A couple of weeks ago I decided to stop using AI image generation. You can read about that here. I want to know more about your attitudes towards AI generation, whether text, image or anything else. Do you use them currently? Do you see them becoming useful? Or do you see them as a threat or problem?
No judgement here - I just want to take the temperature of the room on this one. Of course, there are a lot of writers in the community, but I’d also love to hear from anyone on the other end - as a reader, or viewer, what is your response to encountering AI creations?
No, I had issues with the AI tools from the beginning. I know several prominent artists in the SFF community, some who work with abstract digital tools, and they were all concerned about AI image generation, because of the plagarism aspect of training those tools.
Then ChatGPT emerged, and the use of it by some of those "get rich quick" sorts promoting the use of it to make money fast in writing. This piece isn't new, either. Last summer there was much discussion in a Kindle Vella forum I'm part of about how one company was spamming Vella and generating false buys in order to game the Vella bonuses.
The fact that ChatGPT and other writing AI have been trained using the writing of others without express permission is extremely problematic. The degree to which some writers are already using these tools in generative work is extremely unethical and problematic, and I'm going to call out people who use AI for prompts, expanded thesaurus, or "different perspectives" on their work. They're calling upon a tool that has been unethically trained and putting ethical concerns aside, in my opinion. I've also read and heard examples of the alleged "help" from AI drafting.
It homogenizes the voice of the writer. It purges their unique perspective into a mishmash.
Don't get me wrong. I don't fear the AI. It is stronger than human intelligence in crystallized intelligence--knowledge acquired--and in processing speed and memory retrieval. But this is a mechanical superiority over HI, caused by processing power that can grind away at the cognitive wheel. HI is superior to AI in the application of fluid intelligence--that intuitive and deductive leap that creates out-of-the-box solutions to all sorts of problems. I'm currently reading some psychometricians (psychologists who study cognitive assessment measures and devise them) on the subject of AI, and that's the leap that some believe AI cannot take.
Even given that HI is superior to AI in fluid intelligence and will probably remain that way, there is still the ethical implications of using an unethically trained tool. AI will not produce a good work of creative writing soon, if ever. The reports from Clarkesworld editor Neil Clarke and other discussions I'm privy to in a private forum suggest that the AI works are poorly written and easily identified.
But. AI has many potential uses in the creation of instructions, generating certain types of promotional text, and standardizing certain documents. Were I still involved in special education case management, I'd be advocating for its use and training to construct Individual Education Plans, simply because a.) it has the potential to reduce the heavy paperwork load on those individuals and b.) the quality of some of those IEPs is staggeringly awful. It would become possible to change IEP updates from a process taking an hour to several hours to a much quicker, uniform, and more useful document.
HI will still need to be used to check and correct this sort of writing. But that is AI's potential in writing. Not in creative work.
That said, every piece of creative work I am sending out from now on, whether a self-published novel, a Substack post, or short story cover letter is going to contain the following statement:
"No generative AI has been used in the conceptualization, development, or drafting of this work."
It's frustrating the way new tech at the moment seems to be so easily hijacked. Much like how blockchain concepts were shoehorned into NFTs and the various scams and delusions associated with them, now ChatGPT has been immediately co-opted by people trying to make a quick buck from churning out bad writing (which is especially amusing, given how hard it is to make money from writing...).
I use all of it and I’m obsessed. ChatGPT has been so much more useful than Google for researching my essays and novel chapters. Even if I have to fact check, it’s better than clicking on a hundred links that are all SEO content that don’t contain what I’m looking for. And yes I have been loving Midjourney for creating images for my novel chapters!
Elle, this is one way I use it as well, as a replacement for search engines. It's a starting point that refines a vast amount of data into very concise points. I then take it from there and do more research to track down sources when necessary.
I've not really seen how to use ChatGPT as a search engine, as I don't find myself in a position where I can trust its responses - and it always delivers responses with utmost conviction, making it even trickier. Whereas 10 links in a Google search at least have immediate context: it could be an actual academic paper, or a piece of journalism from a respected publication, or it could be an opinion on a random blog, or a crowdsourced Wikipedia article. All with their own merits and issues, but adding up to a broad understanding of a topic.
I don't quite see how ChatGPT replaces any of that, other than through a sort of filtered merging, with an added element of uncertainty.
Conceptually I can understand it as a starting point for further research, but when that starting point is on such flimsy foundations I'd be worried about being pointed down a false path from the very start.
You have to treat it like you're the professor and it's your student. For example, I formulated the following query: "What are the prevailing theories on how the pyramids were built? Please cite one source for each theory in question." It did exactly that, describing each theory, then at the end providing one source for each. That's not possible with any search engine with any sort of coherency.
I can see these tools being very useful eventually. At the moment, I think they are quite problematic - sharing misinformation, using copyrighted artwork, etc - and require some real oversight. I don't use them and don't really want to, but I accept that people want these tools and they will be used.
Agreed - I think that's largely where I'm at, currently. I'm not against using such tools in the future, especially if they can be developed in more convincingly ethical ways. MidJourney for a time I found really interesting and useful as a concept engine.
The problem, really, for me, was that I slowly slipped into relying on MidJourney to quickly make an image because I'd run out of time - rather than using it as a point of inspiration. I'd imagined using it to help with my own art, but that didn't really happen. The lack of intent behind MidJourney images also is an odd one, in terms of using it as an inspiration source.
I am currently too lazy to learn how to use it. But I am sure it will get easier. I imagine AI will be great for most people until it takes over and eliminates “flawed organic life.” Thankfully, I am already old. 😉🤓
I dislike generative AI as an artist -- the way they've been trained and used to this point has worsened the already ubiquitous problem of art theft and the devaluing of the art people create. Anything that encourages an environment of disrespecting artists on any level is a no-go in my books. The culture of AI is not something I want to be part of.
I wonder if there's a way in the future for an AI to be trained on your own material. If I could train an AI more directly on my work (written or visual), and then have it help me with concepts, that could be an interesting tool. But that doesn't seem viable - they seem to rely on massive datasets, which wouldn't work for individuals.
