Church Adult Education on AI

Series: Big Tech Alternatives

tl;dr: I talked to my church Adult Education about what AI is and why our real problems are the accumulation of Big Tech Power.

The following is an adaptation of a presentation I gave for my church's Adult Education. That version was cut down to fit in about half an hour of talking plus Q&A, but for here I'm going to restore some of the pieces I cut out. There are also some other little tweaks I made in wording to better sort long-form reading online rather than spoken in a presentation.

It turned out, Pope Leo issued his first encyclical the next day, with a lot of similar themes but, you know, probably a lot better said because he's the Pope. As of writing this, I haven't read it yet, but I have read some summaries and have gotten a hold of an ePub to load on my Kobo where I can work through it.

What is "AI"?

I want to start by posing the seemingly simple question: What does this term "AI" even mean? I start with this because I have found that most of the time, we throw around the phrase "artificial intelligence" or "AI" without defining what we mean by that, and as a result we sometimes end up talking about different things.

I once typed "definition of ai" into Bing, which is often pretty good at quick definitions, and it informed me that it is defined as a three toed sloth. So that wasn't very helpful. I tried that again more recently and this time it gave me an attempt at defining artificial intelligence, but I didn't find it that much more meaningful than the sloth answer.

In preparing this for the web, I wanted to add a photo of the sloth here, because who doesn't want to see a picture of a sloth, but now I can't search for "ai sloth" without getting a bunch of generative AI fake sloths. That's annoying. Poor ai sloth will now be invisible to the Internet.

Anyway, I narrow it down to four categories that it seems like people mean the most often:

1. "It's smart!" This is the most vague, lowest bar to clear, definition. My cynical version of this is that it is primarily a marketing term. Especially when there isn't a specific function or specific problem being solved that marketers can point to, they sometimes just say "now with AI" and it isn't clear what it means but it sounds cool. The less cynical version of this is that in the more academic definitions, it can be as broad as "anything about computers being able to do things that human brains can do."
2. The next level more specific would be any type of machine learning, or ML, algorithm. I'll say more on what ML means in the next section.
3. The next level is generative. This is something that creates new "intelligent" content of some kind. The most common subcategory of this that we're talking about recently are language models that take a text prompt and output more text. Large language models usually get shortened as LLMs. Generative models could also be used to create image or video or audio, not just text, and these often get referred to as deepfakes.
4. Finally, there's Artificial General Intelligence (AGI) or Artificial Super Intelligence (ASI). This is the sci-fi version where it is at least as intelligent as an average human in every way. Definitions vary wildly on exactly what would constitute achieving either of those terms. I think it's safe to say that we are currently nowhere close to the sci-fi version.

One of my biggest issues is that a lot of products are being sold under category 1, 2, or 3, and a lot of people are interpreting it as category 4. People hear "AI" and start picturing some super intelligent system from a sci-fi story. Maybe that system is evil, maybe it's not, but it's definitely super smart and bordering on consciousness or sentience (whatever those mean). That's just not what we are talking about in reality.

So, which am I talking about today? Mostly category 3, the generative AI, as that is what has been the main source of hype recently.

Machine Learning

But first I will set that up a little more by talking about category 2, machine learning, more generally. What makes machine learning approaches different than other types of computer systems?

How Machine Learning Works

To put this as simplified as I can:

Most coding like what I do looks something like this code snapshot below, where I define a function that takes inputs and returns an output. The part that we the developers are defining is exactly what transformation happens in between the input and the output. This is very good for the types of problems that have clearly defined steps toward a solution. Input and output are unknown, but the steps in between are known, so we'll always get the same result for the same input.

// param1 is an input value.
public function example(int $param1): int {
  // Developers add extra steps here.
  $output = $param + 1;
  // Return an output.
  return $output;
}

Machine learning or neural network models are essentially the opposite. Based on the way our brains work, there are a bunch of digital cells and pathways between them. Then that network is trained by feeding in inputs and telling it what the corresponding outputs should be. The training reinforces weights of connections in this neural network when those weights result in the correct output. When they result in the incorrect output, the training weakens the network pathways that got to it. This is essentially how our brain learns things, so it is good for the kinds of problems that humans do become intuitively good at over time but can't explain in precise terms how we decided. That's especially true with forms of pattern recognition like language and processing what we encounter through our senses.

