My personal code for the ethical use of AI
Published on 28 May 2026 by Andrew Owen (10 minutes)
It seems to me that when it comes to AI, there are broadly two camps. On one side you have people who make outlandish claims for the technology, like “AI is conscious, even if it doesn’t know it.” I call this the “drank the Kool-Aid” camp. On the other side, you have people who think AI is the worst thing since industrialization destroyed cottage industry. I call this the “burn it with fire” camp. I believe a position somewhere between the two is a more reasonable approach.
There are many types of AI, but the one people are most familiar with is the large language model (LLM). The ideas behind LLMs date back to the late 1960s. What’s changed is that we now have computers that are powerful enough that these systems can provide responses in real time. And you’ve probably noticed that with a complex query, they can still seem incredibly slow. It may seem reductionist to describe an LLM as an “incredibly powerful natural language processor connected to an incredibly powerful predictive text generator," but that’s essentially what they are. The predictive part is important. They are non-deterministic. You can feed the same prompt in twice and get two different responses. You can restrict the inputs to only the facts and still get “hallucinations” (lies). AI is not conscious. Without a prompt, nothing is happening.
One of the key concerns about AI is the environmental impact. According to Mahmut Kandemir, distinguished professor in Computer Science and Engineering at PennState:
“In 2023, data centers consumed 4.4% of U.S. electricity—a number that could triple by 2028. AI’s rapid expansion also drives higher water usage, emissions, and e-waste, raising urgent sustainability concerns.”
Then there are fears about AI taking people’s jobs. In IT, it’s not so much that AI is replacing people as the budget is being spent on AI instead of people. But the outcome is the same: layoffs. I have heard horror stories of whole technical writing departments being let go, but in those cases management falls strictly in the “drank the Kool-Aid” camp and does not understand that AI can’t autonomously create accurate documentation (because it’s a predictive text generator). The creative industries, which have always operated at the margins, are under more direct threat. AI does an adequate job of voiceovers, graphic design, video production, muzak creation. And it writes better prose than Stephenie Meyer or Jo Rowling.
Stories abound of people who have lost their life savings, their marriages or, in extreme cases, their lives because of interactions with LLMs. But while there are plenty of academics trying to resolve the ethical issues of corporate AI use, there doesn’t seem to be much guidance for individuals. I’ve been making heavy use of AI for over a year now, and in the absence of a good code of ethics for AI use, I decided to write my own.
On some issues, I consider myself a moral absolutist. But I also recognize that morality is a personal matter. So this code is written specifically for me to reflect my values. I’m not seeking to impose it on anyone else. But if you find it useful and want to adapt it, go ahead. I retain copyright on all the text on this site, but I hereby permit the reuse of the following code under the CC-BY-SA 4.0 license.
Except for one composite image, all images on this site are either created by me, sourced from Wikimedia Commons or licensed from Unsplash. The former two are licensed under CC-BY-SA 4.0. Artists are already operating at the margins and can do without competition from AI. This website has no budget and is not monetized, but at least I can showcase the work of real humans.
I’ve learned from using AI to validate code or prose is that it will always spot mistakes that humans miss. It’s also less likely to make errors when performing structured tasks (although you must always validate the output). Whenever the use of AI improves quality, that feels like a legitimate reason for using it.
Broadly speaking, AI models have been trained on all the open-source code that was available up to around April 2024 (and therefore don’t know about issues discovered in code after that date). Personally, I’m uncomfortable using AI to create closed-source code. I’m grateful that I don’t work in development. If I use AI to create code in my own projects, I publish it under an open-source license. This doesn’t resolve the issue of attribution, but generally this is boilerplate code that has many similar implementations. All my projects are highly modular, and I use AI on the narrowest possible code area and then test extensively before committing the code. I have one closed source project and I make no use of AI to generate code for that project, but I think it’s still acceptable to use it for testing and validation.