Or, indeed, whether an AI could be trained using actual tutoring like humans receive when studying art. A more deliberate training, in other words, without the more plagiaristic elements.
Yeah, the way current AI work doesn't jive well with small datasets, and I know there's no way I could produce even 30 images with a consistent look, nevermind thousands to train an AI on; not only would I get tired, but my art is always changing as I develop. The best way to achieve this help right now would probably be good old-fashioned photo-bashing and moodboards! These also have the advantage of being made by you as the person who knows what you want -- no repetitive reiterations to find something vaguely right. The ends have always existed, the means just take a bit more time.
If you wanted an AI to work on a smaller dataset to give you something you could have made, I imagine you'd need to rebuild generative AI from the ground up -- the fundamental concepts of the stuff coming out right now is always going to depend on scraping the Internet for masses of data. As such, although AI generally may be theoretically useful in the future, the versions we have now just aren't something I consider ethical and probably never can be on an inherent level.
It does seem that way. Though I also get that I'm speaking mostly from ignorance, when it comes to the technical aspects of how these things work. :)
Photobashing, though - yes! Since stopping using MidJourney I've reworked the front cover for my book, and went the photobash route for that. Found a bunch of stock images, bashed them all together, and then did hand-drawn (digital) line work and watercolours over the top. I was pleased with the result and a) it was much more unique than anything I'd produced with midJourney last year and b) it FELT much more 'me'.
There is a difference between using it as a tool and using it for everything to not do any work. I can already see that there are plenty of people who are doing exactly that.
I've used writing tools a little bit to "prime the pump" on some writing for my day job. Often they just got me unstuck by showing me exactly what I didn't want to say. Useful, I suppose.
I guess I just find the work they produce uninteresting and they kind of miss the point, at least for me. The point of art, one of the points, is to facilitate a relationship between and among people It's not just the thing, it's the person and their world and the interaction therein that led to the thing that is now in conversation with a different person in a different world.
I suppose sooner or later we will become more machine and machine will become more human and we will relate to one another as equal-self to equal-self, but we're not there yet.
I think that's the crux of it for me, too. AI generation really needs to be done by a properly sentient AI (whatever THAT really means) for it to be of interest to me. Otherwise it's just noise. Art is communication, which is about connecting different minds. If it's not coming from 'a mind' it's lacking something for me.
I used the Dream by WOMBO app for quick image generation for my substack posts, then read negative articles about its effect on artist and theft of styles, so got uneasy and quit. I still use filters that make my photos look like paintings.
I use other forms of AI a fair bit. I have machine learning-driven stuff help my computer run video games faster. Another tool cleans up my audio for Zoom meetings and automatically hides anything that isn't human voice. Those tools aren't pretending to make art, though, and that's where things get a bit weird with the likes of MidJourney and ChatGPT. There's a fundamental shiftiness at their core, where they're pretending to be something they're not.
And if I understand how ChatGPT works, by predicting words based on the analysis of an impossibly vast amount of PAST words, then the doomsayer in me worries that what seems futuristic will paradoxically beholden our grandkids to the past even more than our current crop of nostalgia loving human creators do now!
Part of me gets it. Mashups are cool, who doesn’t want to read Hemingway’s werewolf story!
Also, it seems we’re at a point in literary history where culturally diverse voices and styles finally have as much respect as “the canon” but I worry if we’re not careful, a flood of AI-generated text could wash away those gains and fill our screens with recycled-Frankenstein-type content instead.
Confession: I did ask ChatGPT to tell me a werewolf story in the style of Hemingway and it was neither a good werewolf story nor a good Hemingway homage. Ergo! Because now I really really really want to read Hemingway’s werewolf story, it’s become one of my many works in progress. I just have to reread a bunch of werewolf stories and Hemingway stories to prepare and research. And even if my story isn’t any better than ChatGPT’s - I’m very much looking forward to the research/prep I need to do to attempt it. I won’t let a software program take that away from me :)
I use midjourney-generated images for the main illustration of my substack posts. I see it as a tool to express myself better. Just like a painters brush + paint are tools that make it easier to make a drawing than painting with bare fingers and charcoal.
I also use it to make coloring pages for my kids based on their own desires/prompt ideas.
I’m not really for or against in general. It really depends on the use etc. I do feel that when someone uses a specific artist’s style in the prompts they should mention them as a source/inspiration at least.
The fact that I am a digital artist with sensitivity to all things visual may be a factor in that I 1) didn't like the process of generating AI art, and 2) don't find it appealing to my eyeballs.
As for the text generating AI tools; it's only been 3 weeks since something prompted me to experiment with ChatGPT and I use it on some occasions.
For some things, it's faster to get what I'm looking for than by using Google. And I also like to take it as a 'brainstorming pal'.
But I do not like and can't personally use the text it produces - it's not my voice and I have a strong repulsive feeling in my body to use those formulations as my expression.
A hot topic for sure. I dabbled a bit with Nightcafe and stopped because the output is too generic and the anatomy and architectural perspectives are always messed up unless you use starting images. I did do some "evolving" of existing real images to make them look futuristic. The results were mostly goofy but sometimes excellent, and that's where people get anxious, and rightly so. A new artist putting up his oil paintings at a small local art gallery can find his work evolved online the next day, thanks to phone cameras and AI… not cool.
And ChatGPT? For writers, Dramatron might be more interesting. Although, ChatGPT has its uses as a tool to get information which you can also get via Google. I let it finish some text once, and it was able to come up with something generic. Then again, looking at the media landscape, would you always be able to tell which script has been written by ChatGPT?
I tried to solve a problem coding a health bar with hearts for my Interactive Fiction Demo in Twine with ChatGPT. It proposed many solutions, none of which worked, and every line of code it produced did not follow the correct syntax. I kept training it and made sure it had the latest release versions and story formats, to no avail. Some solutions looked cool and elegant, but the code was rubbish, and I ended up losing three hours and then did it myself in 10 minutes. You can check out the result on my substack (first post).