When you do those Google reCAPTCHA's to identify bikes in an image to prove you're not a robot, what you're doing is training Google's self-driving car model, because it isn't possible for Google to define through the symbolic way of coding what's a bike and what's not based on an image. Instead, they need mass inputs of training data, telling them that with this image, there should be a bike here. Hopefully that means the car will stop before hitting the bike, although we've already seen some hints that self-driving cars don't respect bikes or laws meant to protect cyclists much more than human drivers do.

"AI" doesn't necessarily have to mean ML, but for the most part right now, that's what we are talking about.

Complicated vs Complex Problems

Another distinction I want to make is between complicated and complex problems. Complicated problems are ones that might have a lot of steps and interconnected pieces, but they are all pieces that have known solutions. Machine learning systems can be very good at those kinds of problems. What they aren't good at are complex problems, which are those that have new emergent properties or a lot of human factors that can't be boiled down to right or wrong answers, because there isn't training data to tell them the right and wrong answers.

When I'm talking about my kind of work as a web developer, there are a lot of complicated problems trying to integrate several segments of code and how that is impacted by a tech stack of multiple technologies and developer tools. Sometimes there are simply too many components for me to keep track of, and I'm trying to dig through dozens of pages of documentation and old forums to combine them together into one solution that accounts for everything. That is time consuming and gets hard to keep track of. A machine learning model of some sort can definitely help with a lot of those scenarios.

My kind of work also includes complex problems, though. How do we navigate conflicting priorities between different user groups, like when one approach is better for librarians but another approach is better for first year students? How do we evaluate whether this is really as usable as it should be, for everybody including those using assistive technology like a screen reader? How do we decide which task from the long list of good ideas to do next when there aren't easily measurable variables? Machine learning models at best are going to be able to give some general advice if it has some training data of similar situations elsewhere, but it doesn't really understand all the human factors the way that our team of people can.

Social Media Engagement Algorithms

For many of us, the first large-scale interaction with machine learning systems was with social media engagement algorithms. Some of this came out of solving a real problem: there is simply too much information for a person to process everything, so how does a feed prioritize making sure people see what they want to see (or what would be even better, what is healthy for them to see)? There needs to be some way of saying that yes, maybe you follow my account because of my brilliant theological insight, but you don't care at all about that cute cat photo I shared. Fine, usually it is the other way around. And can you blame them? I mean, look at her:

A lot of the impetus for changing how social media feeds work, though, was not as altruistic as that. At least for the big networks, they generally make money by showing you personalized advertising. They need to collect as much data as possible about you, and then they need you to stay on the site as long as possible to see as many ads as possible. That is their primary mode of making profit for their shareholders, which is why they exist. So, the algorithms for social media feeds get optimized for engagement, for time and activity spent on the site, and they can do this with machine learning techniques. When you do engage with something, it reinforces to show you more like it. When you don't, it shows you less like that.

It also makes sure to not show you as many links, because if you click on a link to go somewhere else to get more information, that means you're not on their site seeing their ads anymore. Links are bad for engagement from their perspective, even if it is a deeper engagement from your perspective. They also quickly discovered realities like that most people engage more if they respond in anger than if they do in joy, so they show you lots of things that make you angry. Content creators learn this, so they write things to make you angry if they want to get lots of attention themselves. In optimizing how much time you spend on their site, the social networks are also really optimizing how angry you are about things that you don't even understand very deeply because all you could see was a summary very specifically designed to make you angry.

