• er under ombygning - vi er tilbage mandag med nyt udseende. Henover weekenden er alt vores indhold åbent, men man kan ikke logge ind og debattere.

Using AI to overcome writer’s block: “You can write a novel together with Dostoevsky”

Martin Pichlmair and Charlene Putney got the idea for their creative AI-powered writing tool LAIKA when they started diving into artificial intelligence as a “weekend project” during the coronavirus pandemic. Illustration: Carl Nielsen

If you dream of writing a classic whodunit crime novel with Agatha Christie or making a vampire game with a story written in collaboration with the author of the original Dracula novel, Bram Stoker, then there may be hope ahead.

With the writing tool LAIKA, writers of all kinds can get access to an AI-powered text generator that can mimic the writing style that has been fed into the artificial intelligence. At present, LAIKA already supports writing styles of a large number of deceased authors whose work has entered the public domain.

They are called “brains”.

“They are like a brain in a vat,” says Martin Pichlmair, one of the two people behind LAIKA, with a sly smile.

“We have Edgar Allan Poe, Kafka, Marcus Aurelius. You can more or less write a novel together with Dostoevsky.”

LAIKA was created by the Danish-based couple Martin Pichlmair, who is originally from Austria, and Charlene Putney from Ireland.

Charlene has a background as a project manager at Facebook and Google, and she has worked as a writer on a number of games—including the critically acclaimed role-playing game “Divinity: Original Sin II”. Martin Pichlmair also has a background in the gaming world. Among other things, he has founded the game studio Broken Rules, and he is also an associate professor at the IT University of Copenhagen’s Games Programme.

Artificial collaborator

The IT University is also the place where LAIKA came into existence. During the coronavirus pandemic, the couple began diving into artificial intelligence as a “hobby project in the weekend”, Charlene Putney explains. They started holding workshops where people could learn how to use artificial intelligence as a writing tool. Namely, artificial intelligence could act as a collaborator when it was not possible to meet other people.

“As a game writer, I was used to writers’ rooms, where seven people sit and write together. After I came to Denmark, I suddenly had to work alone. The blank page problem became harder and harder to get around,” Charlene Putney says.

The couple’s workshops received really good feedback, but it was clear that it was difficult for the participants to use the AI tools available.

“So we thought, what if we could take these tools we were experimenting with and make them faster, easier, and more accessible,” Charlene Putney says.

With the help of the IT University, the two entrepreneurs sought funding to get the project off the ground. They got that through the Innovation Fund Denmark’s InnoExplorer programme, whose goal is to turn knowledge and research results into concrete products. Now that the funding has run out and LAIKA is in closed beta after half a year of “intense development”, Martin Pichlmair and Charlene Putney have set up their own company. Although the IT University is still associated with the project, the founders now have full control over what will happen to LAIKA in the future.

A new starting point

LAIKA works in such a way that the program can take an incomplete sentence and complete it based on the writing style that has been fed into the artificial intelligence.

“The ideal way to work with LAIKA is that you have, for example, a half-finished novel, or a large text document that describes your game world if you are a game writer. Then you feed all that text into the program, and LAIKA—not by understanding the text, but by understanding the distance between concepts—can complete half-sentences, with something that, for example, fits into the game world you’ve fed into the AI,” Martin Pichlmair says.

LAIKA is created from a modified version of the GPT-2 language model. Although the GPT-2 language model has been fed enormous amounts of text, it is a somewhat smaller language model than, for example, its big brother GPT-3, and there is a good reason for that.

“The really big language models, for example GPT-3 and Bloom, are really good at coming up with a conclusion to what you’re writing in very neutral, contemporary English, because that’s what they’re trained to do,” Martin Pichlmair says.

“But the thing that’s most interesting for writers is writing with a certain type of voice. Therefore, we use the smaller language model and fine-tune it to imitate a specific source material.”

The artificial intelligence can help provide new ideas when the author feels stuck in a text.

“You always get a new starting point,” Charlene Putney says.

“One of the exercises I always do when I teach writing is to sit down and write for ten minutes without stopping at any point. Even if you’re writing something stupid, you’re moving forward. Most writers feel that if you’re constantly trying to write the right thing, you’ll end up stuck. LAIKA can be of help here. You’ll be given three options to complete a sentence. Even if you can’t use any of the options, at least you’ve arrived at a new point. You’ll move forward.”

In addition to providing inspiration, LAIKA also advances the writing process itself.

“Most writers have a process where they write a little, then they edit a little, then they write again, then they edit and so on,” Martin Pichlmair says.

“To a certain extent, you can say that it is the writing parts that are the most difficult. That is why we have tried to make a change where there is more emphasis on editing because the machine does some of the typing for you.”

An example of a sentence completed by LAIKA. The black words are written by Charlene Putney, and the red ones are generated by LAIKA, which has been fed the Grimm’s Fairy Tales. Illustration: LAIKA

However, both Martin Pichlmair and Charlene Putney are very aware of the fact that it is an inspiration tool—not one that increases productivity. For that, it is too much of a “toy”, the couple says. That is one of the conclusions from their user feedback from approximately 1000 beta testers who have used the program so far. The same study also found that the authors who have used LAIKA believe it to be highly useful.

“About 90 percent believe that working with LAIKA stimulates their creativity,” Charlene Putney says.

As previously mentioned, with LAIKA it is possible for a user to upload their own “brain” by feeding the AI their own texts containing a minimum of 10,000 words. Editing what your brain comes up with provides some new perspectives, Charlene Putney says.

“LAIKA can imitate your voice so well that you start to recognize things that you don’t like about your own way of writing,” Charlene Putney says.

“For example, I like ending all my sentences with ellipses... But it could also be too many italics, too many commas, sentences that go on for ages. It shows you what you’re doing and puts it right in your face.”

Lack of EU regulations

There is currently a lot of focus on copyright in connection with the rapid development of generative artificial intelligence. It has also been a problem that Martin Pichlmair and Charlene Putney have had to deal with several times in their work with LAIKA.

“We would really like to know all the details about the texts on which the language models are trained, but very few actually share that information,” Martin Pichlmair says.

“Having said that, it’s not as relevant to our program as it is, for example, to AI-generated images or AI-generated code, because we have fine-tuned the algorithm in LAIKA so much that there is a really large degree of randomization. So you won’t end up in a situation where you’ve suddenly taken whole sentences from one specific source.”

They have ensured that the authors’ work is in the public domain, so that LAIKA does not risk running into legal problems.

“We’d rather not play the lead role in the lawsuit over AI writing tools and copyright that will surely come within the next three years,” Charlene Putney says.

However, the couple is unhappy with the fact that there is still no EU regulation. It would be somewhat safer if there was some legislation to comply with, they say.

“It would be nice to have some kind of a public benchmark to fall back on, because right now it’s a matter of us judging that it’s fine. But we are in unexplored territory with artificial intelligence, so a lot can happen on the legislative side in the coming time,” Martin Pichlmair says.