Can computers write novels?

In My Humble Opinion
1

In this thread, the note is casually dropped that computers can currently write novels. This seems like utter bullshit to me. Hell, I would be amazed if computers can currently write an original metaphor, let alone develop a whole plot, something which depends upon human motivations.

I recall reading a story of a group that convinced a television news program to report on a computer program they had developed that could write novels. I believe it was in the style of Agatha Christie. The whole point was that it was impossible, that science reporting was overly credulous and scientifically illiterate. I doubt we have come far enough that any of these points are no longer true.

I could not find that story, but I found two others. The first is similar to what I recall, a novel (“Just This Once”) written in the style of a mystery writer, but it is in the style of Jacqueline Susann. This was from 1993. There is no payoff of it having been a fake, though. In fact, the author apparently sued the programmer for royalties.

Then, there is a Russian novel, “True Love”, written in 2008. It is a reworking of the ideas of Anna Karenina, written in the style of a Japanese author.

None of the articles I have been able to find on any of these works say how much the humans are doing and what the computer is doing. Sure, if the humans come up with the whole plot and what happens in every scene, and then write that into a code the computer can read, they could have a computer that generates text similar to a particular author. That would then need to be cleaned up to correct the constant and terrible errors. At that point the machine is a very advanced thesaurus.

Nothing I can find actually goes into what the process is like though.

So, can computers write novels? Do any of you have any information on how the ones supposedly written by computer were actually produced?

2

Based on this article, novels have been created by the following methods:

So it looks like most computer generated novels are just pulling text from human-written sources and throwing it at the wall, resulting in books that don’t make sense.

3

I remember reading computer-generated short stories in the early 1970’s, when the artificial intelligence programs were still at their most primitive. The stories were funny just because they were so silly. I doubt that any can be found on-line today, but I take a quick look.

ETA: Nope. Some quick google searches with various likely-sounding keywords. All I get is cites for more modern CG stuff and especially CGI stuff.

4

There’s a big difference between yes, technically this is a novel of sorts and an actual novel that would get published/anyone would want to read. The “novel” should be judged on its own merits, not for the novelty of its origins.

5

Maybe Kurzweil will write a novel after he downloads his brain into a computer.

6

Plotting is pretty mechanical. If you went through TV Tropes and picked out the various character types, genres, etc. and layered in other options for setting, happy/sad ending, etc. you could have something that could make something that would generate whole storylines, of arbitrary length.

Making all of that interesting is the part that takes some effort.

7

I could see computers writing screenplays for the SiFi channel, but that’s about it.

8

Don’t be insulting. Some of my best friends are computers. They’d core dump at the very thought of sharks in tornadoes.

9

It is kind of like the computer-written technical papers. They only get into scam conferences which never bother to look at the submissions. You don’t need a PhD in the field to know these things are fakes.

10

Computers can write stories and vignettes. I didn’t mean to imply they’re any good, or would consistently pass a Turing test. Generally the longer the story, the worse they are because they eventually slip into word salad – like a chat bot. Scheherazade is the best I’ve seen. You can read about it here along with some other notes about computer generated metaphors. Here’s a sample of prose:

Don’t expect any deep insights into the human condition, unless it’s an accident. My main point in the original thread was if you combine these sorts of statistical analysis with an actual AI (instead of dumb, rote algorithms) then humans are cooked.

For other computer generated material with some semblance of thought behind them, there’s the [I like my relationships like I like my source, open]“I like my X like I like my Y, Z” generator (PDF),](http://homepages.inf.ed.ac.uk/s0894589/petrovic13unsupervised.pdf) which does actual analysis on relationships between words to craft the joke. Humans found the results funny 16% of the time, which is probably better than my hitrate.

Some samples:

Here are some other computer generated jokes.

Not exactly Bill Burr, but it might make a kid laugh (or an adult groan, which is a sort of win for a pun).

A site you can end up wasting a lot of time on is the Bot Or Not? poetry page. It’s usually easy to spot the bot, but sometimes you get snookered…and it’s especially disconcerting when you think a person is a bot. Check out the leaderboards. Here’s a sample of bot poetry that was voted as being particularly human. The very end cracks me up:

So true, so true!

11

Since a novel is defined as a work of fiction of a particular word length (40,000 words for Hugo and Nebula ballots; 70-90,000 for a commercial publisher), I’m sure a computer could write one. The question is whether it’s a coherent narrative.