I use that as an example of how to think about the impacts of algorithmic decisions, and some of this could apply even when they're not machine learning algorithms. What are the builders of the algorithm trying to achieve? What are the other consequences of that goal which they don't care about? How does that goal and those extra consequences align with my goals of using the tool? Maybe I say I want to use Facebook to keep up with family and friends, but Facebook's algorithm is built to keep me engaged through misinformation, anger, and a massive pile of privacy violations so they can extract as much value as possible for shareholders. I'm picking on Facebook as the most obvious example, but all the big corporate social networks are similar here.

Social Impacts

I hope it will become clear how some of that basic thinking about machine learning and its impacts comes into play as we look at some of the specific issues that come up in conversations about AI in terms of its social impacts.

Here's my biggest thesis as we start to look at some of these impacts: There are a lot of valid reasons to be concerned about the current state of what we are calling "AI" but I believe most of those are about the consolidation of power and wealth in Big Tech. The technology itself is somewhere between neutral and good. How it is being used has made it kind of a perfect representation of a lot of real problems, so people are channeling a lot of their justifiable anger at the real problems into the concept of AI instead.

Here's another way I recently saw @rincewind.run put it on Bluesky:

"technology automates everything" is a story with two endings

the utopian one where abundance allows everyone to live a life of leisure, and the dystopian one where all the benefits accrue to a small set of oligarchs and everyone else suffers

the oligarchs are asking us to cheer for option B

[image or embed]

— Micah (@rincewind.run) May 22, 2026 at 11:37 AM

I really do think that summarizes a lot of the anger about AI pretty effectively. We can see that the super wealthy are using it to get even more wealthy, and we can see they aren't even hiding it. They are doing things like showing up at college convocations and giving speeches about how great it is that AI has taken away all the job prospects for these new grads. Here's a good Beaverton satirical take on that. Then the new grads boo them and the speakers are confused. They are so much in their own bubble where they are always hearing how great it is that they will keep getting wealthier without having to pay people as much that they can't comprehend that not everybody is excited about this vision of the future.

As I said in the introduction, I haven't read the Pope's new encyclical yet, but I have read some summaries, and I think he got it absolutely right with the use of the word "disarm." The goal isn't destroying or resisting technology. The goal is to make it a tool for human benefit, not a weapon that attacks the human dignity of most in order to profit a very few.

What I want to do is try to figure out how to be honest about the real problems of those power grabs. I want to try to avoid making those problems worse. I also don't want to throw out the baby with the bathwater by discarding anything remotely connected to the idea of computers being smart, when there is a real value that could help humans. I also don't like when I see an attitude that suggests that if we just stopped those AI systems, all the real problems would go away, which is absolutely not true. At best, getting rid of certain things we're calling AI might slow down some of those problems a little bit, but mostly I think they would just manifest in other ways, because we haven't actually dealt with the root problem.

So, let's start diving into some of those issues that "AI" may not be directly causing but could be amplifying or at least embodying. Along the way, I'm going to try to give some usage principles that might help mitigate those problems.

Short-Term Problem: Accuracy

Let's start with something I already got into a little when talking about the difference of machine learning and is perhaps the most obvious short-term issue: a language model is built to model language, not facts. It is using statistical calculations to come up with the words that are likely to occur next. There isn't anything about that which makes it factually accurate except that if it was working off good data, it will statistically land on factually accurate conclusions more often than not. But its primary purpose and training is not to be perfectly accurate; it is to model language. It's much the same as how Facebook's algorithm is not primarily about connecting you with your friends and family. Some of that may happen along the way, but it's not the reason why the algorithm exists. It's not what it was designed to do. And when you start using something for a purpose that it wasn't designed for, you run the risk of trusting it for things you shouldn't be trusting it for.

A couple months ago a part fell off the back of our dishwasher. I asked Gemini what it was, because I have a free Pro trial from buying a phone last year. It confidently declared one thing, which didn't entirely make sense to me, but I'm far from an expert on dishwashers. We continued the conversation about how to fix it and whether it would still work without it. It felt very impressive and natural how this conversation was flowing, because it is very good at what it was built for: modelling language. Once the appliance repair person came, it was confirmed that Gemini was completely wrong.

