AI and content creation

<< Silicon psychopaths

Setting the scene >>

Hello again. As promised last week, this week I’m going to focus on the impact of generative AI in “content creation.” This loose definition will allow me to talk about my experiences of AI in my work, personal, and writing lives (they are all the one life, just to be clear). And, for further clarity, this is very much a collection of my personal insights. I’m obviously not an expert in AI. That hasn’t stopped most other commentators, however, so it won’t stop me.

But that important caveat means that I really (even more so than normal) would love to hear from folks who actually do know what they’re talking about.

To bring you up to speed, last week we talked about killer robots, a.k.a. the impact of AI in military operations and technology. I’ve read a lot of wonderful fellow bloggers give their two cents on generative AI, I’ve seen its utility and limitations, and am here to give my own perspective.

I’m going to talk about how AI itself is nothing new, rather a new linguistic framework we’ve put around a much older idea which has been with us for some time. Then I’ll talk about how AI can help with the creative process, before moving on to the downsides and my conclusion.

Before we start, let me encourage you to subscribe to the blog using the link below. As always, I’d love to see your comments below, or you can contact me via webform here or email here. Finally, if you’ve enjoyed this post, please share it with a friend.

If you enjoy this blog and want to support it, please consider a donation. Keeping this blog going doesn’t cost much, but it isn’t free either, so any help would be very much appreciated👍

AI is just a new term for “statistics”

In my day, we had AI, but we had a different name for it: statistics

—AI expert

The more technically knowledgeable is the person you speak to, the simpler the answer they give you for how AI works. The guy I worked with who said the above is one of the most knowledgeable of all.

Artificial intelligence is something which most of us are slightly familiar with and use on a daily basis (see next section if you disagree). Let’s define some terms, because I really want to focus on generative AI in this article, and it’s probably what you want to read about too. But AI is a much broader term, encompassing all sorts of clever computer behaviour. Here’s one way to look at the distinction:

The above is a bit simplistic; you can have generative AI that doesn’t use neural networks. But it works for most of our familiar generative AI ecosystem such as ChatGPT, Claude, Gemini, etc.

Now that we’ve defined it, how does it work? Let’s ask it:

That’s… not great. Maybe someone else out there has done a better job of explaining the process?

I guess not. It looks like someone just vomited up a bunch of half-digested words, pictures, and icons onto a page. Fancy some “Grearning,” anyone?

Right, time to give it a go myself:

I’ll back this up with a military metaphor. When I did my basic training, one of the first things I learned was how to shoot. We started with practical indoor lessons: here’s the rifle. Here are the names of parts. Here’s how to take it apart. Here’s how to put it back together. Here’s how to fire it (dry, still indoors, no ammo), and this is exactly what you do when you hear a particular order on the range.

After a few weeks of this, we were ready for the range. They stood us in a line facing individual cardboard man-sized targets. A little bit like the one below, but with fewer circles, no numbers, and a white patch in the centre.

They had experienced soldiers supervising each firing line and they went through the live practice exactly how we’d practised it indoors. To this day, if you gave me a Steyr rifle, a magazine full of ammunition, and something/someone to shoot at, I could do a decent job of it. This doesn’t happen very often, to be clear.

Okay, that’s one way to learn. But it’s boring and procedural and narrow and constrained. Let’s try the machine learning way. Bring the raw recruits straight to the range and hand them rifles and as many full magazines as they can carry. Put up a bunch of targets. They don’t have to be in a line; put them all over the place to simulate actual combat conditions. Tell the recruits to blaze away.

If they hit a target, reward them. Give them some food, maybe, or a morning off inspection. If they keep hitting targets, maybe a weekend pass and a promotion is in order. If they miss, punish them. You can shout, then escalate to punishment exercises. If they accidentally shoot an instructor or a hapless civilian out for a walk, feel free to put them on extra inspections and punishment details.

Pretty soon you’ll have competent shooters. They might not understand why, but they know that hitting those man-sized cardboard targets with the white patches is good. You can send them straight into the fight, where you can be sure they will engage and defeat armies of cardboard targets.

Obviously we don’t do this with military recruits, but it’s an effective way to train a computer when the stakes are lower, e.g. deciding which word comes next in a sentence. Do this for long enough, with enough data, and a powerful enough computer, and you have ChatGPT. It can simulate conversations with an astonishing degree of fluency, all based on the probability of what comes next.

