Generative AI is the use of artificial intelligence to take some inputs (e.g. a prompt, a matrix of noise, etc.) and generate some form of media (text, image, video, or audio).
Overall, it's something I think the utility of has been overstated, it's an industry that’s been over invested in, and it's a danger to our environment worse even than crypto. But, after the bubble pops, I think it might become a useful tool in limited contexts.
Also, full disclosure that despite my general concerns with the industry, I do have professional experience building tools that utilize generative AI (although with the recent pushes to shove it in every product, which software developer can’t say the same?). I more fully disclose my personal use of generative AI at the bottom of this page.
This is an ongoing trend in computer science that has been fairly controversial, to say the least. I'm a programmer working at a company that uses GenAI, which certainly brings about some mixed feelings.
For what it's worth, I'm not going to be going into detail on how GenAI works. There's plenty of resources out there, and I'm not interested in explaining how machine learning, transformers, diffusion, etc. all work. This page is going to more go over the ethics and impact of it all.
Common Examples:
- LLMs like ChatGPT
- Some also take voice and video input, like Gemini or ChatGPT-4o
- Art generators like Dall-E or Midjourney
Training Data
I have some complicated feelings towards copyright. If it weren't for capitalism, I'd be totally down with the default being that information ought to be free, but I think an exception for requesting content not be used for AI training. Philosophy Tube argues in What Ethical AI Really Means that “ethical AI cannot exist under capitalism” which I think is fair. Relatedly, Alexander Avila makes a reasonable leftist case for why even the AI hate may be being channeled towards efforts that still just serve capitalist interests.
For right now, though, the policy should absolutely be to expect explicit consent for use in training, or wait for copyright to expire. It's quite frustrating that companies double dip and withhold their creations from the public domain yet completely ignore copyright when there's profit to be made.
Their indiscriminate scraping has also put a lot of strain on a lot of websites, including Incremental Social.
For what it's worth, this data on AI sentiments seems to mirror my thoughts, where China is a stark outlier in how much they think AI is helpful, while also being quite socialist as well as not caring much for copyright.
Environmental Cost
Obviously, machine learning is a relatively expensive process. There are a LOT of operations to be performed, and despite the optimizations from chip manufacturers, models continue to get larger and larger and require ever more resources to train and use. But there's levels to it, and it doesn't seem like (all) forms of generative AI are necessarily prohibitively bad for the environment.
LLMs are, compared to the other forms of generative AI, not too bad and offer the most clear pathways to optimize further.
Training varies upon a lot of factors but is in the realm of dozens to hundreds of annual household emissions each, and uses as much water as several households in a year. And these numbers are expected to grow substantially as models proliferate and grow in size: https://generative-ai-newsroom.com/the-often-overlooked-water-footprint-of-ai-models-46991e3094b6. This is concerning - I hope the hype dies down, the bubble pops, and this growth doesn't end up coming to fruition. It'd also help if people collaborated and shared more rather than reinventing the same models over and over, but capitalism/copyright are unlikely to allow that.
The actual usage of an LLM is much smaller, in the realm of performing a couple Google searches. With quantization, smaller specialized models, etc. the actual usage of the models is fairly negligible.
Other forms of AI generation are much worse. They retain the high training cost, but are also so expensive to run that in some cases it'll even exceed the training costs, both in terms of energy and water.
And I should clarify: we have already been on a very bad path with respect to the environment, even before AI. We have a LOT to do to reduce consumption of resources and switching to renewable sources of energy.
Quality
A big question surrounding generative AI is whether it's even worth the above issues. These images are often not very good to look at, videos lack temporal consistency, and LLMs hallucinate regularly.
Previously I was a believer in the argument that this would rapidly change and the quality of outputs would improve so rapidly that quality arguments would be disproven and AI generated content wouldn't be “clockable” as such. And sure, looking at videos they look a lot better than a couple years ago... But they also feel like they've started stagnating. And they're certainly still clockable, with regular mistakes that are easy to pick up from a discerning viewer. Images and text still have certain styles (yellow filter and outlined text for images, emdashes and negative parallelisms for text) that make it highly recognizable as well. To be honest, I'm having low confidence that models will be able to be made that fix these problems, or if they do that it won't just have other styles we'll learn to pick up on.
In any case, I do think AI is being used for situations where it's simply not the best tool for the job. I don't enjoy reading generated text and don't want generative AI movies, books, music, or other forms of art.
I think AI can be used in specific use cases well, though. In particular I think it's fine when answering queries on text that's included in the models context: this includes tasks like asking for feedback on human made text (like an article), summarizing notes you took in class, or even asking questions based on those notes.
LLMs are once again basically the only form of generative AI I'm even a little understanding towards, though. There are simply far fewer (if any) use cases for generated art etc. as a tool rather than just creating slop, and as mentioned they also use much more energy and water.