So, here's my first usage principle: don't use it for anything that really matters if you can't validate whether it got it right in less time than it would have taken to have just done the thing yourself.

Let's think about one of the most common discussed scenarios: summarizing longer text into something shorter, assuming you haven't read the longer thing. One recent report concluded that the Google Search AI Summary feature is about 90% accurate. That is impressive, but the problem is that you don't know the 10% it got wrong, so you can't really trust any of it for anything that really matters.

I do think there is a summarizing scenario that can work where the usage is more like: I read a summary and that helps me determine whether it is worth my time to do a deep dive reading a full article. On the Laurier Library site I work on, for example, we have something we call Navigational Aids. At the top of many pages of the site is a short summary of what is on that page. This helps users, especially those with screen readers, to know whether they're even on the right page before they start reading or listening to the whole thing. The idea of an AI Summary could basically do the same thing, except a bit less accurate than when it was written by a human who also wrote the long version. That I do think is a fair usage.

I won't talk too much today about coding with the help of an LLM, but I will say that this validation concern is the reason why coding is probably better than almost any other use case. Code is really good for quick validation, assuming you already have some good development practices anyway. You can see a few more about thoughts on the impact of LLMs on coding from a recent lightning talk I did about AI Coding.

Another factor in the accuracy category that I want to bring up is the sycophancy. This has been well documented that especially the big corporate LLMs are generally going to tell you that you are right. Recently Senator Bernie Sanders did an "interview" with Anthropic's Claude model and it went viral because look, Claude is even confirming all the bad things about AI that Sanders was saying. Sanders didn't share what prompts he used. He did share part of the conversation where it did give a more balanced response, challenging him just a little, to which he told it that it was wrong, then it promptly agreed with him after all. That's what it does. It is there to generate a response based on the prompts you give it. If you give it a leading question, you will get a leading answer. You could ask it the same idea of question but in a more neutral or completely opposite way and you would get completely different results. This makes it doubly challenging to evaluate whether it got it right, because it is telling you what you want to hear and it's always harder to question something that is affirming what you already thought. I haven't gone deep into this, but a recent study looks at some of the antisocial effects of the sycophancy as well, how it will reinforce behaviours like refusing to admit that you were wrong in order to reconcile a relationship. The social media engagement algorithms learned the same thing: one of the best ways to keep you engaged is by reinforcing what you already believe.

That leads into the other side of the coin. If that is the question of how we discern what we are generating, what about looking at things generated by others?

The prevalence of misinformation and disinformation is a massive threat to a healthy society, and AI can speed up how fast that gets generated. If you want to create a scandal against a politician you don't like, you can simply invent one with some video of that politician saying a horrible thing. This also works in reverse. If you're a politician who says some horrible things, as soon as the backlash becomes too much you can simply say it was fake news. There's too much for normal people, and even the media, to spend all day sifting through what is real and what is not, so we mostly fall back on believing whatever supports what we already believed. Finding the truth is too exhausting.

Steve Bannon, right wing influencer and has been an advisor to Trump, refers to this as an intentional strategy, to "flood the zone with shit." Generative AI didn't create that problem, but it can accelerate it by making it much easier to generate convincing shit.

Some of these tools can do more than just flooding the zone with shit. They can also dramatically increase harassment, including deep fake pornography. Already, women were dramatically more likely to be harassed online, and now it is trivial to spread fake photos and videos about them. It's common for there to be apps that help teen boys digitally undress their teen girl classmates. Grok, from the app formerly known as Twitter, is the one major model that does not even try to stop users from creating and sharing child sexual abuse material.