In the end, ChatGPT, Bard, etc., are tools that are not going away, they are only going to get better and if you wonder what's out there right now? Well, theresanaiforthat dot com.
Yeah, you're absolutely right that AI generation isn't going away. It's going to be fascinating to see it develop, and also observe where it is adopted (and where it isn't). It's all coming in so hot at the moment that it's impossible to really tell where it's headed.
AI is already everywhere, ad campaigns, copywriting, movie editing, scriptwriting (dramatron), research, you name it, it's there. Not using it will be a conscious choice you may be able to afford, but companies will look at the $ and decide they don't need testers or designers anymore. Teachers? AI. Doctors? AI. Presidents? AI... my initials... ;)
I think it depends a little on the application and tech, as not all AI/machine learning stuff is the same. The generative AI we've seen in MidJourney and ChatGPT relies on the huge data sets, which is where the ethical problems come in. AI-enhanced tools that have been in Photoshop, or like DLSS or noise cancelling stuff from NVIDIA, I see as being slightly different and far less problematic.
I can see it birfurcating sooner rather than later: strands of society and business that go heavily down the AI route, where efficiency and process matters most. And then another strand where creativity and human connection matters most, where AI will play a more minor part. I'm not saying that the latter is 'better', incidentally - just different.
The question will be "is this thing supposed to form a connection of some sort between people, or is it supposed to get a job done?" And that will determine much of the application.
It's getting stuff done (?) for now. For connection forming, this "thing" is not advanced enough, yet. Although Japan has invested in automating elder care for over two decades now (see MIT Tech review), they haven't made any significant progress. And if there is an Elder Care Bot waiting for us, it's not a comforting thought...
A word of definition - I'll assume, Simon, you're specifically referring to GENERATIVE AI tools - those that seek to "create" with minimal input from a user. I make this distinction as, among other tools, I'm typing this on a smartphone using SwiftKey - a keyboard app which can be set to read one's emails and social media accounts to learn how I wrote and use AI to improve its autocorrect and predictions.
Thus, yes, I do use multiple AI based/enhanced tools, from SwiftKey to On1 PhotoRaw to Vegas Pro and Topaz Video and beyond.
Here we get into different "classes" of AI tools. On1 PhotoRaw uses AI based tools extensively - especially with version 2023. However most of the tools are based around "smart masking." Let's say I have a photograph of a person standing on a beach, looking out across a lake. On the far side of the lake are hills with some buildings. PhotoRaw now uses "AI smart masking." If I select the "AI superselect" tool it will attempt to isolate elements as I mouse over them - the person, sky, foreground beach, background hills, buildings, water; even the foliage. Once an element is isolated I can either left click then move my mouse over another element to add to the selection, or I can right click and assign effects or filters to that element. In practice, this means if I want to tweak the color balance of the trees, I mouse over the trees, right click and add the color balance effect. However, this is basically speeding up the tedious process of masking (note also the AI masks aren't always perfect, but I have manual masking tools to clean up). The software isn't making all the decisions for me.
Ok, fine, there's an AI exposure tool, but I rarely use it because the AI doesn't adjust images to my personal taste. I may use AI exposure in culling to get a better idea of what an image can look like, but I reset and redo shots I'm keeping.
But PhotoRaw even now has "AI presets." Click a preset and it does X to the water, Y to the land, Z to the sky, A to the person, B to the buildings and C to the foliage. Pretty sweet. Of course it's applying masked, non destructive filters to all the elements, and I can go in and tweak everything. Still, I LOVE these. Now I can work on a single image and really quickly stick the same looks on everything else in the set without tedious masking or copy/pasting filter settings. Thanks, AI tools! You've removed most of the boring bits and let me get onto the fun bits!
That's the type of AI tool I like. It's doing the annoying tedious bits for me - pretty well, usually 99% there, sometimes only 80% - and letting me get onto my creative work.
Generative? Not for me. I'm not the best artist, photographer, editor, anything-else-I-do by far, but I worked to develop what skills and techniques I have. With a Generative thing I don't feel I have control - because I don't. I also get bored because I'm sitting waiting for the random AI thing to spit out, and, more often than not, it isn't giving me what I want.
Simon, I like to think our discussion on WhatsApp two days before you published this influenced the article. For any not-Simon person who may read this, the discussion was on how certain of my friends were admiring some YouTube channels that do quick, boring slideshows of AI generated content. Think "Star Wars, but it's 1950's."
Not only are the images generative AI, but the videos are straight static slideshows without the editor even bothering to "Ken Burns" pan/zoom or even crossfade between slides. They're straight static images with no rhyme, reason or attempt to order the images into a narrative flow. They are likely the images taken in the random order generated, dumped on a timeline and rendered out and it's so boring and utterly fucking lazy that I'm actually offended this utter shit is garnering 10's of thousands of views.
So to prove a point to these friends I allocated one hour to generate as mamy "Avengers, but it's 1920's" images as I could. To take a step of effort beyond what these YouTube channels did (where their "Princess Leia in the 1920's" images all had different faces) I even specified actors! Errol Flynn as Tony Stark, Lou Costello as Hulk, Eddie Anderson (best known as "Rochester" from Jack Benny's shows) as T'challa. Among other things the limits of the AI model's training became clear. Let's just say the AI didn't know who any of my non-white actors were, and Stable Diffusion has only one black male face - high cheekbones and a goatee. The racist training of the AI is a different topic...
But, to finally get to my goddamned point after paragraphs of setup, I moved on to "Avengers, but it's 1950's. Here we ran into issues. I gave myself an hour to generate prompts. 45 minutes were wasted, because no matter which prompt I used the AI was utterly unable to ever understand that "Peter Lorre as Hulk" means "Give me a shirtless, green muscular man with Peter Lorre's head." (Yes, I tried that as a prompt.) Basically I ended up both bored after an hour of watching a scrolling Discord thread, AND annoyed I didn't get the image I wanted and PISSED because I could have done it myself in Affinity Photo in a couple of minutes.
