An Eye on AI
AI - and if you've been hiding under a rock for the past few years (I wouldn't blame you) - that's artificial intelligence, has become A Big Thing, no longer restricted to computing labs or behind the scenes with Big Tech, but actually available to us members of the hoi poloi. There is understandably a lot of concern about the potential for AI to replace huge swathes of working humanity, leaving them to scratch around for a living while big companies offload services to Human Intellect 2.0, and also a lot of noise has been made about the potential for copyright infringment during training of the AI 'models'. I can't comment on the workings of the tech itself and as anyone who's ever looked into copyright infringement knows, it's a minefield of interpretation, much more art than science. Anyway, a month ago and with a book almost in the bag, I decided to have a look for myself, and in this post I'm going to report what I think I've found out. AI is always evolving, much like we its creators, and my experience is limited to Chat GPT (other AIs are available) so my thoughts may well change. But as of October 2023, here's what I found.
I'm now much less concerned about copyright issues than I was, at least in the field of text-based AI generation. A major issue remains, which is whether training an AI constitutes storing copyright material electronically; that's obviously a big no-no, but do the algorithms actually store the data itself so that it could be reverse engineered, or an 'impression' of the data, filtered through an aggregate of existing 'memories'? I can't answer that, but the output of Chat GPT when asked to generate text is cliched, clunky, and internally inconsistent. A book written solely using Chat GPT responses to user 'prompts' (i.e. the questions you ask or tasks you give it) is going to need a lot of editing, even if you spend a lot of time crafting those prompts. After a month's use refining my prompts, Chat GPT still generates text that I would have been embarrassed to have produced as a 15 year old (which is saying something). It's shlocky B-movie hyperbole, the worst Hollywood has to offer. Most intriguingly, Chat GPT can't 'remember' what it's already done sometimes in the same pararaph: a lonely traveller might suddenly be accompanied, bested enemies might reappear. It also can't reproduce a book, even if it's obviously been trained with it. I asked Chat GPT to complete the sentence "In a hole in the ground there lived a..." and twice it summarised either the whole of The Hobbit or Bilbo's lifestyle, once it created something utterly new, but it didn't start spewing out Tolkien's actual text. You can find summaries of The Hobbit online already, so there is no direct copyright infringement in the sense that the AI will give you a free copy of the book.
Where I have found Chat GPT to be genuinely useful is in two areas: i) complex research questions, and ii) plotting. I'll deal with these in turn.
Research is something I really enjoy, so it's a real rabbit hole for me. Even so, despite professional training as a researcher and a lifetime's fossicking in the backwaters of Triviana, answering some questions just takes a very long time because nobody has addressed those questions directly. You might read a dozen books, scraping fragments out of each one, and then have to peer at the results for months from different angles, reading even more books on peripheral issues, to get what you want. Chat GPT is pretty good at answering even complex questions, but its database only goes up to Sep 2021, and the free version is not on the net, so it misses things. It also has that persistent problem with consistency. I asked it which metals were known in Antiquity, a question to which I know the answer, and while it was correct it gave me a list of ten metals, a very round number which I had not asked for, listing tin, lead, and copper more than once to make up to 10. Sometimes it just gets the answer wrong, or misunderstands even a very simple prompt. Chat GPT also seems keen not to offend anyone, so it steers clear of controversy, especially in religion. It's designed to be Woke-proof, with all the negatives that entails for objectivity. One great thing is that you can interrogate it further about an answer, so it's more than just a summary of Wikipedia. So B+ on that, or even A- at a pinch.
For plotting I've asked Chat GPT to come up with plots in different settings. It does a pretty good job, though it's very Hollywood in its approach to things: there is a utopian bent, happy endings, personal growth, teamwork with a band of like-minded individuals who don't hate each other, and more specifically for anything sf or fantasy related: a massive focus on McGuffins like mysterious artifacts. It's still been useful for me though, because it acts as a sounding board for ideas, and offers alternative views of the same set-up. Both of those capabilities are drivers of creativity, and several times I've come away thinking in new directions and building on my own ideas in new ways. Even writing the prompts is useful because it forces you to think in very structural terms about a plot, something that can be hard to do if you're 'in character'. Finally, I shy away from the Hollywood hyperbole, and so it's been useful to see how some of my 'tame' and non-epic ideas can be viewed in that light. Definitely an A here.
Final point: Amazon now wants indies to say whether a work is AI generated (AI made the text, you edited it) or AI assisted (er... anything else?). As I've suggested there, I think the latter term is pretty useless, and as AIs get incorporated more and more into e.g. search engines, it will be impossible to say that any work is not AI assisted in some way: is it even possible to do that now? Only if you never search the net for information.