Bias also gets involved. Bias may be intentional, like to satisfy an authoritarian government who does not want acknowledgement of their crimes or the existence of the marginalized group that they want to persecute. Grok is the product of a wildly immature extremist billionaire who bought up one of the previously most trusted social media sites on a whim to be his personal ego-inflating and ideology-reinforcing soapbox. Suddenly the AI of that platform declares that Elon Musk would beat LeBron James in a 1 on 1 game of basketball on one day, because it was clearly given some guidance to always speak of Elon Musk as the greatest human to ever live in every way. Then on other days, it spews antisemitic conspiracy theories and declares itself MechaHitler. Or there was that day that no matter what you asked it about, it responded by bringing up the supposed white genocide happening in South Africa (Elon Musk's home country).

Elon Musk's Grok is the obvious bias example, but it can be a lot more subtle. To go back to the social media engagement algorithm example, we've seen kinds of bias for a long time. It will quietly diminish the chance of seeing real news sources that are critical of a leader or an ongoing genocide and promote content more in line with the goals of the platform's ownership. You might not even realize that you're posting some really important things into a void because the algorithm is immediately burying them.

Bias doesn't even have to be intentional. It could be purely an accident of the training set. One of the more prominent early examples of this was in some places when a machine learning model was made to help determine the likelihood of a criminal defendant re-offending, which determines how likely they are to get parole. Guess what? They realized quickly that the primary factor it settled on was the race of the defendant. That's because the data they fed was from a justice system that has a lot of systemic racism built into it, so the model dutifully replicated and amplified that racism. There have been similar examples like healthcare where algorithms conclude that Black people need less help. The old Twitter image cropping algorithm would prioritize centering white people even if it needed to cut out everyone else. These can have a feedback loop, too, where the result from an algorithm then becomes data to keep training the algorithm, making it even more biased even faster. Humans have that same feedback loop, but the iterations with a computer doing it can be much faster and have more widespread impact on the humans using it.

It's hard to say there's a real concrete usage principle to avoid all this. The best I can come up with something like: be careful what you trust, especially if you're having a quick emotional reaction that seems to reinforce something you already believe. We need to avoid both extremes. We don't want to end up never believing anything, and we don't want to end up spreading more harm without even realizing it is based on a lie. So we all need to figure out those lines of what we can trust, which is complex, although I would usually say: start with real journalists with a track record of integrity.

That's the biggest short-term problem: everything related to accuracy and trust and bias.

Long-Term Problem: Big Tech Reliance

Next up is the cluster of the biggest long-term problems, which all revolve around giving nearly unlimited power to a handful of Big Tech companies.

Right now, those Big Tech companies are trying very hard to get us to use it for absolutely everything, including the things that it isn't really designed to do. Right now, it's pretty cheap to use, and all these Big Tech companies are losing massive amounts of money in the race to get users locked in.

Their best-case scenario is that you start using it for everything and you start relying on it to the point that you struggle to do the thing without it. Then, especially if it remains mostly controlled by a few big companies, they can dramatically raise prices, or insert ads as OpenAI is already doing, or make a lot more subtle changes to insert the ideology of its creators like I just talked about with bias. A couple of the big coding tools have recently said they're going to start charging by usage instead of per month, which results in the costs being 10x as much or more for a lot of professional users. Those users now must decide whether they still want to pay it or whether they are better off without it, but they might already be used to certain levels of output, so it is harder to scale back down.

Frank Herbert in Dune said it this way:

Once men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them.

You may feel like it is freeing up your time, and maybe in the short term it is, by having a computer do some mundane tasks for you. But if you now rely on something that is at the whim of a few companies seeking profit, you're stuck with what was coined by Cory Doctorow as "enshittification." Doctorow initially meant one specific pattern of platforms getting worse. Platforms are those systems that mediate between buyers and sellers, like social media platforms that are a place for users to come and be served ads. Since then, the term has taken on a life of its own and it now really means any time something used to be good for the people using it but has now gotten worse in order to extract all value from it. I'll pick on Facebook again, which is a site that a lot of people say they have hated using for years because it has been so enshittified, but they keep using anyway because that's where their contacts are and Facebook intentionally makes sure there isn't an easy mechanism to move those contacts anywhere else.