I don't like the AI tools because they aren't very good. I'm better. But I'm slower. I've already lost work to generative AI tools. A friend working for a major game studio is fighting to keep 50 jobs from being replaced with an intern and a generative AI. I wouldn't give a rat's arse if people used the tools, but the bottom line is they are going to cost jobs (hey programmers, ChatGPT can code!)! Which only matters because we remain on an antiquated, millenia old model across most societies of tokens (money) earned through labor for food and shelter, and AI art tools, AI self driving cars, robotic factories, et al remove jobs without creating new jobs to replace them.
These disruptive technologies have appeared for centuries - sewing machines took work from seamstresses, automated factories took work from laborers, synthesizers took work from musicians (don't argue with this, the average Broadway or West End orchestra was 50-100 musicians before 1980, now it's 15-30. The synthesizers and backing tracks took over half the jobs). The AI revolution will take more.
And that's a real problem when we're on the "work to eat" kick.
Yup this comment just took a turn into advocating a "Universal Basic Income" brand of socialism!
Anyways, we've also discussed, ad nauseum, the ethical connotations of the training models, via scraping of copyrighted work.
I'm thinking of recent Corridor Digital experiments where they got it right and wrong at the same time.
They trained a Stable Diffusion model with their own faces so that SD could consistently create the Corridor Crew in other images. This is kind of "right" because the AI was trained on custom images they took.
Then they used the AI to turn live action into anime, but they trained a custom data set they scraped from a specific anime - and justified every fear the working artist has. They specifically, willfully and with premeditation took copyrighted worked and trained an AI to emulate it to do themselves in the style of the AI.
I don't THINK Corridor meant to offend, but they haven't fully thought through this. Man, they're literally getting worldwide bad press for the project. I think they'd rather get positive press for things like Wren's amazing scale videos. Niko's comments about AI "democratizing art" don't help.
For one thing, the hardware requirements for custom training an AI model are staggering-they are using $8k+ machines-and the other option is buying runtime on a Google cloud. Most people use Google Colab, thus does the trillion-dollar conglomerate make money.
Side note: OpenAI, maker of ChatGPT, is an Elon Musk company. He's Co-founder of the company behind one of AI art tools... I think Dall-E. Using these tools supports evil. Just saying.
Digression over - "democratizing art?" Call me elitist, but, NO! Develop some skill! Practice! Learn! If you aren't any good, maybe you shouldn't do it! Typing a prompt into an AI and taking the image or text it spits out doesn't make you an artist or writer. Practice might.
AI leads to what I call "Next Gen Syndrome," because I figured this out as a teenager watching Start Trek: The Next Generation. The rule is: "Overreliance on computerized tools makes you stupid and lazy." Because the Next-gen crew are basically idiots most of the time. There's the episode where the Enterprise contracts a computer virus and nobody thinks to reboot the computer and re-install from the clean backup. At all. Until Data "dies" of the virus then suddenly sits up and announces his brain has rebooted off the clean backup. There's the episode where these crystal node things are growing on Enterprise's systems and the holodeck becomes Enterprise's "subconscious," and our heroes walk up to the holoactor putting together a jigsaw puzzle showing Enterprise overlaid with a giant crystal, and the first officer, counseler and chief engineer reveal they don't know what a fucking progress bar is because none of them have any idea of what this puzzle could POSSIBLY represent. They are idiots because they let them computers do everything.
I see it in people today - and that includes myself. I used to be the waking phone book. I knew all my friend's numbers, work numbers, locks business numbers. People used to call me to get everyone's numbers cuz I had then memorized. Now, of course, my phone has a phone book and I haven't bothered to memorize numbers. Relying on the tech made me stupid.
As a new generation of "artists" starts using more generative AI their work will become bland, boring and stupid. We can't stop it. We're already seeing it on YouTube (chasing the algorithm) or pop music (an analysis of music from the 1920's to the mid 2010's objectively proved modern music is less complex in tune, tempo and structure than music before 1990). We're seeing it in TV shows and movies planned by committee. Generative AI is only going to accelerate the homogeneity of art into boring blandness.
Yeah, Corridor made a big mis-step with that anime video. Felt very queasy watching it, and the flimsy claims of 'democratisation'. And the end result wasn't great, either.
That's definitely what I liked about MidJourney, at least for a while. Over time I felt I was being sucked into that homogeneity, rather than it actually enhancing my art - but that is more to do with me, I think. I've ended up in a place - for now at least - where I'd rather manually gather stock images and inspiration and compile vision boards. Takes longer, but feels more deliberate and purposeful.
I have started to use it more for software programming and I can see the value immediately. While associate programmers with only a few years' experience may not find it as useful, because you don't know what you don't know, I can very precisely narrow down a problem domain and have a polished solution with little editing. The code is production ready in minutes vs. hours. In the past I would have been concerned about this progress, but it's a large leap from that to AI is going to steal my job.
One thing I mentioned in my recent post about this, is simply the disruption it's going to cause in traditional publishing. Con artists will find ways to manipulate systems using AI as their slave assistant. They will overwhelm the little guys who can't keep up with the influx of submissions and attempts to make a quick buck. I'm trying to recognize the trend and make adjustments to my strategies as a writer.
That's interesting, I've read far less about its applications for coding.
I listened to Ed Miliband's podcast today in which he was interviewing Bernie Sanders, who briefly touched upon AI and the risk to jobs. Sanders' point was that AI tools aren't a problem, and tech is never the problem - it's the way they're implemented and introduced into society. If AI tools can be used to help workers be more productive in the same amount of time? Great! If they're used to reduce the working week to 2 or 3 days while maintaining wages? Great!
Of course, that's rarely how it goes.... (and this is where we circle back to the problem with billionaires and megacorps being the ones investing and inventing this stuff)
No, I had issues with the AI tools from the beginning. I know several prominent artists in the SFF community, some who work with abstract digital tools, and they were all concerned about AI image generation, because of the plagarism aspect of training those tools.
