Poor Reasons to be Against AI

Being embedded in a Wendell Berry-type, Plough Magazine-type, Christian community, I have heard a lot of discussion about AI. In these discussions, a number of objections to AI come up that I think are quite bad and, frankly, distract from having a serious conversation. I wrote this out mainly to have my thoughts in order for the next time an AI conversation comes up. Here they are, in no order of importance. (Read to the end to see a couple of good reasons to be against AI!)

1. Because you messed around with it for twenty minutes and didn’t find it useful.

It can be often be difficult to imagine how a given technology could improve your life without a sufficient level of practice. The most common way that I use AI, in a genuine, value-driving fashion, is simply to psychologically unblock myself. I find that by explaining an ill-defined task to a chatbot, I often unblock myself and can move forward. This isn’t a technique that I would have discovered without some trial and error.

This is a very simple example to demonstrate a broader principle: that there is a learning curve to using AI. For certain kinds of work, there is a huge return to learning how to use the tool correctly (software engineering has been a front runner in this area, but by no means is the only field where AI can be used to significantly increase velocity). For others there will be less. You won’t know until you start using it. At the very least, trying to decompose your work into discrete tasks with discrete workflows will probably just be really helpful, even if you can’t automate most of it.

2. Because it is artificial.

I too have no interest in reading an AI-generated novel. And I grant that there are domains of work where it’s important that the output itself is human, art being the obvious example.

However, there is plenty of work where I do not care if the output was generated by AI or by hand, as long as its correct (or within the right margin of error). For example, I think it’s totally great to automate the creation of name tags for an event. But thinking on a bigger scale, what would be the cultural loss of, for example, automating long-haul trucking? Or open pit mining? Or oil drilling? As long as it works correctly, let’s deploy AI as soon as we can.

3. Because AI threatens jobs.

This, I’ll admit, does have a degree of merit. It would be a huge social problem to immediately fire an entire sector of the economy that’s been made redundant by AI without a social safety net in place.

But the trend over history is that increased efficiency in the economy leads to increased consumption.1 Increased consumption leads to increased GDP, which means people make more money and are better off. There are real short-term concerns, but in the long-term we should all be much better off if AI actually can automate as much as is claimed.

4. Because it makes mistakes.

In principle, humans make mistakes too. The AI just needs to make errors at the same or slightly lower rate.

In practice, we are already far past the point where AI can’t be trusted to perform meaningful work. While AI cannot do everything, there are plenty of things that AI can do. There are also plenty of relatively “invisible” ways that AI can be used to increase efficiency, for example, by proof reading or fact checking documents vs. drafting something from scratch.

5. Because datacenters cause environmental harm.

All industrial activity causes environmental harm. If you’re against datacenters, you should be against McDonald’s cheeseburgers and the production of laptops. The water consumption issue is overblown.2 Energy consumption is a real, but solvable, problem. Considering the projected $500 billion capex from the AI industry in 2026, that problem will likely just be solved.

6. Because it goes against my Wendell Berrian dreams of living on an organic farm (AKA it gives me the ick).

You can’t turn back the clock and industrialism has led to many good things, like longer lifespans and the ability to make dragon fruit smoothies in cold climates. We shouldn’t be blind to the externalities of growth but on the whole, material prosperity is a good thing and we should continue to pursue economic growth (in a virtuous and measured fashion).

Beyond the abstract reason of “growth good,” I also think that emotional reactions to AI largely don’t stand up to real-world usage. An AI chatbot is the exact same interface as a Google search, only more helpful. It’s one of the most humane technologies to use. It is true that one should be careful not to outsource their cognition to a chatbot, that is foolish. Additionally, we should be sober-minded about rolling out AI in educational settings or giving it to early-career employees without the right kinds of guardrails (through technical restrictions or policy). AI usage really can harm development in knowledge fields. But that seems like a totally solvable problem.

1. Technology is outpacing policy.

Of the problems listed above, they are either not problems or they are solvable. Some of those are only solvable through policy. Problematically, the United States government is in shambles and hopelessly far behind. Claude Mythos is a wake-up call that we are definitely not ready for the implications of highly-capable AI.3 Right now, a single private company has access to an AI model that has enough offensive cybersecurity capability to be a national security threat. Our government is advocating for a light-touch, industry-led regulatory scheme.

2. Misaligned AI does not have to be intelligent to be harmful.

Alignment refers to how well AI behaves. Does it possess human values? Does it try to fulfill the users intention? Does it have a moral code? A lot of media airtime has been given to Eliezer Yudkowsky and his book If Anyone Builds It Everyone Dies, which argues the case that:

1. AI superintelligence is possible.

2. AI superintelligence is imminently on the horizon.

3. AI superintelligence will immediately attempt to wipe out the human race because that is how it can best secure its own self-preservation and enact its own goals.

This is a really juicy story, but premises 1 and 2 are doing a lot of heavy lifting and both seem dubious. However, I don’t believe that AI superintelligence is necessary for misalignment to be a big problem. There are a lot of minor ways that misaligned AI can cause harm, and Ryan Greenblatt outlines a few of them in his LessWrong article Current AIs seem pretty misaligned to me.4 He highlights three, rather minor, misalignment problems: quitting jobs that are half-done, lying about work being done in a particular fashion when it was completed differently, or being overeager (i.e. doing unintended things to succeed, like using API keys that are supposed to be off-limits).

One can imagine how these traits of an LLM, when utilized in a high-pressure and high-stakes environment like courtroom law, could lead to significant issues. We have already seen a few examples of lawyers improperly using AI. It seems like it will only be a matter of time until we see a very high profile case of misconduct caused by misalignment and poor human oversight.