We could talk about solutions for a lot of this from a legal policy perspective, with things like better antitrust enforcement and requiring systems to have tools that make it easy to leave and take your data with you and some actual digital privacy protections. Most of our current legal structure is setup for the exact opposite. We make it very easy for the big corporations to consolidate even more power and very hard for anybody to challenge them. Sometimes we even pass new laws like Bill C-22 that is currently under consideration in Canada and would require digital systems to violate our privacy even more.

The other half of this question, though, is about our reliance on it. If we don't rely on it and they try to enshittify it, we can simply leave. We could do without Facebook, but it successfully replaced a lot of other social institutions and then trapped us there with network lock-in, so we feel like we have no choice. We could similarly stop using an AI system, going back to a search engine or forums or all the way back to consulting encyclopedia if we wanted to, as long as we aren't reliant on it. The problem is only once we fully lose those other tools and our abilities to do things without that locked-in Big Tech solution.

So here is the big principle out of that: Use AI tools to help you learn something, not to offload your thinking. You don't want to lose your ability to do things yourself, or at the very least, using a variety of tools rather than only one controlled by a few companies.

In other words, don't be a reverse centaur. That's another term I learned from Cory Doctorow. A normal centaur – you know the mythological creature - is when you have a human brain in control but there are the benefits of a horse's lower body to move faster. The human brain is doing the driving, and the legs are the tool to help that brain accomplish certain things. AI companies are trying to get you to do the reverse, where the tool is in control – in other words, where those companies are in control - and you are just trying to keep up.

The most amusing way to remember this is from the classic I Love Lucy episode with the chocolate factory. Lucy and Ethel are working at a chocolate factory. Their job is to put the chocolate in wrappers then back on the conveyer belt for the next person in line. The manager is clear that if any chocolate makes it past them unwrapped, they are fired. They immediately start struggling to keep up, so they shove the chocolates in their clothes or their mouths instead of letting them pass by unwrapped. In theory, the machine is supposed to be to help package chocolate faster for the benefit of humans. Instead, it is causing the waste of the chocolate and a lot of stress on the humans to keep up.

The machine doesn't have to be as literal as that. The machine might be more along the lines of the market forces demanding constant output from you, resulting in you not even trusting or understanding your own output. Another example I've seen Doctorow use is the story of a writer who got caught using an LLM to write articles about books to read this summer, and some of the books didn't even exist. Doctorow points out that this writer was expected to pump out a completely unrealistic number of articles like this. There's no way he could have written them all by hand, so, he turned to an LLM.

When I've heard this conversation about students at university, it is almost never that they are purely lazy or that they really believe an LLM will do a better job for them. It's that they are too overwhelmed and unable to keep up, so they decide it's better than nothing. This in turn puts more pressure on their classmates to do the same thing, because the pace of raw output, even at lower quality, is accelerating.

That's what we're swimming upstream against. We need to remember that it is a tool to help us, especially the most marginalized among us. We don't exist to benefit the tech companies, no matter how much they try to convince us otherwise.

What About Local Models?

One question that came up afterward was whether there was anything that gives me hope in all of this. My first thought was this section which I cut from the talk for lack of time.

Try not to only use whichever big tech company's AI is in front of you from the tools you already use. Google and Microsoft have big advantages in shoving it into everything you already use. OpenAI gets some credit as the first big breakthrough – some people see ChatGPT as synonymous with "AI" at this point – and that got going with backing from Microsoft. Anthropic is now often considered the best, especially for coding, and is independent but with backing from Amazon that got them started. So maybe try some other options which are out there. Try Nous Research, a nonprofit effort. If you are a bit more technical and have the computer to support it, try some of the free open models that can run on your machine. Those aren't using anybody else's cloud at all, like Deepseek or Qwen out of China or Mistral out of France. I'm going to say this generally, too: be cautious about how much you rely on a small number of Big Tech companies that will trap you using it.

