Why am I still yelling about AI?

AI models have been regularly trained on human works, including copyrighted works, without permission or attribution. This is not only works that were freely available through legal means. AI companies have also used pirated works in training⁵. And because their training processes are not generally open, I'm left to speculate about how much they haven't been caught for. However, the problem doesn't stop there. It's worth noting that even when work is freely, legally available on the internet, you don't get to just do whatever you want with it. Beyond merely being used for training, AI is sometimes able to reproduce portions of copyrighted works, which, depending on the work, could easily violate the copyright or license for that work.

It goes beyond using works without permission, though. At least one generative AI company has been caught scanning, then destroying books, including rare or valuable ones. This is the destruction of human art for profit. This is the destruction of human art for exclusive profit. You may want to wait to see what the courts say⁶, but I know where I stand.

By the way, these companies would also just love if the rules applied to everyone but themselves.

AI companies have also significantly worsened the problem of web-scraping bots on the internet by deploying hordes of aggressive and relentless bots that ignore established conventions. This is a significant way in which these companies gather data for training, and it has a very real effect on internet infrastructure, increasing hosting costs and impacting website performance and avilability. The Wikimedia Foundation has written on how this has affected them, and for-profit content delivery providers like Akamai and Cloudflare, which back huge swaths of the internet, have warned about it. Of course, smaller communities suffer more. As just one example, the CHICKEN Scheme wiki has seen occasional but recurring degraded performance as bots and crawlers relentlessly hammer it.

Training and hosting generative AI models also has a significant environmental and human impact. A report by the UN found that AI's power and water use matches that of millions or billions of people. This can significantly impact climate change as well as the availability of resources for humans. The development of data centers also threatens to displace or disrupt local communities, unfortunately but unsurprisingly disproportionately affecting marginalized communities.

Time and again these companies have proved that they do not care about us.

Outside of ethical qualms specific to AI, I also generally strive to lean away from proprietary software and towards open-source software because I fundamentally believe transparency and community directly lead to more reliable and accessible software and information, and a better future for human beings. This alone makes me somewhat suspect of generative AI models provided by billion dollar companies with an established history of harmful or predatory practices. This also makes their relative popularity in open-source communities a bit surprising to me.

AI companies, like any other big, profit-oriented corporation, are not above serving ads to their users. In my experience, ads that are indistinguishable from content have become quite common. Or, ads are pushed from free tiers into paid tiers, except maybe at the highest level. I don't see why AI would become the exception to those tactics. This is especially dangerous because the nature of interaction with chatbots or coding agents might make it easy to miss sponsored content under the premise of "conversation" or "just code." These companies are also not above any other strategies that will make them money, such as providing services to the U.S. Department of Defense. This is particularly concerning given the technical aspects of the software; a highly non-deterministic and fallible technology potentially being used for waging war and killing is a fairly terrifying prospect.

It directly follows that these companies also aren't above collecting your personal data, which is a real cause for concern. Of course, the bigger, historically well-stablished names in this space like Microsoft and Google have built their empires upon this. It's not about whether they might collect your data, it's that they are already. I have seen absolutely no reason to trust the newer companies more with this. Especially concerning to me, it anecdotally seems like people are a bit too comfortable directly giving personal information to AI models, maybe in part because it can feel like talking to a "person." This including some of the same people who were once apprehensive of uploading personal information to social media. As a reminder: it's never been about personally having something to hide; it's about protecting the right to privacy for us all.

Sometimes an argument is made that LLMs make software or computing more "accessible," by allowing anyone without e.g. the time or money for an education to build software or access information. I have a few problems with this idea. First, the amount of hardware required for the data centers hosting big-name LLMs is literally driving up the cost of consumer electronics required to create or use software. Second, getting consistently high-quality software from AI that rivals that of a skilled developer often has people investing tons of money into it anyway⁷. Finally, the open-source community already does a great job of making software resources, including education, available to all, given you respect the terms of their licenses. This is the primary way that I found my way into that field. Granted, that still requires time, but most worthwhile things do. Outside of software, physical libraries serve a similar role in making information and art accessible, as do volunteer-run projects like Project Gutenberg and Wikipedia.

It's possible that ethically trained, open-source models could address some of these issues, but I don't actually know what that landscape looks like currently, and concerns about the behavior and capabilities of the technology would remain.

It's tempting to think of generative AI as "just" another tool. Even Linus Torvalds, creator and lead developer of the Linux kernel, has compared AI to compilers and editors, implying that "intelligent" people don't shy away from "new and interesting" tools.