Psychosis
A very concering aspect to LLMs is their ability and propensity to feed into people’s delusions. “AI psychosis” and general sycophancy have me very concerned about the use of AI to replace conversations with real people, particularly therapists. It’s also soured my attitude towards fictophilia. I hope society can shed its individualism and mental health crisis exacerbating these issues, and address the sycophancy on the model level, perhaps using some of the solutions Maggie Appleton discusses.
Dead Internet Theory
AI slop is ruining the internet, filling it with low quality content, misinformation, etc. for ulterior motives. I want a internet with human made content with artistic intent. We're going to need to employ tools like digital locality to avoid bots and AI generated content.
Accessibility
One big argument I've seen for AI is that they can help people who would otherwise struggle with certain tasks. Someone bad at writing can use AI to clean it up, or someone who's bad at art can make a game that requires it.
I do not want to be ableist and generally consider most affordable made for accessibility quite reasonable. It as an argument for AI doesn't quite make sense to me - I think a spellchecker will typically suffice, and there's plenty of royalty free art and music out there for your personal projects that don't require burning a forest to make. But as the technology improves perhaps I can be persuaded it's okay to use for cleaning up writing and stuff. Until then, I'd suggest just asking for feedback but making the changes manually.
Honestly, I do think natural language interfaces will become increasingly common for a variety of tools, where appropriate. They're similar to a command palette in that they make it easier to access a plethora of actions.
Shaming
I still use LLMs, and not even exclusively at work where I’m essentially compelled to use it. I think it’s reasonable to talk about the issues with LLMs without shaming its users, many of whom have needs being unmet through other means. Shame, similar to abstinence-only education, is not an effective tool of preventing generative AI usage. I like this video (which even includes an interview with Devon Price!) going over this stance in detail. Of particular interest is her point that those who know more about generative AI tend to be more discerning with its usage, including what tools they use and which companies/models they support.
Opportunities
In an article I find veering a bit towards techno optimism, Douglas Rushkoff argues that while in its current form the powers that be are trying to utilize generative tech in ways that preserve their position in the status quo, AI (and the destabilizing “wobble” it creates) offers opportunities to undo a lot of the bad assumptions that have been instilled in society since the industrial age or earlier.
I disagree with their optimism for generative AI, but agree with the things they believe can be improved past their colonialism-influenced current day status. So I wanted to discuss each of them here and a bit of why I think the solution to fixing these problems lies outside of generative AI.
Jobs
Jobs are social constructs, ones society has not had forever and will not have forevermore. Automation will certainly play a part in that (as would removing a lot of “bullshit jobs” that only exist due to capitalism's insistence on everyone having a job), but generative AI specifically will likely not.
Without the need for social or financial capital, using generative AI to create content will no longer make sense. The results will be error prone enough to not be used for anything high stakes. While I think it can help in some specific use cases, like barefoot developers solving the long tail of user wants, I just don't see many jobs that will be considered still necessary to be done where generative AI will be the right tool.
Education
Education is something I've discussed here at length. I'm obviously highly critical of its current state for a myriad of reasons. However, his vision for how generative AI will affect education is it becomes the teacher. I don't think knowledge and education works that way. Knowledge is not static, and the way it should be exchanged and built up upon needs human minds coming together and forming novel ideas from their unique perspectives, not the regurgitation of previous knowledge.
Language
I think this is the “opportunity” I disagree with the most. Rushkoff talked about an interview they did with a researcher putting indigenous intelligences into LLMs, and an LLM generated a response asking to “model its responses in something other than language”. I will not argue against the idea that language, particularly eurocentric languages, are wholly incapable of modeling certain kinds of intelligences. But a large language model will be working in some form of language (although admittedly with a broader definition of “language” than most might be assuming).
But this is not a revelation brought upon by AI. We did, in fact, already know that language is a social construct, and many humans have described the limitations of language, and how much more limited some languages are than others. AI is not necessary for us to realize this, nor is it the best way for us to fix it - as, due to the nature of its training, it is a manifestation of all the issues with language.
My own use
I'm going to end with some transparency about my own use of generative AI. I'm not here for purity testing and I am not “pure” myself. We live in a society, after all, and criticizing it does not preclude us from continuing to live within it. I’m not making excuses for my own use, nor am I trying to draw a line suggesting anything more/worse than what I’ve done is where it becomes unacceptable. All of the above is levied at systems, not the individuals using the technology. The following is roughly chronological.
Personally, I have done the following:
- Built software that uses generative AI to enhance or otherwise modify images and video at my job.
- Used ChatGPT for answering questions, reviewing my writing, getting arguments for or against various ideas, and even to essentially tell me I'm right in a handful of arguments.
- Subscribed to ChatGPT plus for 6 months back in 2023.
- Setup OpenWebUI and ollama for local processing (although I've never been able to run models with high enough quality to really be useful for anything, so nothing from this ever actually got used).
- Used Claude code to write software at work
- Used Opencode off and on to write some software for personal projects since April 2026, including subscribing to their Go plan.
To elaborate on that last point, I have a more thorough documentation on my page on vibe coding.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.