All in all I've been reassured by AI's limitations at text generation from the perspective of copyright infringement and writers having no future, and enthused by its ability to drive creative thinking. As I said at the beginning, things are changing fast, but right now I think AI is a useful tool. One danger: asking it to do stuff gives you (or me...) a false sense of productivity and creativity. I've definitely found that the best stuff has come out of my own head after using Chat GPT, in thinking over the output and my approach to the input. We're not obsolete just yet...
I'm now much less concerned about copyright issues than I was, at least in the field of text-based AI generation. A major issue remains, which is whether training an AI constitutes storing copyright material electronically; that's obviously a big no-no, but do the algorithms actually store the data itself so that it could be reverse engineered, or an 'impression' of the data, filtered through an aggregate of existing 'memories'? I can't answer that, but the output of Chat GPT when asked to generate text is cliched, clunky, and internally inconsistent. A book written solely using Chat GPT responses to user 'prompts' (i.e. the questions you ask or tasks you give it) is going to need a lot of editing, even if you spend a lot of time crafting those prompts. After a month's use refining my prompts, Chat GPT still generates text that I would have been embarrassed to have produced as a 15 year old (which is saying something). It's shlocky B-movie hyperbole, the worst Hollywood has to offer. Most intriguingly, Chat GPT can't 'remember' what it's already done sometimes in the same pararaph: a lonely traveller might suddenly be accompanied, bested enemies might reappear. It also can't reproduce a book, even if it's obviously been trained with it. I asked Chat GPT to complete the sentence "In a hole in the ground there lived a..." and twice it summarised either the whole of The Hobbit or Bilbo's lifestyle, once it created something utterly new, but it didn't start spewing out Tolkien's actual text. You can find summaries of The Hobbit online already, so there is no direct copyright infringement in the sense that the AI will give you a free copy of the book.
Where I have found Chat GPT to be genuinely useful is in two areas: i) complex research questions, and ii) plotting. I'll deal with these in turn.
Research is something I really enjoy, so it's a real rabbit hole for me. Even so, despite professional training as a researcher and a lifetime's fossicking in the backwaters of Triviana, answering some questions just takes a very long time because nobody has addressed those questions directly. You might read a dozen books, scraping fragments out of each one, and then have to peer at the results for months from different angles, reading even more books on peripheral issues, to get what you want. Chat GPT is pretty good at answering even complex questions, but its database only goes up to Sep 2021, and the free version is not on the net, so it misses things. It also has that persistent problem with consistency. I asked it which metals were known in Antiquity, a question to which I know the answer, and while it was correct it gave me a list of ten metals, a very round number which I had not asked for, listing tin, lead, and copper more than once to make up to 10. Sometimes it just gets the answer wrong, or misunderstands even a very simple prompt. Chat GPT also seems keen not to offend anyone, so it steers clear of controversy, especially in religion. It's designed to be Woke-proof, with all the negatives that entails for objectivity. One great thing is that you can interrogate it further about an answer, so it's more than just a summary of Wikipedia. So B+ on that, or even A- at a pinch.
For plotting I've asked Chat GPT to come up with plots in different settings. It does a pretty good job, though it's very Hollywood in its approach to things: there is a utopian bent, happy endings, personal growth, teamwork with a band of like-minded individuals who don't hate each other, and more specifically for anything sf or fantasy related: a massive focus on McGuffins like mysterious artifacts. It's still been useful for me though, because it acts as a sounding board for ideas, and offers alternative views of the same set-up. Both of those capabilities are drivers of creativity, and several times I've come away thinking in new directions and building on my own ideas in new ways. Even writing the prompts is useful because it forces you to think in very structural terms about a plot, something that can be hard to do if you're 'in character'. Finally, I shy away from the Hollywood hyperbole, and so it's been useful to see how some of my 'tame' and non-epic ideas can be viewed in that light. Definitely an A here.
Final point: Amazon now wants indies to say whether a work is AI generated (AI made the text, you edited it) or AI assisted (er... anything else?). As I've suggested there, I think the latter term is pretty useless, and as AIs get incorporated more and more into e.g. search engines, it will be impossible to say that any work is not AI assisted in some way: is it even possible to do that now? Only if you never search the net for information.
All in all I've been reassured by AI's limitations at text generation from the perspective of copyright infringement and writers having no future, and enthused by its ability to drive creative thinking. As I said at the beginning, things are changing fast, but right now I think AI is a useful tool. One danger: asking it to do stuff gives you (or me...) a false sense of productivity and creativity. I've definitely found that the best stuff has come out of my own head after using Chat GPT, in thinking over the output and my approach to the input. We're not obsolete just yet...
No comments have been added yet.
