Ethics of Generative AI

The other month I wrote more generally about how I’m navigating AI hype and an uncertain future.

I fell into the trap that a lot of these “my take on LLMs” posts do and I didn’t touch on ethics there. In retrospect, I should have written this one first. I will go back and link it at the top of the previous one.

Let me preface this by saying that I think you can make a strong argument that any use of generative AI is inherently unethical. I actually agree with this take, even though I do use LLMs in some very specific and intentional ways.

Ethical behavior in the real world is always a compromise. I wish I could claim that my behavior is perfectly in line with my ethics, but clearly that’s not the case.

I drive a car. I eat animal products. I purchase products that are wrapped in plastic.

All of these are ethical failures. I do my best to be mindful and minimize the negative impact these actions have, but I still do them.

I very much respect people who don’t do these things, and extend that respect to those who choose not to interact with generative AI at all. I would never try to convince anyone otherwise.

The Ethical Pitfalls

Theft of other people’s creative labor.

LLMs and other generative models can be thought of as lossy compression of information and its relationships to other bits of information. This is an oversimplification, but I think it’s a useful one.

When you query a model, it searches that compressed information for related bits and re-assembles it into plausible-seeming outputs of text, images, or sound.

Much of that information came from billions of hours of careful thought and intention by creative people who were sharing their talents with us. This exchange was part of a social contract that we would support them by buying their art, listening to their music, using their software, and reading their musings.

In some cases that work seems to have been outright stolen, in other cases licenses were not respected and mechanisms like robots.txt files were completely ignored.

This will all play out in court, but even if compressing and then regurgitating data you don’t own ends up being legal, it’s certainly not ethical.

This is a line I draw. I do not use LLMs to generate writing, images, or (I can’t believe people actually do this) music. I also do not use it to wholesale generate code for features or functionality or do architecture or planning for how I might structure my application. In the same way that using it to generate an illustration is stealing from the work of an artist I don’t know, asking it to architect a feature is stealing from some engineer I don’t know with zero attribution. This also carries the very serious risk that code that is not compatible with your license is laundered into your project.

What I do use them for is to help me understand code that is not mine, explain a concept in a language or framework I’m not familiar with, apply a pattern that I’ve already designed to another part of the code, or help track down a bug I’m stuck on that is inevitably something simple I have become blind to (the curse of the solo technical founder).

For example, I might use them to generate the Rust structs that map to a database schema I’ve designed where I already have half a dozen examples of similar structs to copy. This is rote work that I wouldn’t learn anything from and will not result in stolen IP in my code. I would not ask it to design the schema, or the initial shape of the structs that represent my data models. This saves time, doesn’t erode my understanding of the code, and doesn’t steal others’ work.

I also use them on posts like this to help fix grammatical and spelling issues. When I do this, I review the suggestions and make the changes by hand if they make sense to me.

For me, these uses raise minimal ethical concerns and I don’t think they are contributing to any significant atrophy of my skills.

Sending money to companies that are making the world worse.

I honestly have a hard time coming up with a single way generative AI has made the world better.

• It has flooded the internet with low-quality content, disgusting imagery, and insecure software at an unimaginable rate.
• It has sucked up the funding for startups so that every new company is a boring Claude API wrapper.
• It’s made it so that computer and electronics hobbyists are unable to afford GPUs, memory, and storage devices.
• It’s causing tremendous anxiety for young people who have no idea how they will get a foot into the economy after school.
• Companies are using it as an excuse to lay off thousands of employees.
• As Yuval Noah Harari predicted in Nexus, it is being used to make and justify government decisions about life and death.

In exchange, it has saved us an hour a day of reading Stack Overflow answers and search results.

Every penny that is sent to OpenAI, Anthropic, and the others is fueling this.

I’ve found that, for my uses, the open-weight models like GLM-5 and MiniMax-M2 are generally good enough. I use these via OpenCode (in plan mode) exclusively and do not send any money to OpenAI or Anthropic (although I did in my earlier experiments). If you use the paid versions of these models via OpenCode Zen/Go, they claim to have a zero-retention and training policy. This is not an endorsement, it’s just what I use.

Flooding the public discourse with low-quality writing, images, recordings, and software.

I don’t think I need to expand on this, since everyone who uses the internet has experienced it firsthand.

The worst thing about this is that the onslaught of slop has made discovering quality work all the more difficult. It’s much harder to get noticed because of all the noise.

As you can tell by my distinctly flawed (though fully human) writing ability, I do not participate in this.

The cure for this is humans. Talk to your friends to learn about new things, and when you are lucky enough to discover something genuine and high quality, share it with people who might care.

Creating a culture of dependence.

The goal of these AI companies is pretty clear - become essential to individuals and businesses so that when the subsidies end and the enshittification begins, we are too dependent on the product to do anything about it.

There are two ways they can accomplish this:

1. Sell businesses on the idea that they reduce their head count. This obviously implies that the remaining employees are expected to do multiple times the work they were doing before “augmented by AI”, but without additional compensation.

2. Sell individuals on the idea that if they do not leverage these tools they’re going to be “left behind”. I’ve seen with my own eyes how this can lead to skill atrophy and complete dependence. Ezra Klein used a phrase recently in I Saw Something New in San Francisco that I thought captured this idea quite well - “cognitive surrender”.

There’s not a lot I can do about #1, except to support software and businesses that are not shoving it unnecessarily into their products (like Vivaldi).

For #2, I’m definitely not worried about being “left behind”, but I am tremendously worried about skill atrophy. As I said, I try to avoid that by being very intentional about how I use it.

Power and data centers.

If we actually build all of these AI data centers, by some accounts we’re looking at ~106GW of additional power needs in North America alone. This is an incredible amount of power we’ll have to generate for something that has few if any benefits to society.

The power needs of these facilities are going to make electricity more expensive for average households, and increase emissions of greenhouse gases as new generation facilities come online.

Still, as technology enthusiasts, I think we need to consider this critique more broadly. How many computers do you have in your home running 24/7? I have at least 3. They do things that would be hard to do without, like serve as a router and firewall for my home network but, also things that are pretty frivolous. Is that an ethical failing? Probably yes, but I’m unlikely to give up those conveniences and hobbies.

Likewise, I think running prompts 24/7 to generate cat memes to flood social media forums is a lot different than running an agent to review your open source project for security vulnerabilities. So while I generally agree with this issue, I think there is more nuance here to explore.

I will always argue against unnecessary data center growth pushed by investor speculation, but I don’t think that the average person’s usage of these tools is the driving force here.

In summary

• I never use LLMs for creative work. I write, code, and make music with my brain and hands. For the things I’m bad at, I support people who are good at it. If I need an illustration, I will hire an illustrator. If I need copy, I will hire a copywriter.
• I now exclusively use open-weight models. One day, these will be more affordable to self-host, and using them helps undermine the proprietary monopoly.
• I don’t make slop. I use my brain. I care deeply about what I put out in the world.
• If I’m learning something new, I never rely on LLMs. Powering through the friction is how you come out the other end an expert.
• I always frame my usage so that I would not miss these tools if they go away. I do not need them to do my work.

You might not have a choice

If your employer is pushing (or requiring) you to use Claude Code, you may not have a choice here. You have to feed yourself and your family. If you have to use these things all day at work, I’d encourage you to exercise the skills that are important to you on your own time.

If you enjoy coding, work on a side project without LLMs. If not, maybe start a personal blog and write about things that are important to you or that you enjoy.

Good luck out there friends.