Then ChatGPT emerged, and the use of it by some of those "get rich quick" sorts promoting the use of it to make money fast in writing. This piece isn't new, either. Last summer there was much discussion in a Kindle Vella forum I'm part of about how one company was spamming Vella and generating false buys in order to game the Vella bonuses.
The fact that ChatGPT and other writing AI have been trained using the writing of others without express permission is extremely problematic. The degree to which some writers are already using these tools in generative work is extremely unethical and problematic, and I'm going to call out people who use AI for prompts, expanded thesaurus, or "different perspectives" on their work. They're calling upon a tool that has been unethically trained and putting ethical concerns aside, in my opinion. I've also read and heard examples of the alleged "help" from AI drafting.
It homogenizes the voice of the writer. It purges their unique perspective into a mishmash.
Don't get me wrong. I don't fear the AI. It is stronger than human intelligence in crystallized intelligence--knowledge acquired--and in processing speed and memory retrieval. But this is a mechanical superiority over HI, caused by processing power that can grind away at the cognitive wheel. HI is superior to AI in the application of fluid intelligence--that intuitive and deductive leap that creates out-of-the-box solutions to all sorts of problems. I'm currently reading some psychometricians (psychologists who study cognitive assessment measures and devise them) on the subject of AI, and that's the leap that some believe AI cannot take.
Even given that HI is superior to AI in fluid intelligence and will probably remain that way, there is still the ethical implications of using an unethically trained tool. AI will not produce a good work of creative writing soon, if ever. The reports from Clarkesworld editor Neil Clarke and other discussions I'm privy to in a private forum suggest that the AI works are poorly written and easily identified.
But. AI has many potential uses in the creation of instructions, generating certain types of promotional text, and standardizing certain documents. Were I still involved in special education case management, I'd be advocating for its use and training to construct Individual Education Plans, simply because a.) it has the potential to reduce the heavy paperwork load on those individuals and b.) the quality of some of those IEPs is staggeringly awful. It would become possible to change IEP updates from a process taking an hour to several hours to a much quicker, uniform, and more useful document.
HI will still need to be used to check and correct this sort of writing. But that is AI's potential in writing. Not in creative work.
That said, every piece of creative work I am sending out from now on, whether a self-published novel, a Substack post, or short story cover letter is going to contain the following statement:
"No generative AI has been used in the conceptualization, development, or drafting of this work."
It's frustrating the way new tech at the moment seems to be so easily hijacked. Much like how blockchain concepts were shoehorned into NFTs and the various scams and delusions associated with them, now ChatGPT has been immediately co-opted by people trying to make a quick buck from churning out bad writing (which is especially amusing, given how hard it is to make money from writing...).
I use all of it and I’m obsessed. ChatGPT has been so much more useful than Google for researching my essays and novel chapters. Even if I have to fact check, it’s better than clicking on a hundred links that are all SEO content that don’t contain what I’m looking for. And yes I have been loving Midjourney for creating images for my novel chapters!
Elle, this is one way I use it as well, as a replacement for search engines. It's a starting point that refines a vast amount of data into very concise points. I then take it from there and do more research to track down sources when necessary.
I've not really seen how to use ChatGPT as a search engine, as I don't find myself in a position where I can trust its responses - and it always delivers responses with utmost conviction, making it even trickier. Whereas 10 links in a Google search at least have immediate context: it could be an actual academic paper, or a piece of journalism from a respected publication, or it could be an opinion on a random blog, or a crowdsourced Wikipedia article. All with their own merits and issues, but adding up to a broad understanding of a topic.
I don't quite see how ChatGPT replaces any of that, other than through a sort of filtered merging, with an added element of uncertainty.
Conceptually I can understand it as a starting point for further research, but when that starting point is on such flimsy foundations I'd be worried about being pointed down a false path from the very start.
(this is all likely to change of course, as theses systems become more sophisticated and useful)
You have to treat it like you're the professor and it's your student. For example, I formulated the following query: "What are the prevailing theories on how the pyramids were built? Please cite one source for each theory in question." It did exactly that, describing each theory, then at the end providing one source for each. That's not possible with any search engine with any sort of coherency.
I can see these tools being very useful eventually. At the moment, I think they are quite problematic - sharing misinformation, using copyrighted artwork, etc - and require some real oversight. I don't use them and don't really want to, but I accept that people want these tools and they will be used.
Agreed - I think that's largely where I'm at, currently. I'm not against using such tools in the future, especially if they can be developed in more convincingly ethical ways. MidJourney for a time I found really interesting and useful as a concept engine.
The problem, really, for me, was that I slowly slipped into relying on MidJourney to quickly make an image because I'd run out of time - rather than using it as a point of inspiration. I'd imagined using it to help with my own art, but that didn't really happen. The lack of intent behind MidJourney images also is an odd one, in terms of using it as an inspiration source.
I am currently too lazy to learn how to use it. But I am sure it will get easier. I imagine AI will be great for most people until it takes over and eliminates “flawed organic life.” Thankfully, I am already old. 😉🤓
Current status: energetically avoiding
I dislike generative AI as an artist -- the way they've been trained and used to this point has worsened the already ubiquitous problem of art theft and the devaluing of the art people create. Anything that encourages an environment of disrespecting artists on any level is a no-go in my books. The culture of AI is not something I want to be part of.
I wonder if there's a way in the future for an AI to be trained on your own material. If I could train an AI more directly on my work (written or visual), and then have it help me with concepts, that could be an interesting tool. But that doesn't seem viable - they seem to rely on massive datasets, which wouldn't work for individuals.
Or, indeed, whether an AI could be trained using actual tutoring like humans receive when studying art. A more deliberate training, in other words, without the more plagiaristic elements.
Yeah, the way current AI work doesn't jive well with small datasets, and I know there's no way I could produce even 30 images with a consistent look, nevermind thousands to train an AI on; not only would I get tired, but my art is always changing as I develop. The best way to achieve this help right now would probably be good old-fashioned photo-bashing and moodboards! These also have the advantage of being made by you as the person who knows what you want -- no repetitive reiterations to find something vaguely right. The ends have always existed, the means just take a bit more time.
