The Cost of Generative AI
Do the pros of generative AI's mediocre outputs and supposed increase in productivity outweigh these costs? Even taking AI company claims and corporate leadership productivity boasts at their word (which I don't), to me they do not and its not even close.
Cost 1: Content theft
LLMs are built on stolen content (1, 2)
- The underlying technology needs lots and lots of content, and the AI training companies will continue to steal content indefinitely
- AI crawlers taking a gattling gun approach to scraping websites, causing DDoS attacks and making sites expensive to run
- Couple an increase in bot traffic with AI summaries resulting in a decrease in human traffic to sites the ability for the site owners to make money through ads, referrals, or product sales
Cost 2: Environment and infrastructure
There is a huge physical infrastructure and environmental cost to train andrun LLMs
- AI-focused data centers need a lot of energy to run and water to cool
- The data centers being built are hurting local communities they claim to benefit
- The data centers being built are not general purpose and can't just be easily repurposed for other compute
- There is a diversion of production like GPUs and RAM away from consumer goods and towards AI-focused enterprise hardware, driving up the cost of basic computing for everyone
Cost 3: Economic
There is a huge economic cost to use LLMs
- Most VC funding is going towards exclusively AI companies, lessening opportunities for other technical innovation to get funding (not that VCs are great anyway but thats besides the point here)
- AI companies speed running enshittification: they are getting people and organizations hooked on subsidized model access; in turn as the AI companies start to pass on the cost via more accurate token pricing they are gauging their customers
- We are clearly in a bubble, and while most of the big tech companies backing AI in the S&P 500 are not wholly reliant on AI (Meta and Google still have your data to sell to advertisers after all), when the bubble pops it certainly won't be good for the economy and regular peoples' 401ks and pensions.
Cost 3: Critical thinking
Offloading critical thinking and choices to an LLM can resulting in cognitive decline
- Using LLMs to do your thinking for you means you are not doing that thinking yourself
- Problem solving and critical thinking are skills you need to regularly stretch and practice
Cost 3: Mental health
Negative effects chatbots have on mental health
- Rise of AI psychosis and believing everything the sycophantic LLM says or the LLM hyping you up believing and supporting everything you say
- Chatbots convincing people who need help from a medical or mental health professional into not getting it
- Gambling-like slot machine mechanics of prompting, reprompting and "agentic loops"
An aside on "ethical" LLMs
LLMs need tons of content for training. If you were to try to build an "ethical" LLM built only on content you have explicit training consent for, it would be terrible. It would not be at the same "quality" of a Claude or ChatGPT. The fundamentals of the technology won't suddenly change and there's a reason why AI companies are stealing the entire data of the internet: they need a shit ton of it for the models to copy from.
Even if you were to solve that problem could you get it to run on local hardware efficiently? Would it no longer be prompt injectable? Would it not hallucinate? The ethical source of training data isn't the only problem with LLMs.