LLMs may be "new and interesting," but mere "tools" they are not, and it's obviously disingenuous to compare them to the open-source tools of the past:

• Most open-source tools do not require subscriptions that directly fund the wealthiest men in the world, who then cozy up to authoritarian regimes, or any highest bidder, to protect and grow their already unfathomable wealth.
• Most open-source tools don't purport to replace your learning, design, and implementation process entirely, with the ability to influence your ideology and decision making based entirely on a non-deterministic algorithm finely tuned mainly to impress and increase capital.
• Most open-source tools cannot be used, with minimal effort, to push forward massive misinformation campaigns, autonomously hack computers, or rip off entire artist or developer portfolios behind a dubious veil of plausible deniability in mere minutes.
• Many open source tools are opinionated, developed with a specific philosophy and politics, but one that is in explicit and direct opposition to the shady proprietary practices behind LLMs that center greed, data collection, and anti-competitiveness. The same is true, or at least was, of Linux itself.

If inaction is action, pretending "not to care" about the politics of "tools" like AI, or strongly asserting their neutrality, is no better than explicitly endorsing the people behind them and the practices that back them. How can you separate the tech from the corporations, the billionaries that lead them, and their ethics, when using it requires directly funding those things?

On a technical note, there might be something to say about a "tool" that generates provably and frequently fallible output. The misuse of a tool for harm or evil usually only requires the tool to be "good enough", whereas the use of a tool for good or excellence usually requires a higher standard. Within that framework, AI is terribly imbalanced.

One problem I have with generative AI technology is that it's effectively the ultimate instant gratification machine. The question here isn't whether it's provided ethically, or technically capable, but what impact it is having on human psychology. At the end of the day you have to ask if it's actually good for you and those around you, because our collective health and humanity are absolutely paramount. Otherwise, what are we doing here?

Admittedly, I don't know a lot about psychology. I just wonder if having something that gives us a lot of what we want, whenever we want, pretty much instantly, is doing our brains any good. Several studies suggest that frequent AI use can be harmful to critical thinking skills. This also sort of just makes sense intuitively. We know that low-effort, instant high-reward activities can lead to long-term harm. This is basically how addiction works in general. It's kind of no wonder, then, that a billion-dollar industry has been established around AI, in the same way that there are lucrative industries around sugar, drugs and alcohol, social media… instant gratification is known to work as an accelerator for unregulated (or badly regulated) capitalism, often leveraged regardless of the harm to the consumer.

Instant gratification is probably not always bad. "Everything in moderation," they say, but moderation is pretty far from what we're seeing when it comes to AI, as it gets increasingly pushed down our throats in every job, on every website, in every form of media, and in our hobbies.

AI is quite good at reinforcing our biases and delusions. The fact that it's designed to "feel like" a person, to adapt to and emulate individuals' speech patterns, and agree with you, can make it especially dangerous. It can pull people in by making them feel less lonely, or provide some proxy for socialization that absolutely does not mirror or provide the same benefits as socializing with real people. This is like the echo chambers reinforced by social media, but on steroids. Now, every individual can have their own echo chamber, with effects ranging from blindly helping people feel validated in their problematic opinions or interpersonal arguments, to encouraging murder and suicide.

For what it's worth, I don't think software really gets a pass from this just because it's a "computer" thing; software is also a "human" thing. The effects may or may not be as severe, depending on how essential or critical the software in question is, but AI can absolutely serve to reinforce biases and delusions in software design as well.

One pretty well-accepted platitude about AI is that in order to ensure quality, output must be carefully reviewed for accuracy, correctness, or veracity. I think this is easy to say and accept, but I have doubts as to how easy it is to do in general and how much it is done in practice.

My impression is that, even without AI, carefully reviewing is a more tedious task than creating. Add to that the fact that AI can clearly generate output at a much faster rate than human can keep up with. Then consider that it often generates output that seems good enough or appears to work well enough. All of this makes it seem like the way the technology works almost directly disincentivizes careful review. At the very least it certainly makes it much trickier to implement correctly.

Human art and works, whether created physically or digitally, are wonderfully unique, wacky, and diverse. One of the wonders of browsing an art museum, a library, or the early internet, was to stumble upon wildly different points of view and aesthetics.

AI works are… kinda bland? They're getting better and better at fooling us, but, again anecdotally, people have somehow remained decent at picking up on AI-isms in digital art, writing, website design, code, etc. How many times every day do people post their blog or code repository online only to be correctly suspected of having generated it with an LLM? There's kind of a signature and prevailing same-ness that permeates throughout.

Maybe this won't always be the case, but regardless, I worry that deferring more and more of our design and output to algorithms that spit out slight variations of statistical averages is reducing a level of authenticity, expressivity, and culture from human works.

I could imagine some pushback on this point, claiming that if used carefully, AI does not have to generate samey output. That's cool and all, but the majority of people are not and will not use it that way if they can get away with it, and its increasing use and integration paired with its rate of output seems to skew this in the direction of homogeneity anyway. Besides, at that point, make your own art. Squeeze your own goddamn lemons.