If you wanted an AI to work on a smaller dataset to give you something you could have made, I imagine you'd need to rebuild generative AI from the ground up -- the fundamental concepts of the stuff coming out right now is always going to depend on scraping the Internet for masses of data. As such, although AI generally may be theoretically useful in the future, the versions we have now just aren't something I consider ethical and probably never can be on an inherent level.
It does seem that way. Though I also get that I'm speaking mostly from ignorance, when it comes to the technical aspects of how these things work. :)
Photobashing, though - yes! Since stopping using MidJourney I've reworked the front cover for my book, and went the photobash route for that. Found a bunch of stock images, bashed them all together, and then did hand-drawn (digital) line work and watercolours over the top. I was pleased with the result and a) it was much more unique than anything I'd produced with midJourney last year and b) it FELT much more 'me'.
An acquaintance of mine does this https://imaginationrabbit.substack.com/p/so-i-trained-a-stable-diffusion-model
There is a difference between using it as a tool and using it for everything to not do any work. I can already see that there are plenty of people who are doing exactly that.
I've used writing tools a little bit to "prime the pump" on some writing for my day job. Often they just got me unstuck by showing me exactly what I didn't want to say. Useful, I suppose.
I guess I just find the work they produce uninteresting and they kind of miss the point, at least for me. The point of art, one of the points, is to facilitate a relationship between and among people It's not just the thing, it's the person and their world and the interaction therein that led to the thing that is now in conversation with a different person in a different world.
I suppose sooner or later we will become more machine and machine will become more human and we will relate to one another as equal-self to equal-self, but we're not there yet.
I think that's the crux of it for me, too. AI generation really needs to be done by a properly sentient AI (whatever THAT really means) for it to be of interest to me. Otherwise it's just noise. Art is communication, which is about connecting different minds. If it's not coming from 'a mind' it's lacking something for me.
I used the Dream by WOMBO app for quick image generation for my substack posts, then read negative articles about its effect on artist and theft of styles, so got uneasy and quit. I still use filters that make my photos look like paintings.
I use other forms of AI a fair bit. I have machine learning-driven stuff help my computer run video games faster. Another tool cleans up my audio for Zoom meetings and automatically hides anything that isn't human voice. Those tools aren't pretending to make art, though, and that's where things get a bit weird with the likes of MidJourney and ChatGPT. There's a fundamental shiftiness at their core, where they're pretending to be something they're not.
And if I understand how ChatGPT works, by predicting words based on the analysis of an impossibly vast amount of PAST words, then the doomsayer in me worries that what seems futuristic will paradoxically beholden our grandkids to the past even more than our current crop of nostalgia loving human creators do now!
Part of me gets it. Mashups are cool, who doesn’t want to read Hemingway’s werewolf story!
Also, it seems we’re at a point in literary history where culturally diverse voices and styles finally have as much respect as “the canon” but I worry if we’re not careful, a flood of AI-generated text could wash away those gains and fill our screens with recycled-Frankenstein-type content instead.
Confession: I did ask ChatGPT to tell me a werewolf story in the style of Hemingway and it was neither a good werewolf story nor a good Hemingway homage. Ergo! Because now I really really really want to read Hemingway’s werewolf story, it’s become one of my many works in progress. I just have to reread a bunch of werewolf stories and Hemingway stories to prepare and research. And even if my story isn’t any better than ChatGPT’s - I’m very much looking forward to the research/prep I need to do to attempt it. I won’t let a software program take that away from me :)
I use midjourney-generated images for the main illustration of my substack posts. I see it as a tool to express myself better. Just like a painters brush + paint are tools that make it easier to make a drawing than painting with bare fingers and charcoal.
I also use it to make coloring pages for my kids based on their own desires/prompt ideas.
I’m not really for or against in general. It really depends on the use etc. I do feel that when someone uses a specific artist’s style in the prompts they should mention them as a source/inspiration at least.
The fact that I am a digital artist with sensitivity to all things visual may be a factor in that I 1) didn't like the process of generating AI art, and 2) don't find it appealing to my eyeballs.
As for the text generating AI tools; it's only been 3 weeks since something prompted me to experiment with ChatGPT and I use it on some occasions.
For some things, it's faster to get what I'm looking for than by using Google. And I also like to take it as a 'brainstorming pal'.
But I do not like and can't personally use the text it produces - it's not my voice and I have a strong repulsive feeling in my body to use those formulations as my expression.
A hot topic for sure. I dabbled a bit with Nightcafe and stopped because the output is too generic and the anatomy and architectural perspectives are always messed up unless you use starting images. I did do some "evolving" of existing real images to make them look futuristic. The results were mostly goofy but sometimes excellent, and that's where people get anxious, and rightly so. A new artist putting up his oil paintings at a small local art gallery can find his work evolved online the next day, thanks to phone cameras and AI… not cool.
And ChatGPT? For writers, Dramatron might be more interesting. Although, ChatGPT has its uses as a tool to get information which you can also get via Google. I let it finish some text once, and it was able to come up with something generic. Then again, looking at the media landscape, would you always be able to tell which script has been written by ChatGPT?
I tried to solve a problem coding a health bar with hearts for my Interactive Fiction Demo in Twine with ChatGPT. It proposed many solutions, none of which worked, and every line of code it produced did not follow the correct syntax. I kept training it and made sure it had the latest release versions and story formats, to no avail. Some solutions looked cool and elegant, but the code was rubbish, and I ended up losing three hours and then did it myself in 10 minutes. You can check out the result on my substack (first post).
In the end, ChatGPT, Bard, etc., are tools that are not going away, they are only going to get better and if you wonder what's out there right now? Well, theresanaiforthat dot com.
Yeah, you're absolutely right that AI generation isn't going away. It's going to be fascinating to see it develop, and also observe where it is adopted (and where it isn't). It's all coming in so hot at the moment that it's impossible to really tell where it's headed.