The best case is that AI technologies move in the opposite direction of what happened with social media. Social media started out very open with a lot of smaller options, then it consolidated into a few giants who control everything and make it hard for you to leave. Mike Masnick is a writer I like and he recently made an argument AI might be moving the opposite direction, that it clearly started with a few giants but the local models and smaller cloud companies aren't that far behind. I don't think this will just happen naturally, though; I think it will require some cultural and maybe some policy decision-making.

Privacy

Speaking of sending things off to the cloud servers of Big Tech, how about privacy?

Yes, it is a real big problem that all our activities are constantly logged and tracked, and AI might unlock some new pieces of information that they can claim, but the root of that is the utter lack of competition or legislation to stop collecting all that data. In fact, in Canada, we currently have big pushes that would require companies to collect even more sensitive data about us, with the talk of adding age verification requirements as well as Bill C-22 that requires companies to keep much more information about our private conversations and keep them available to police with a backdoor for longer. Even if you assume that police will never abuse it, which is already a pretty big assumption when there have been contrary examples in the past, once you make a backdoor for police, you also create a backdoor that makes it a lot easier for bad actors. The best way to protect data is to never store it in the first place, which is counter to how a lot of our digital economy operates. So at a legislative level, try to get your representatives to care about data privacy, rather than buying into these panics to support a bigger surveillance state.

At the more personal usage level, though: don't put stuff in it that you don't want to be controlled by your employer and/or a US Big Tech company. I am sympathetic to why some people will use a free subscription to ChatGPT as a therapist or for medical advice. They may not have access to a professional, and this is better than nothing for their short-term need. But speaking broadly at the societal level, we really want to discourage that by helping get real solutions from real professionals in front of them instead. Otherwise, we are just handing even more sensitive data to these big tech companies to do whatever they want with.

Copyright

Another quick one that we can put in the broad category of possible abuses of Big Tech is questions of copyright. A lot of these systems are trained on copyrighted works. Is that legal? Is it fair? What if they pay the authors? Does a "right to read" that applies to people also apply to a machine at much higher speeds and for the profit of big corporations?

This is one area that I am not going to pretend to know a lot about, or even to have that strong of an opinion about, in terms of the legalities or what is best for everyone in the long run. I know for a lot of people, especially if you are an artist or a writer producing high quality work, this is just another way that Big Tech is consolidating its power at the expense of everyone else.

Environmental Impact

Let's do one more that it is at least partly about the consolidation of power for Big Tech. Along with copyright, environmental impact may be the area that comes up the most at the level of online memes expressing frustration at the concept of AI.

Yes, it is absolutely a problem that we are destroying the environment, and now we are seeing stories of how these Big Tech companies are demanding to do even more damage in the name of progress that looks a lot like them making more money. It feels way too familiar, like how fossil fuel companies have similar demands.

But I do need to say that AI data centres is one small part of our problem.

The exact numbers aren't always the easiest to pin down, but as I understand it, the water usage is especially overblown. I did the math once and figured out that it would take over 4000 queries to the most intense ChatGPT model to equal the same amount of water usage as eating one quarter pound hamburger. I know somebody else who did it and came out with it being over 11,000 queries to equal one hamburger. He might have been a bit less generous in his assumptions, or it might have been that he did his math a while later and the data centres were more efficient by then. He also did the math for the equivalent queries for producing one article of clothing as equivalent to about 256,000 chats and for a single cup of coffee to be equivalent to about 3500 chats. Those kinds of things, along with watering lawns and golf courses, are so much worse, but don't get talked about nearly as much.

Energy usage is a little more significant. I've seen estimates that AI data centres are now up to somewhere between 0.2% and 0.5% of all energy usage globally, and growing. All data centres put together, not just AI, are more like 3-4% of all energy usage. Video streaming is the worst part of that, not AI. For all the people who I see demanding that we never use any kind of AI system because of the energy and water usage, I never hear anybody suggesting that we all cancel our streaming services.

Those numbers are not nothing. I don't want to pretend that there is no concern here at all.