As Fred Pohl said, a moderately fast typist (50 wpm) can easily write the 70,000 words needed for a novel in a week. The problem is getting them in the right order. And that’s exactly the problem with a computer generated novel.

If there were a formula for novel writing, everyone would be writing successful ones. But there is no formula (“There are nine and sixty ways of constructing tribal lays/And every single one of them is right!”) and also so many factors you have to think about – basically, with every sentence you write, you need to consider dozens of factors (What does this say about the character? How does this advance the plot? Is this something the character would say/do? Does this make something I’m planning to write five chapters later more difficult to pull off?) that you need to get fixed in the editing process. There is no simple analysis, and I doubt anything other than a truly self-aware computer could manage it.

12

Quote:
I like my relationships like I like my source, open.

I like my boys like I like my sectors, bad.

I like my women like I like my gas, natural.
These are great.
I guess I’m probably a computer.

13

I’ve often whished someone would use techniques like to this to create a movie script; then produce said move exactly as the script is written (no cheating with a script doctor) just to see how fucked up the end result would be. :smiley: Am I alone?

14

I’m a programmer and a writer. I wasn’t joking. You could dynamically generate plots that would be indistinguishable from something a human could make. I used to generate roleplay scenarios by rolling a dice, over a table of options.

15

Thanks for your replies everyone. The results are exactly was I suspected would be true, but very different from the articles I have been reading. Sage Rat, I can see now that writing a plot outline would be fairly easy. The trouble is incorporating that with text that flows, moves the plot forward, and reveals more about the characters.

16

There was a SF short story that I think I saw in John Campbell’s magazine back in the day, which posits an SF writer in a disagreement with his editor over changes to his latest work. The writer drags the editor to see “Rollo,” an experiment done by a professor friend of his. Rollo, a monkey, has parts of his brain wired to a computer, which uses the computing power of the monkey brain to analyze writing. (A takeoff on the “infinite monkeys on infinite typewriters” thought experiment.)

The writer contends his work, as submitted, is the best way to tell the story; the editor wants a scene change in the middle. The story is read to Rollo, who picks up the cadence, rhythm, and ‘voice’ of the narrative. At the point of the story in contention, they stop and let Rollo take over. He continues for a few lines, then introduces a scene change.

The editor is jubilant, saying this proves he was correct. But the writer says that this is exactly what he wanted to prove: that a soulless, calculating machine approach does indeed put a scene change in. But doing it that way, he argues, means that the editor’s job is useless, that any machine can calculate how to edit. It takes real human insight, compassion, and depth to know when to break the rules. He points out that if the editor disagrees, he’ll basically be admitting his job can be done by a souped-up monkey. Shaken, the editor agrees to leave the story as is, and leaves in search of a stiff drink.

The professor is impressed with the writer’s arguments, but asks, “Tell me, what would you have done if little Rollo typed it your way?”

Aggrieved, the writer admits, “Damn it, professor, that’s what I thought he was going to do.”

17

I recently read about a company that says they can write marketing copy with software. Related, but very different, I admit.

18

Now that I would totally believe.

19

You could probably do it, yeah. I would think you’d need a pretty solid natural language engine and then a program that could use some sort of fractal to create a story, ie start with a beginning and pick an ending. With just a start and an end, it’s pretty linear, so your fractal could split it in half and refine each half with something else interesting. You’d do that again for each half. Instead of changing locations of dots on a line you’d be modifying plot points, such that each iteration involved smaller and small changes to the overall plot. When you’ve worked down to small instantaneous changes in plot it would be easier to calculate a series of events that would allow the story to progress from one to the next, like speech or interactions. You’d probably want to run over the whole things a few times after to smooth it out, but it could probably work. If each plot point was assigned a “type” ie what sort of thing happens, events could be correlated together through the story. For example, a plot point could be classified as “type_A interaction” and you could assign characters or locations to type_A. Any where a type_A interaction occurred would be related to other occurrences, so that you didn’t just have a series of random events. I think in essence the secret is generalizing the relationships that people have to each other, places, objects and events and making sure you keep track of them so that the story feels organic enough. You’d probably end up with a really wonky story as it would be mostly RNG driven but it would at least be relatively consistent.

20

You left out the dino and the laser gun.