AI is already everywhere, ad campaigns, copywriting, movie editing, scriptwriting (dramatron), research, you name it, it's there. Not using it will be a conscious choice you may be able to afford, but companies will look at the $ and decide they don't need testers or designers anymore. Teachers? AI. Doctors? AI. Presidents? AI... my initials... ;)
I think it depends a little on the application and tech, as not all AI/machine learning stuff is the same. The generative AI we've seen in MidJourney and ChatGPT relies on the huge data sets, which is where the ethical problems come in. AI-enhanced tools that have been in Photoshop, or like DLSS or noise cancelling stuff from NVIDIA, I see as being slightly different and far less problematic.
I can see it birfurcating sooner rather than later: strands of society and business that go heavily down the AI route, where efficiency and process matters most. And then another strand where creativity and human connection matters most, where AI will play a more minor part. I'm not saying that the latter is 'better', incidentally - just different.
The question will be "is this thing supposed to form a connection of some sort between people, or is it supposed to get a job done?" And that will determine much of the application.
It's getting stuff done (?) for now. For connection forming, this "thing" is not advanced enough, yet. Although Japan has invested in automating elder care for over two decades now (see MIT Tech review), they haven't made any significant progress. And if there is an Elder Care Bot waiting for us, it's not a comforting thought...
The next 50 years will be wild!
Ah. Forgot to comment here.
A word of definition - I'll assume, Simon, you're specifically referring to GENERATIVE AI tools - those that seek to "create" with minimal input from a user. I make this distinction as, among other tools, I'm typing this on a smartphone using SwiftKey - a keyboard app which can be set to read one's emails and social media accounts to learn how I wrote and use AI to improve its autocorrect and predictions.
Thus, yes, I do use multiple AI based/enhanced tools, from SwiftKey to On1 PhotoRaw to Vegas Pro and Topaz Video and beyond.
Here we get into different "classes" of AI tools. On1 PhotoRaw uses AI based tools extensively - especially with version 2023. However most of the tools are based around "smart masking." Let's say I have a photograph of a person standing on a beach, looking out across a lake. On the far side of the lake are hills with some buildings. PhotoRaw now uses "AI smart masking." If I select the "AI superselect" tool it will attempt to isolate elements as I mouse over them - the person, sky, foreground beach, background hills, buildings, water; even the foliage. Once an element is isolated I can either left click then move my mouse over another element to add to the selection, or I can right click and assign effects or filters to that element. In practice, this means if I want to tweak the color balance of the trees, I mouse over the trees, right click and add the color balance effect. However, this is basically speeding up the tedious process of masking (note also the AI masks aren't always perfect, but I have manual masking tools to clean up). The software isn't making all the decisions for me.
Ok, fine, there's an AI exposure tool, but I rarely use it because the AI doesn't adjust images to my personal taste. I may use AI exposure in culling to get a better idea of what an image can look like, but I reset and redo shots I'm keeping.
But PhotoRaw even now has "AI presets." Click a preset and it does X to the water, Y to the land, Z to the sky, A to the person, B to the buildings and C to the foliage. Pretty sweet. Of course it's applying masked, non destructive filters to all the elements, and I can go in and tweak everything. Still, I LOVE these. Now I can work on a single image and really quickly stick the same looks on everything else in the set without tedious masking or copy/pasting filter settings. Thanks, AI tools! You've removed most of the boring bits and let me get onto the fun bits!
That's the type of AI tool I like. It's doing the annoying tedious bits for me - pretty well, usually 99% there, sometimes only 80% - and letting me get onto my creative work.
Generative? Not for me. I'm not the best artist, photographer, editor, anything-else-I-do by far, but I worked to develop what skills and techniques I have. With a Generative thing I don't feel I have control - because I don't. I also get bored because I'm sitting waiting for the random AI thing to spit out, and, more often than not, it isn't giving me what I want.
Simon, I like to think our discussion on WhatsApp two days before you published this influenced the article. For any not-Simon person who may read this, the discussion was on how certain of my friends were admiring some YouTube channels that do quick, boring slideshows of AI generated content. Think "Star Wars, but it's 1950's."
Not only are the images generative AI, but the videos are straight static slideshows without the editor even bothering to "Ken Burns" pan/zoom or even crossfade between slides. They're straight static images with no rhyme, reason or attempt to order the images into a narrative flow. They are likely the images taken in the random order generated, dumped on a timeline and rendered out and it's so boring and utterly fucking lazy that I'm actually offended this utter shit is garnering 10's of thousands of views.
So to prove a point to these friends I allocated one hour to generate as mamy "Avengers, but it's 1920's" images as I could. To take a step of effort beyond what these YouTube channels did (where their "Princess Leia in the 1920's" images all had different faces) I even specified actors! Errol Flynn as Tony Stark, Lou Costello as Hulk, Eddie Anderson (best known as "Rochester" from Jack Benny's shows) as T'challa. Among other things the limits of the AI model's training became clear. Let's just say the AI didn't know who any of my non-white actors were, and Stable Diffusion has only one black male face - high cheekbones and a goatee. The racist training of the AI is a different topic...
But, to finally get to my goddamned point after paragraphs of setup, I moved on to "Avengers, but it's 1950's. Here we ran into issues. I gave myself an hour to generate prompts. 45 minutes were wasted, because no matter which prompt I used the AI was utterly unable to ever understand that "Peter Lorre as Hulk" means "Give me a shirtless, green muscular man with Peter Lorre's head." (Yes, I tried that as a prompt.) Basically I ended up both bored after an hour of watching a scrolling Discord thread, AND annoyed I didn't get the image I wanted and PISSED because I could have done it myself in Affinity Photo in a couple of minutes.
