Check the end for free giveaways from indie authors!
Do you actually know how something like ChatGPT works? If the discussion emerges, do you have the language and the knowledge of the structures and toolsets behind the general “AI” umbrella to break it down into its composite parts and arguments? Do you know what AI…is? Do any of us?
Let’s consider a few different vectors when talking about working with “AI” as a broad toolset.
- The product – The thing that is created with the “AI” tool.
- The human ethics – Whose labor (if anyone) is being replaced or exploited with this tool?
- The environmental impact – Comparative water and electricity use to other tools.
- The reliability of that tool.
- The ease of use for the individual.
- The “human growth” element – The new information or skills gleaned from creating the product.
- And, as a comparison point, the tools that already exist to create the same product.
As our first thought experiment, let’s pretend we’re writing a blog post with Microsoft’s current generation built-in speech-to-text software. Later, we’ll compare it to ChatGPT in a general sense because they’re both, ostensibly, part of the same Azure OpenAI ecosystem.
The first phase is using your microphone to collect your audio, then analyzing it via acoustic modeling through a neural network (keep in mind we’re simplifying the hell out of this). It then combines this acoustic model with machine learning and a language model (small or large as the case may be).
This lets you open a software of choice, set the cursor, speak, and the words will appear in the writing software with correct homophonic interpretation. Ie: it can figure, via context, which “there/their/they’re” you’re using.
To create this same product without the use of AI, you would do so by typing out the words manually on your keyboard of choice or by using a hard coded logic-based language model software like Dragon NaturallySpeaking. There are also some additional products that are non cloud-based AI language models that you might use.
But every one of these methodologies produces roughly the same product. It’s your words on the screen. The generation of the product is wholly your intellectual work, and, more importantly, intellectual growth, expansion of knowledge, and semantic practice. You’ve become better at the task of writing if only a little bit.
It is just the physical tool that is different. The reliability of each of these tools might vary a little depending on access to the internet for the cloud-based language models or the functionality of your hardware. As will the ease of use depending on your specific physical circumstances. Those would be very personal things to consider not broad trends.
So not a lot of differences in the AI vs non-AI tool across those vectors.
What is the human ethical vector? In ye olden times if you needed something written but you didn’t want to or couldn’t do that particular physical labor, you would hire someone that you would dictate to. That service still exists. Doctors can speak into a tape recorder and there’s someone out there listening to it and transcribing it. Realistically, though, this is not the kind of thing a normal, everyday person has ever had regular access to. So if you are one of those everyday people who needs to utilize dictation, using an AI-based software isn’t replacing anybody. You weren’t going to pay someone to begin with. So AI isn’t taking anyone’s job in this capacity.
Then we have to address the language model itself because language models are trained on a corpus of work. From what both OpenAI and Microsoft claim, this body of work is either public domain or legally licensed. They aren’t “stealing” any work to train the model on in the traditional sense, but you could probably make a case for the transgression of End User License Agreements and reasonable expectation of not being scraped. If I post to A03, I didn’t implicitly agree to my text being used to teach a language model, but I also don’t recall clicking a box that says “don’t touch my shit.”
That’s a very complex issue that we’re all still trying to catch up with both ethically and litigiously.
In terms of environmental impact, there’s still a lot of information the public doesn’t have access to. But we do know that in comparison to non-AI tools, the extra computational power consumes more electricity and water for cooling. So you’d have to make the decision whether replacing a manual tool with an “AI-driven” tool is worth an unknown environmental cost to create the same product.
And that’s a conversation that has a lot of nuance. A prosthetic limb uses plastics and alloys that a biological limb doesn’t, but that doesn’t mean you deny people the ability to walk just because it uses resources. So if a dictation software lives in that same carousel, then it has to have the same balance of pros and cons. There’s just a lot there to think about and no easy answers.
But what if instead of AI-dictation, we “write” our theoretical blog post with a prompt into ChatGPT. This is a software that uses the same LLM and, therefore, should have roughly the same environmental impact if not more (because it’s wholly generative).
So those are comparable vectors. What about the others?
A bit of writing from ChatGPT is middling to bad. At a line-to-line level, it’s of median quality compared to every other piece of writing in the world. This is because it’s, ostensibly, just copying writing that already exists. ChatGPT can’t think. It just recognizes patterns in pre-existing written work. It’s well-know that it’s prone to internal inconsistencies and semantic hallucination. It does not reliably produce a high-quality product compared to what a human brain is capable of.
And while, in this specific scenario, the only person’s labor is your own, it’s built on the back of pre-existing work. You’re essentially just collaging together semantic ideas formed by other people instead of creating your own. While at the same time denying yourself an intellectual growth opportunity.
So the trade-off for these things is that generating text is supposed to be “easier.” But if if you talk to die-hard advocates for “generative AI,” for lack of better phrasing, they’re very insistent that prompting is a skill unto itself. So let’s say we throw them that bone, that prompting requires specialized knowledge and, while it’s not hard, it isn’t “easy.”
So you’re putting in some manner of effort and using large caches of resources to put together words that aren’t really your own to create a mediocre product and you didn’t even learn anything in the process.
Then what’s the point? The benefits of “generative AI” over other tools start shrinking under the first pass of workflow scrutiny.
And that is the conversational space where we have to live when we start talking about the deleterious nature of certain AI “tools.”
Because there are arenas where things like machine learning and neural networks (things we think of as AI) can perform tasks that humans can’t. A computer-vision operated fish-lice laser (monitored by a human) is going to be faster and more precise than a human operator. And if it can learn to be more precise as it goes, fuck it sure, why not? Computational edge detection was the thesis of my pre-college internship in 2007, and I literally published research on it. That shit is wacky and a prime candidate for computational learning in a way that is actually useful at scale.
Grammar and spell checking used to be just hard-coded logic driven. Now they’ve transitioned into AI-driven. In some ways it’s better, in some ways it’s worse. In either case we’re stuck with it.
The average anti-AI person still recognizes that there are going to be use-cases where AI actually is a superior tool compared to pre-existing ones. They’re willing to introduce nuance.
The hyper-polemic AI-advocates don’t care about this conversation. They’re not going to hear “you’re creating half-stolen, intellectually lazy mediocrity at the expense of increased resource expenditure and that’s not the same as a cancer-detection model.”
The conversation has to be for everyone else.
The parent who uses Claude to come up with a fair, randomly assigned chore chart for her kids when the average Girl Scout was doing that at the age of nine with slips of paper pulled out of a hat. It’s the person who uses ChatGPT as a search engine when Google, at it’s dumbest, gives you the exact same information at a tenth of the resource cost and backed up by an actual, citable source so that you can verify the information is correct.
We have to reach the people who are offloading their thinking to a language model, and in the process, forgetting how to use their brain. If we’re going to do things like fight fascism, we can’t lose our access to critical thinking. And the way the average person seems to have sunk into AI, that’s a real risk.
Collectively, we’re too stupid to have a good grasp on how we can use things like LLMs, neural networks, diffusion, and machine learning in a way that actively benefits us long term. We’re throwing it at the wall to see what sticks, and manual creative and thinking systems are falling apart underneath it.
Being able to have that conversation about the nitty gritty of what AI is and what it’s actually doing isn’t the solution to “good” AI use, but it gets us closer to figuring out what that “good” AI use actually is.
Free Reads
Check It Out!