Lots of commentators make the point that AI can’t think, it can only make inferences based on statistical likelihood. I don’t want to go down a philosophical and neurological rabbit hole, because I’m patently unqualified to comment, but I don’t think we can say for certain that human language understanding is substantially different. There are important differences, though, between you and an LLM. Angie Wang’s Is My Toddler a Stochastic Parrot? cartoon in the New Yorker (Nov 2023) is an excellent and emotional read. One quote from it is:

A toddler has life, and learns language to describe it. An L.L.M. learns language but has no life of its own to describe.

Generative AI is great at using language (and, to a lesser extent, images) to give back a decent average of what’s out there already. How is that useful for me, as a writer? Let’s turn to that next.

AI can help with content creation…

First thing first: AI can’t create content. All it can do is describe what’s already there. This does not make it useless. In fact, and this is sure to be a controversial opinion with some, it can be quite a useful tool for a writer or other content creator.

Here are some ways I’ve used generative AI in my writing

• Names in fiction. I write short and novel-length fiction, but one thing I consistently suck at is coming up with names. I don’t know why, but I do. ChatGPT is a quick and easy way to give me ideas.
• Military Realism movie examples. This blog is supposed to be about military tech in movies. Sometimes I get too wrapped up in the point I’m making (like today!) and forget to link it to film and TV. Other times, I might have one or two great examples of the trope I want to talk about, but feel the need for a few more just to flesh it out. ChatGPT can spit back ideas to make my research³ more focused.
• Pictures for blog. If you’re a fan, you’ll have seen that I sometimes use a ChatGPT picture in the banner or body of my articles. This is unbelievably handy when I can’t find or create one otherwise. Here are just a few examples. And yes, I know I should give the work to a starving graphic designer, but y’all don’t pay me enough for that.
• SEO titles and descriptions. WordPress (which hosts this site) has a handy feature where you can auto-generate metadata which is optimised for search engines. This way each post has a real title, which you can see (and is usually some stupid pun), and a search engine optimised one (this is the title which you see in the tab heading). It does the same for picture alt text: another boring but essential bit of website plumbing⁴.
• Explaining a concept. This is useful for worldbuilding. Like with the movie examples, it’s handy for my own background and just replaces a series of Google searches. Need a run-down on the challenges of terraforming Proxima Centauri B? The lifecycle of the boll weevil? Economic conditions in eighteenth century Bohemia? ChatGPT can give me a 95% answer, which, honestly, is about 15 p.p. more than I ever need.

There are many things I don’t use ChatGPT for. Most obviously, my actual writing. At risk of stating the obvious, I enjoy writing. Letting an LLM do it for me would utterly defeat the purpose, like subcontracting someone to go on holidays for me. That’s before we get to any questions of quality: let’s talk about those next.

…or really get in the way

“Would you believe this was written by ChatGPT?” Yes. I would. Too often with AI we fall into the trap identified by Samuel Johnson a quarter of a millennium ago:

The quality of the AI response depends on the quality of the input. If you spend time crafting a set of quality prompts and iterate this with the LLM, you will get a better output. At some stage, you’re probably better off writing it yourself.

With the use cases I mentioned above, there are problems with using AI. ChatGPT usually hallucinates scenes from films to match what I’m looking for. This has led to more than one frustrating rabbit hole where I’m searching online scripts or clicking through feature-length films to find one perfect scene to illustrate my point, before realising that it doesn’t exist. A simple Google or YouTube search can be more time-effective. Sometimes I just need to accept that I don’t have the breadth of examples I would like.

Pictures are another pain point where I often wonder whether it’s worth it. I spent at least fifteen minutes last week trying to get ChatGPT to make the cover banner image into a long and thin picture. Despite telling me multiple times that it could do this, this is something that’s outside its capabilities. I liked the picture, but it still looks crap as a banner:

Unlike with a test output, the first picture you get from ChatGPT is usually the best, because each subsequent one consumes and regurgitates the previous output in a trend toward mediocrity and generalisation.

It reminds me of South Park’s “Time Immigrants,” and not just because “They Took Our Jobs!”:

Mediocrity isn’t the only problem with AI, however. Because it’s a massive bubble whose unprecedented investments need massive future returns, technology platforms are pushing AI “features” down your throat.