I don't like the AI tools because they aren't very good. I'm better. But I'm slower. I've already lost work to generative AI tools. A friend working for a major game studio is fighting to keep 50 jobs from being replaced with an intern and a generative AI. I wouldn't give a rat's arse if people used the tools, but the bottom line is they are going to cost jobs (hey programmers, ChatGPT can code!)! Which only matters because we remain on an antiquated, millenia old model across most societies of tokens (money) earned through labor for food and shelter, and AI art tools, AI self driving cars, robotic factories, et al remove jobs without creating new jobs to replace them.
These disruptive technologies have appeared for centuries - sewing machines took work from seamstresses, automated factories took work from laborers, synthesizers took work from musicians (don't argue with this, the average Broadway or West End orchestra was 50-100 musicians before 1980, now it's 15-30. The synthesizers and backing tracks took over half the jobs). The AI revolution will take more.
And that's a real problem when we're on the "work to eat" kick.
Yup this comment just took a turn into advocating a "Universal Basic Income" brand of socialism!
Wasn't done yet.
Anyways, we've also discussed, ad nauseum, the ethical connotations of the training models, via scraping of copyrighted work.
I'm thinking of recent Corridor Digital experiments where they got it right and wrong at the same time.
They trained a Stable Diffusion model with their own faces so that SD could consistently create the Corridor Crew in other images. This is kind of "right" because the AI was trained on custom images they took.
Then they used the AI to turn live action into anime, but they trained a custom data set they scraped from a specific anime - and justified every fear the working artist has. They specifically, willfully and with premeditation took copyrighted worked and trained an AI to emulate it to do themselves in the style of the AI.
I don't THINK Corridor meant to offend, but they haven't fully thought through this. Man, they're literally getting worldwide bad press for the project. I think they'd rather get positive press for things like Wren's amazing scale videos. Niko's comments about AI "democratizing art" don't help.
For one thing, the hardware requirements for custom training an AI model are staggering-they are using $8k+ machines-and the other option is buying runtime on a Google cloud. Most people use Google Colab, thus does the trillion-dollar conglomerate make money.
Side note: OpenAI, maker of ChatGPT, is an Elon Musk company. He's Co-founder of the company behind one of AI art tools... I think Dall-E. Using these tools supports evil. Just saying.
Digression over - "democratizing art?" Call me elitist, but, NO! Develop some skill! Practice! Learn! If you aren't any good, maybe you shouldn't do it! Typing a prompt into an AI and taking the image or text it spits out doesn't make you an artist or writer. Practice might.
AI leads to what I call "Next Gen Syndrome," because I figured this out as a teenager watching Start Trek: The Next Generation. The rule is: "Overreliance on computerized tools makes you stupid and lazy." Because the Next-gen crew are basically idiots most of the time. There's the episode where the Enterprise contracts a computer virus and nobody thinks to reboot the computer and re-install from the clean backup. At all. Until Data "dies" of the virus then suddenly sits up and announces his brain has rebooted off the clean backup. There's the episode where these crystal node things are growing on Enterprise's systems and the holodeck becomes Enterprise's "subconscious," and our heroes walk up to the holoactor putting together a jigsaw puzzle showing Enterprise overlaid with a giant crystal, and the first officer, counseler and chief engineer reveal they don't know what a fucking progress bar is because none of them have any idea of what this puzzle could POSSIBLY represent. They are idiots because they let them computers do everything.
I see it in people today - and that includes myself. I used to be the waking phone book. I knew all my friend's numbers, work numbers, locks business numbers. People used to call me to get everyone's numbers cuz I had then memorized. Now, of course, my phone has a phone book and I haven't bothered to memorize numbers. Relying on the tech made me stupid.
As a new generation of "artists" starts using more generative AI their work will become bland, boring and stupid. We can't stop it. We're already seeing it on YouTube (chasing the algorithm) or pop music (an analysis of music from the 1920's to the mid 2010's objectively proved modern music is less complex in tune, tempo and structure than music before 1990). We're seeing it in TV shows and movies planned by committee. Generative AI is only going to accelerate the homogeneity of art into boring blandness.
Yeah, Corridor made a big mis-step with that anime video. Felt very queasy watching it, and the flimsy claims of 'democratisation'. And the end result wasn't great, either.
I haven’t used or noticed AI in writing-related activities, although I did read what you wrote about deciding not to use it.
Generally I’m afraid of AI and really don’t want to see humans become obsolete, so I won’t be using it for anything.
The hardest thing to do as a creative is to start with a blank canvas Ai allows you not to do that.
I feel writers and artist should treat Ai the same way we treated our vision boards or inspiration Pinterest.
You use it as a way to model what you are after but you still need to do the work yourself
That's definitely what I liked about MidJourney, at least for a while. Over time I felt I was being sucked into that homogeneity, rather than it actually enhancing my art - but that is more to do with me, I think. I've ended up in a place - for now at least - where I'd rather manually gather stock images and inspiration and compile vision boards. Takes longer, but feels more deliberate and purposeful.
I have started to use it more for software programming and I can see the value immediately. While associate programmers with only a few years' experience may not find it as useful, because you don't know what you don't know, I can very precisely narrow down a problem domain and have a polished solution with little editing. The code is production ready in minutes vs. hours. In the past I would have been concerned about this progress, but it's a large leap from that to AI is going to steal my job.
One thing I mentioned in my recent post about this, is simply the disruption it's going to cause in traditional publishing. Con artists will find ways to manipulate systems using AI as their slave assistant. They will overwhelm the little guys who can't keep up with the influx of submissions and attempts to make a quick buck. I'm trying to recognize the trend and make adjustments to my strategies as a writer.
That's interesting, I've read far less about its applications for coding.
I listened to Ed Miliband's podcast today in which he was interviewing Bernie Sanders, who briefly touched upon AI and the risk to jobs. Sanders' point was that AI tools aren't a problem, and tech is never the problem - it's the way they're implemented and introduced into society. If AI tools can be used to help workers be more productive in the same amount of time? Great! If they're used to reduce the working week to 2 or 3 days while maintaining wages? Great!
Of course, that's rarely how it goes.... (and this is where we circle back to the problem with billionaires and megacorps being the ones investing and inventing this stuff)