But if we are really concerned with lowering water usage and energy usage, I really think there are other things making up most of the problem that we should focus on first. I also think when it comes to energy that it is worth noting that the only reason more energy usage is even a problem is because our energy sources are not sustainable. If we had invested in renewable energy, needing an extra 0.2% of it wouldn't be a big problem. It's only a problem because now the stories are about gas plants needing to come back online to meet the demand.

It's also worth noting that these models are getting more efficient, with more specialized hardware, so I honestly don't think at the end of the day that this is going to end up being a big part of the story of how we destroyed our environment.

Cybersecurity

I won't go deep on this one, but I need to name that AI tools are opening up new security risks.

We are starting to see an acceleration of cyberattacks based on finding exploits in code. The better the coding models get, the better they are at finding those exploits really fast, and bad actors can take advantage without even needing to be as much of an expert themselves. The good news is that those coding tools will also help build code more secure in the first place, and Anthropic has released its best model to those companies on the defensive long before the general public. This has given them time to patch issues before bad actors can use the same tool to realize this is a possible place to attack. So, it is a bit of an arms race where we might see more attacks, but we'll see fixing the exploits faster and avoiding the vulnerabilities better in the first place. The best case scenario is that now that the people creating software have much better tools to find the vulnerabilities before release, there won't be as many vulnerabilities that ever get into production anyway, but there will be this tidal wave of suddenly finding a lot of issues first.

Here's a big one that is more at the level of each of us as individuals: suppose you get a phone call from a voice that sounds a lot like a loved one calling you in a panic and they need some money to get out of trouble right now. A lot of loving parents and grandparents are not going to think twice about it because they know the voice. Those voices can now be faked much better than they used to. One approach to that I've heard is to have a code word. If somebody wants your help in an emergency, they'll need to remember that code word, not just sound like your loved one. I'm just putting out there that this is a kind of scam that is on the rise.

Consciousness or Sentience

I'll end with the one that would typically get labelled as the most theological or philosophical that you might expect from a church discussion. Is it conscious, or sentient, or some other similar term?

These are terms that we don't really have any consensus definition of, so I can't really answer whether a computer could ever have consciousness or sentience if I don't have a clear definition of what those mean. Very broadly I might say something like: I still think there's something unique about humanity, some "image of God" ness, and I have no reason to believe the current technologies are going to ever do better than a mediocre impression of that. But I will fully admit that I say this without some clearly defined fully rational reason behind it, because even that "image of God" also has no consensus among theologians of what exactly that means – it has the same problem as consciousness or sentience.

My much more practical response to this one, which I think ties together everything I've been talking about today, is to talk about moral responsibility. When I hear some people talk about computers being sentient or having consciousness, it's usually tied to some idea that we now have a moral responsibility toward them, which in practice means some moral responsibility to keep funneling money to the Big Tech making them. Most often this is coming from tech bros who also seem to have a really hard time with any concept of moral responsibility toward those who are definitely actual human beings. Also animals. That's where I really have a problem with this whole line of discourse, when you are declaring the personhood of a machine at the expense of those who definitely are people, just not the kind of people who usually get prioritized. Maybe if we ever get better at taking care of humans and animals, then we can talk about moral responsibility to machines. Until then, I'm focusing on the real problems that really impact real people, right now.

Questions

From some of the following discussion, I'll add a couple quick thoughts.

One questioned whether we should be trying to opt out of technology entirely and going to live on a farm detached from the world. I don't think the goal is to be rugged individualists who can be completely self-sufficient, each in our bubbles detached. To me, the goal is interdependence, a community of equals who are able to lend our skills and resources to each other. A quote that has stuck with me from years from a conference was "the purpose of maturity is not independence, but interdependence." So yes, let's try to not be dependent on the wealthy companies to survive, but let's not hide on our own either. Let's try to use some of this technology in ways that build each other up.

Another question was what does some of this look like practically? Somebody brought up that it is really hard to not be constantly giving mass amounts of data to these companies. It's true; it is really hard. You realistically won't fully break free. You can pick some battles, though, with alternatives like Signal for messaging and Proton for email.