in my posts, this will be on the order of sentence fragments, not paragraphs.
Rule 2: The Research may be Shared🔗
For articles that require either technical research or collecting facts, I may use AI tools as part of the research process. There is no hard rule on this, except that when AI contributes something, I will not use it blindly, but will understand the sources and methods it used, challenge the results where necessary, and will generally ensure that I have a good feeling that the important things are correct.
Rule 3: No AI-Generated Images🔗
This rule is a bit more subtle than you might think. I will not use images from AI image tools like Midjourney, ChatGPT, etc., for this blog - I usually find them fairly cringe and frankly, having one of them as a cover image makes me not want to click on a link. Of course, I may use Mermaid.JS diagrams and other computer-generated images, and I may have the code for them be generated by an AI model. Also, the same rule does not necessarily hold for all talks linked here - the Keycloak security talk slides contain AI-generated images.
Rule 4: The Stack is Fair Game🔗
Bluntly speaking: No one is paying me for writing this blog, and thus, any time I spend on it is time I could be spending on different things I enjoy. Thus, I try to use my time for the parts of the process I enjoy, and outsource things I do not enjoy to AI. Luckily, writing original text is something I enjoy - but wrangling a static site generator, messing with CSS, or understanding a templating language is not. This blog is based on Zola with the Arabesque template. The deployment pipeline and pretty much all customization of the template was done with AI, including some small gimmicks like the dynamic images for social media previews, and I will keep using Claude Code to follow new Zola releases, fix breaking changes, and so on.
Rule 5: AI Use is Disclosed🔗
This rule is why this article exists. I want to make transparent how I use AI, both in general, and on a per-article basis. For this purpose, I find the AI Influence Level (AIL) framework by Daniel Miessler quite helpful, and have adapted it into this blog. Briefly, the framework defines 7 different levels of AI involvement in a text, from AIL 0 (no AI involvement at all) to AIL 6 (an AI decided that the article should exist, and no human was involved). Each of my blog posts will have an AIL label, and a note at the bottom that gives additional details (see the bottom of this post for an example).
I expect that my articles will mostly be between AIL 1 (“Human-created, Minor AI Assistance”) and AIL 2 (“Human-created, Major AI Augmentation”). In my interpretation, the scale will cover both the text of the article itself, and also the process of getting to the article (any research or analysis performed with or by AI). Please read the page by Daniel Miessler for more details on the scale.
Note: The way I understand it, Daniel Miessler intended this framework to also be used by other people. However, the page does not list a specific license for it, which is why I avoid reproducing too much of it here, and use my own version of the badges. I have contacted Daniel about this question, and have not yet heard back. Daniel, if you are reading this and want me to remove it, please reach out - my contact information can be found in the imprint.
I use the two words interchangably in this article.