Gen AI Shouldn't Exist, But It Does

When generative AI started to get really big a couple years ago, one of my main responses was, “Oh great, now we have this shiny new technology to accelerate the immolation of our planet.” It was somewhat comforting that it didn’t actually work: AI chatbots provided wrong answers, produced broken code, and generated images of people with six fingers. The hype would pass, and people would slow their mad rush to hurl more greenhouse gasses into the air for the sake of a novelty. Now, though, the models have improved, and the business case for destroying our environment is strong even if the current AI players are part of a precarious investment bubble that shadows the dot com era. Even if several neoclouds, foundation model providers, and possibly Oracle don’t survive, the technology will.

At least in the dot com bubble, there was much more hope about the positive impact of the technology. I started as an AI skeptic because the negative impacts were clear to me. There’s this thing called the affect heuristic, which is a mental shortcut we often take that causes us to base judgement of a thing’s risks or benefits more on an emotional response than concrete information. The affect heuristic is one of the reasons why you have both ends of the spectrum of opinions on LLMs: on one end, LLMs are both useless and terrible for the society and planet. On the other, LLMs are all-powerful, and thinking of the negative effects are only for people who want to live in the past. Now, I think I can both dislike the technology and acknowledge its usefulness. I’ve seen enough of it in the coding world that I think using LLMs to write code will be here to stay in some form. But I still wish this technology did not exist, or at least that our laws and norms can catch up to the reality that it has foisted upon us. Here’s why.

AI Is an Intellectual Property Laundering Maching

If there’s one thing that people have found LLMs to be very good at, it’s copying: copying an artist’s style, creating a clone of a software product, reproducing journalism content that would normally have to be paid for. It’s not an exact copy, but the effect of copying is the same while the detection and enforcement is harder.

Where the law will land on this is still an open question. Westlaw successfully sued a startup that trained an AI model on their content to create a competing product. The New York Times is suing Open AI and Microsoft for training AI models on their content and providing it to users, but that case has not yet concluded. (By the way, this case is why we learned that a Microsoft executive described this scraping of copyrighted material for AI training as “an astonishing theft of unprecedented proportions.”) Meanwhile, artists, authors, comedians, and a host of other creatives are having to face the reality that their work will be ingested by AI at an industrial scale to produce cheap knock-offs of their own work, with no real recourse.

AI Has Outsize Environmental Impacts

The energy requirements for training and running AI models are enormous. We’re now looking at an AI data center build-out that could negate 15 years worth of progress on greenhouse gas emissions, all while we’re set to cross the 1.5°C threshold that the Paris Agreement determined was the level at which we’d start to see severe climate impacts. I’m hoping that data centers rapidly shift to solar, wind, and geothermal power sources, whether they do it on their own or governments force them as a requirement for permission to build. That’s the only way I see the AI boom not making life on Earth significantly worse.

AI Is Opaque

In so many places on the internet, we are beholden to “the algorithm”: the logic that determines what we see. It could be highly tailored to us, trained on our personal data, designed to keep us engaged through emotional manipulation, or designed to put everything a narcissistic billionaire says in front of us regardless of how interested we are in it. The problem with these algorithms is they are opaque, but at least you land at the source material, so even if you don’t have control over how you get there, you can make judgements about what you see based on what the source is. With AI, even that is gone.

Search engines that provide AI responses can provide links to some original sources, but these may only be loosely connected with what the AI output is actually saying. You have no idea what data the AI was trained on that caused it to tell you what it did. Now, companies are experimenting with serving ads specifically to AI scrapers, with the idea that these ads show up somewhere in the AI model’s training data. It’s less of a problem for coding, (unless maybe someone gets malware into the training data?) but for any answer that involves a point of view, you have no idea what’s influencing that point of view on the back-end. Could governments influence AI models with propaganda? The answer to this question is that it’s already happening, silly.

I don’t see a way around this unless we somehow build a framework of trust for AI models: one that fully discloses what data the AI was trained on, or some ability for the AI to disclose all sources that were involved in generating an answer to a prompt. (I’m fairly certain that the latter is not going to happen, given the nature of how LLMs work.)

AI Displaces Workers by Stealing From Them

The advance of technology is inevitable. During the industrial revolution, skilled artisans were slowly displaced by machines that could produce the same output at a lower quality, but faster. AI feels similar in some respects, but it’s combined with the insidious characteristic that in order to be useful, the AI must be trained on the output of the worker it aims to displace. It’s a combination of the industrial revolution and the outsourcing craze in which companies aggressively outsourced work to lower-paid regions, with some companies asking laid-off workers to train their replacements before being shown the door. It’s a paradigm that makes any knowledge work feel like a Faustian bargain where you can get a job at a company, but that job is to continually make your own employment more precarious.

Ironically, the appropriation chain continues beyond the worker/employer relationship. Not only are workers at risk of their output being used to train AI intended to replace them, but the companies themselves are at risk of their intellectual property leaking into the AI models they use and being made available to competitors. In turn, the foundation model providers are at risk of their own models being distilled and cloned by other AI organizations. I just hope that of these three groups, workers get the strongest protections, but I admit that’s a fairly thin hope.

Is Generative AI Useful?

So, is generative AI useful? To me, the answer is yes, but while you can’t let the negative impacts let you jump to the conclusion that generative AI is useless, the negative impacts also cannot be ignored. I think the easiest route would have been for it to not exist at all, but since it does, we have a lot of work to do to make sure that the negaive impacts are addressed. A lot of that will likely depend on having the right leaders in place who both understand the technology and are minded to address the problems instead of bow to the AI industry.