Alright, last post we talked about changes we’ve seen in writing software as well as machine translation. This post is going to discuss the big changes which have, in many cases, triggered this initial conversation.

Audiobook AI Voices

Machine reading of books has been around for a while. I know a number of people who tell software like Alexa or SIRI to read their ebooks for them, skipping the entire process of live readings.

There are issues with software reading like that, mostly in terms of inflection, occasional mispronunciations of uncommon words and, of course, for fantasy the mess that is made up words and names.

For some readers, this is a non-issue and they are willing to accept the trade-offin return for cheap and convenient audio consumption of their favorite works.

What is new, and driven by new companies, is the use of AI voices to create ‘real’ audiobooks. A number of companies have set-up to provide such work, many of them with horrendous terms of use (seriously, I’ve seen one where using their software gives them the copyright of the audiobook, another where you have to pay an on-going licensing fee, and othrs that just prohibit commercial use) but the big change has been Google Play’s introduction of ‘free’ audiobook narration for works uploaded.

Now, you might be wondering how they handle things like fantasy words or the like – and in most cases, it’s a secondary, human check to verify pronunciations or an initial list of words and names and correct pronunciations.

Basically, still requiring a human proofreader. The Google Audiobook system currently runs on a ‘free’ basis with you receiving a 50% cut from sales for these audiobooks but you’re allowed to download and put them up wherever you wish.

On the other hand, places like ACX (Audible) do not allow AI voices (yet!) though some companies, like FindAway have made indications they’re trying to push ahead with them anyway.

Quality wise, it’s just not there yet. I understand in some cases, you can go into the system and help them get better, by choosing the emotional range and stresses for individual sentences and the like; but it’s still not as good as a real narrator. However, there’s an argument to be made that with an AI voice and an author working directly, you might be able to create a more ‘authentic’ experience.  

As others have said, this is coming down the road and is likely going to become a major part of the industry. I’m not a big fan of the idea, I like my narrators, but it is coming. More thoughts in previous blog posts…

But this is just one of the major changes. And not the biggest one since….

AI Created Content (Images)

The huge and overwhelming attention has been to AI art recently. Midjourney has revealed a series of really cool artwork pieces. Someone even won an art award for an AI generated art, after spending countless hours refining his prompts. 

There’s a lot of discussion about the problem with AI art right now, and what it can/cannot do. For example, there’s a fascinating thread about how it is useless (for the most part) for comics, where a continual series of panels are required to tell a story (again, mostly). AI art makes it hard to regenerate the same images, over and over again, or to keep track of a single figure moving across the same feature, but with minor changes.

Though, I wouldn’t be surprised if eventually that could be trained into the AI art systems. It would need a lot more examples of such sequential imaging (possibly ripping apart movies and TV shows to get that context); but it could get there.

And it does some fascinating things at creating weird monsters because it throws up random inputs at one another to form some out there dimensions.

But, right now, the biggest concern for AI art (among publishers) is whether it can be used for covers, which often require a single static image and the increasing volume of such art in stock image sites (potentially causing problems in sourcing). 

It can potentially make for really cheap cover art, since you could get some high quality images for very little cost. However, we come to the human cost of artists being locked out of another way to generate income.

Just as importantly, much of the (current) AI images are very similar in ‘feel’, so it could cause easy customer confusion. Which might work for certain authors but will be a problem for the bestsellers.

The biggest impediment right now, for widespread adoption (beyond social stigma) is the lack of clarity in copyright. Right now, without significant human intervention, in the US, AI generated content is not copyrightable. However, copyright is possible and is owned by the programmers in the UK. Depending on the licenses you get, the way the court cases play out, this could get messy very soon. 

Of course, some artists have embraced the idea of AI art; using them as jumping off points for their own work. Creating their own collages, painting or altering such images to make it work for them. Much like an artist tearing apart magazine pages to make their own art, such work is copyrightable and can be beneficial and see an explosion of creativity.

I’m not sure how I feel about AI art (in general). I’m mostly of the mind that I’d rather use real artists, because I can then get something truly unique with a known quantity (their style). 

On the other hand, I use stock art covers (created by humans); but is it that much different? 

And how will the big traditional publishing companies act? When names in the fantasy writing community are already receiving minimally altered stock art covers, will we see trad pub embracing this technology to provide cheap but good art for their people? 

When will we see the first big 5 AI generated cover I wonder? Because I doubt trad pub companies are going to hold back long.

AI Generated Content (Writing)

This has been pushed a lot and some authors use the AI generated content to create prompts for them to work from. Some authors find it works well for them, giving them jumping off points to finish scenes or fill in descriptions they don’t care to write.

However, the biggest concern is writing full novels. Whilst blog posts are easy enough to generate (and really, it’s just a regurgitation of other work!); such AI generated content cannot write a full book. I’d ben surprised if it could generate a coherent short story anytime soon.

There are significant hurdles, from plot consistency to character arcs to overall plot structures to be combatted. It might be a very long time (and the addition of multiple types of programmed instructions) before software can generate consistently ‘good’ novels.

And then again, we run into the question of author voice. Where a certain person’s view of the world, their real world experiences come shining through and we get a glimpse into another mind, another world that we don’t see. 

AI might be able to ape that, but it won’t be authentic. And it certainly won’t be contemporary (see below for why).

Feeding the Beast

So, one of the things about the way the current generation of machine learning works is that it uses work that has been written/created already. It doesn’t create so much as modify existing work, deriving work into the future.

And yes, humans do that too – with art, with writing, with our decision making – but our series of inputs I would say is likely larger than most AI and more unique. Which leads to the problem of current AI being unable to create ‘current’ or ‘original’ reporting information, instead having to regurgitate alterations of old information.

It also means that, depending on the data set (and programmers are notoriously bad at choosing data sets because they’re not TRAINED to figure out how to pull the right datasets but are trained to program unlike data scientists) the work pulled can just be wrong when you move away from ‘mainstream’ art types.

Thus you get art from Midjourney all looking the same, as if created by a single artist (because it sort of is). Thus you get false positives and that final 20% that makes a work go from decent to great just being missed. 

It also means that, say, if you want brush art work from traditional Chinese paintings,   these AI systems will not be able to replicate them well. The data set being drawn from is just too wide or not wide enough to create such work well. Again, depending on data sets or the kind of information parsing it uses.

Edge cases, work that is not mainstream, is where such data fails.

It is also why places like Google wants you to use their software (for basically free); because that gives them even more data than ever to make use of.  Though how that data is being labeled… if it’s being labeled corrected… who knows.

In short? 

We are likely to see AI artwork make an appearance, in covers, in ebooks as scene setters, etc. We are going to see some pain among artists and voice actors. There’s going to be a few legal cases where people are caught out between the changing laws between each country.

For unique or different artwork, for people with unique styles; they’re likely safer for the time being. This increased competition might see a further fracturing of styles… 

As for us writers? Actual novel writing software and AI writers are a long way away. Decades I would assume, though some derivative work that can be edited to something vaguely passable might appear sooner.

This is getting long. I’ll tackle the next two (which are even bigger topics) in another blog post.

Like the business blog post? Want to support me writing more of them? Want to read ahead (2 weeks) of others? Become a Patron and choose the $2-tier to be able to read the business posts only and ask questions about the business side of writing.