Microsoft’s Copilot tool follows your cursor around. Meta’s AI WhatsApp feature floats next to all your chats. Adobe is the worst of all, in my opinion. Their “AI Assistant” feature takes up half the window of any new .pdf I open. Yes, I’m short on time, but no, I don’t need a generative summary. I need this box to disappear so I can scan the QR code it’s blocking at the ticket barrier.

As a final note before I finish, I should mention data privacy. This is the big one. I don’t use ChatGPT for anything I wouldn’t want the world to know, and I would urge everyone else to do the same. No financial data, no names or photos, and nothing else which I wouldn’t want to see on the front page of The Irish Times. It doesn’t matter that I’ve turned off the “help improve the model for others” option: tech companies have proved time and again that they can’t be trusted with stuff like this. I don’t mind using ChatGPT for ideas related to the blog because that’s freely available online the minute I publish it.

Conclusion: Like with all technology, be aware of AI’s limitations

AI is designed to please us… even if it’s wrong. Think of it like a really enthusiastic but unreliable assistant. Or, in military parlance, a Russian junior officer who is eager to tell his superiors that the attack has gone well (when it really, really hasn’t):

AI can be a useful tool. Like all tools, however, it has its limitations. Writers and content creators are used to using technology judiciously:

• Spelling and grammar check is a useful tool⁵ to make sure you write clearly and legibly. It can give false positives when it refuses to recognise a new word or a name, and false negatives when it fails to spot a homophone. It helps the writer, but is no substitute for proofreading by the author and others.
• Wikipedia is a useful tool for research, saving the writer the necessity of trawling multiple sources or (shock, horror) going into a library. It helps the writer, but is no substitute for reading extensively and carrying out one’s own research.
• Computers are a useful tool for writing and editing. They can save forests’ worth of trees in drafting and re-drafting, but can also contribute to a paralysis of constant editing. It helps the writer, but is no substitute for concentrating on a draft and just writing.

Most of the things I’ve said I use ChatGPT for could be replaced by the simple act of talking. With friends, family, or experts; there’s not much I can get with ChatGPT right now that I can’t with other people⁶. Generative AI can turbocharge your output by giving you a shortcut to this long and complicated process. However, the biggest risk of this tool is that it makes you turn inward and stop interacting with real people. This is a deeper AI problem than writing or content creation, but it’s one that we need to be just as conscious of.

That’s it for this week, folks. I wanted to include AI in business as part of this article, but there was too much to cover with writing alone, so this will have to wait for a third instalment. Watch this space. Thanks for reading and please remember, if you haven’t already, to subscribe using the link below. Please also share this article with a friend. See you next week.

Featured Image: Indiana Jones and the Last Crusade, Lucasfilm

1. Prompt: “Can you draw an eye-catching graphic that illustrates how generative AI model like yourself work? Not too detailed or technical, but with the broad brush strokes technically correct. Would love to see what you come up with!” ↩︎
2. Prompt: “Can you generate a cartoon image of recruits on a firing range, except the targets are in every direction, some recruits are firing wildly into the air or don’t even know how to use their guns, and instructors are ducking for cover, shouting, running away, or even getting hit. It’s chaos!” Reply: “I can’t create or depict violence, injury, or people being harmed, even in cartoon form — including scenes where instructors are getting hit or under fire. If you’d like, I can generate a funny, chaotic cartoon of a firing range training gone wrong — but with a lighthearted, slapstick tone (e.g., recruits aiming in random directions, instructors panicking or waving flags, smoke everywhere, no one hurt). Would you like me to make that version?” Me: “Yes please. The getting hit isn’t the important part, but the sense of comic mayhem and danger is. Is that okay?” ↩︎
3. Yes, I’m calling watching YouTube videos “research.” ↩︎
4. There’s an irony here in using AI tools to improve your article for web search, which AI is simultaneously killing. ↩︎
5. And, as it happens, an example of artificial intelligence which is not based on neural networks. Rather, it’s based on IF–>THEN grammatical rules. This limitation is clear when you’re using an edge case which the grammar check uses a general rule to try to “fix.” ↩︎
6. If you want some metrics: 100% of my blog post ideas have come from other people; 25% of my TV/film examples come from other people, 25% from my own viewing, about 40% come from TVTropes, another 5% from Google, and the remaining 5% from ChatGPT. ↩︎