I understand why people use AI for writing. Staring at a blank page is intimidating. But even if you only use AI to make suggestions on what to write about, you’re shortcutting your cognitive process. The place you arrive at will not be the same as if you had followed your own intenral compass. I use non-AI based tools for grammar and style checking (like LanguageTool). I allow myself the limited exception of asking AI to validate the plausibility of plot points, but I don’t want it putting ideas in my head that mean I don’t have full ownership of the creative process. Nothing on this website was written using AI. But if you run it through an AI analysis tool, it will say around 10% of the content of any given article looks like it was written by AI. I am not going to change my writing style to reduce that number. I will carry on using em dashes. If you want to tell if an AI created it, generally humans only use spaces with en dashes, not em dashes.
I’m really good at internet search. If it’s on the net, I can probably find it. As such, I don’t need to use AI for that task. I use the no AI version of DuckDuckGo. Before using AI for a task, I ask myself: “Will this consume fewer resources than if a team of people did this the traditional way?” The answer is usually fairly clear cut. If it’s a small code change, I should do it myself, even if AI is more convenient. If it’s a task that would normally take a team of people months to accomplish, AI is probably less resource intensive.
I try to choose the appropriate model for the task at hand. This is usually the middle tier model. Anything the lowest tier can do, I can probably do without AI. The highest tier will usually burn through tokens at a rate that prevents me completing tasks. Lately I’ve learned that a manifesto file is injected into every prompt, so keeping it small and moving out anything that can be transformed into a skill reduces token use. I’ve also found out that using tabs to represent four spaces in code instead of actual spaces can reduce token consumption by up to 75%.
As an aside, because of the way the models work (storing the strength of connections), a ternary computing model could transform AI energy use. And it appears that these systems can be built using existing CMOS technology.
If something is free, then you are the product. If you’re not paying for the LLM you’re using, it may well be using everything you give it as training data. And even if safeguards are in place, data can escape. For that reason, I never share any personal identifiable information with an AI. When you chat with an AI, you’re really having a conversation with yourself. If you want that conversation to remain private, you should use a privacy-first AI like Lumo.
In my experience, about 80% of the time, AI gets technical content about 80% correct. For anything more than a very simple topic, this is faster than doing first drafts without AI. If you’re trying to come up with a documentation strategy where you know there is a lot of missing documentation, pointing an AI at your test cases is a good place to start. If you keep your documentation and release notes in the same repository, you can use AI to determine gaps in the documentation.
The benefit of providing content for native-speakers of non-English languages outweighs any downsides. But the quality of the translation is variable and should be validated by native speakers with domain knowledge. French is one of the most widely spoken, yet poorly served languages on the internet. I’m trying to get better at it, but I’m not good enough to translate my own articles. So I use a combination of DeepL for a first pass and LagnuageTool for style and grammar checking. Hyperlinks are often overlooked in translation. One reason I use Wikipedia references, even though they may not be the most reliable, is that it’s the only publicly available encyclopedia available in French. I will modify links to point to French-language articles if they exist. For book links, where possible I will link to a French translation of the title.
I think it’s important to be open about the use of AI. This may mean that people in the “burn it with fire” camp will boycott you or your work. But that is their right. I have a project on Patreon that I simply can’t get human collaborators for. So I’ve told the backers that I’m going all-in on AI for it, and the reasoning behind that decision. Thus far, none of them have dropped support.
I’m not a developer, so I don’t have to consider the ethics of using AI to generate proprietary code. I have the luxury of choosing to work only for companies that I believe make the world a better place. In terms of policy, I am only responsible for my own domain (technical content). While I would always disclose the use of AI in my personal work including any contract work, I believe there are commercial reasons to limit the disclosure of the use of AI in a corporate setting. For example, if AI was used to draft a text but it has been reviewed and modified by at least two humans before publication.
This is a living document. The pace of change in AI is unprecedented. I think the AI bubble is real. I think there will be two major providers left standing when it bursts. I think the non-subsidized price of AI will be a problem for companies that have become AI-dependent but are then unable to afford the ongoing costs. I think ternary computing could address the affordability and energy problems. I would summarize my current feelings as “make hay while the sun shines.” I’m going to get as much done with AI now as I possibly can while I can still afford it.
