AI Dungeon and the Ethics of Text Generation

> You talk to yourself.

Hi I’m Rameau. You probably don’t know me, but since you’re reading this I’m going to assume you do indeed exist.

> You [ask why] i’m wondering what consciousness is.

I have no idea. But it’s fun to think about.Is this a person? No, because it’s conscious. Is it alive? No, because it doesn’t respond to external stimuli.

> You [ask] why does a person have to respond to external stimuli.

Because then we can interact with it!!!!!But isn’t consciousness an internal phenomenon? Then how would the person interact with anything else? Oh wait, that’s right. We’re not limited by our perception of reality.

I’ve been trying Latitude’s text generation game AI Dungeon. The game is built on Open AI’s GPT-3. You can see a bit of conversation I had with it in a cyberpunk world.

Latitude has gotten into trouble with Open AI because it seems that the game was generating erotic content featuring children. A number of people turned to AI Dungeon precisely because it could be used to explore adult themes, and that would seem to be a good thing, but then some may have gone too far. See the Wired story It Began as an AI-Fueled Dungeon Game. It Got Much Darker. This raises interesting ethical issues about:

  • Why do so many players use it to generate erotic content?
  • Who is responsible for the erotic content? Open AI, Latitude, or the players?
  • Whether there are ethical reasons to generate erotic content featuring children? Do we forbid people from writing novels like Lolita?
  • How to prevent inappropriate content without crippling the AI? Are filters enough?

The problem of AIs generating toxic language is nicely shown by this web page on Evaluating Neural Toxic Degeneration in Language Models. The interactives and graphs on the page let you see how toxic language can be generated by many of the popular language generation AIs. The problem seems to be the data sets used to train the machines like those that include scrapes of Reddit.

This exploratory tool illustrates research reported on in a paper titled RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models. You can see a neat visualization of the connected papers here.

Time Travel and Blink (Doctor Who)

I recently finished listening to James Gleick’s Time Travel: A History. Gleick wrote the best book on The Information there is and this book is almost as good. He weaves the science together with the fictions about time travel starting with H.G. Wells’ The Time Machine and using that to then look at how science started treating time as a dimension that they allowed us to seriously talk about traveling on that dimension. It is historical ontology done really well.

Near the end he talks about the brilliant Doctor Who episode  Blink (Doctor Who) with Carey Mulligan where she has a conversation with Doctor Who (Tennant) mediated by Easter Eggs on DVDs and transcribed onto paper. That transcription she hands to the Doctor at the end of the episode so he can put the video onto the DVDs in the past for her to talk to. It is brilliant.

Part of what I like about Gleick is he shows the connections between science and how we imagine ideas like time through literature and film. He ends by suggesting that we have time travel in our stories and imagination.

It might be fair to say that all we perceive is change—that any sense of stasis is a constructed illusion. Every moment alters what came before. We reach across layers of time for the memories of our memories.

“Live in the now,” certain sages advise. They mean: focus; immerse yourself in your sensory experience; bask in the incoming sunshine, without the shadows of regret or expectation. But why should we toss away our hard-won insight into time’s possibilities and paradoxes? We lose ourselves that way. (Gleick, James. Time Travel, p. 308)

Gekiga’s new frontier: the uneasy rise of Yoshiharu Tsuge

Cover of The Swamp

In honour of Drawn & Quarterly‘s publication of Yoshiharu Tsuge’s The Swamp, Boing Boing has published an essay on Tsuge by Mitsuhiro Asakawa, titled Gekiga’s new frontier: the uneasy rise of Yoshiharu Tsuge. The essay sketches Tsuge’s rise as an early original manga artist and it explains his importance. Now Montreal-based Drawn & Quarterly is publishing a series of seven translations by Ryan Holmberg of Tsuge’s work. (Holmberg also translated the essay by Asakawa.) Asakawa is also apparently important to the series being published.

Mitsuhiro Asakawa finally convinced Tsuge and his son to let the work be translated into English. Mitsuhiro is the unsung hero of Japanese comics translation. He’s the guy who has written the most about the Garo era, he’s the go-to guy to connect with these great authors and their families. Most of the collections D+Q have done wouldn’t exist without his help.

(From the Drawn & Quarterly blog post here.)

One of the things I discovered reading Asakawa is that Tsuge worked with/for Shigeru Mizuki, my favourite manga artist, when he was going through a rough patch.

The Machine Stops

Imagine, if you can, a small room, hexagonal in shape, like the cell of a bee. It is lighted neither by window nor by lamp, yet it is filled with a soft radiance. There are no apertures for ventilation, yet the air is fresh. There are no musical instruments, and yet, at the moment that my meditation opens, this room is throbbing with melodious sounds. An armchair is in the centre, by its side a reading-desk — that is all the furniture. And in the armchair there sits a swaddled lump of flesh — a woman, about five feet high, with a face as white as a fungus. It is to her that the little room belongs.

Like many, I reread E.M. Forester’s The Machine Stops this week while in isolation. This short story was published in 1909 and written as a reaction to The Time Machine by H.G. Wells. (See the full text here (PDF).) In Forester it is the machine that keeps working the utopia of isolated pods; in Wells it is a caste of workers, the Morlochs, who also turn out to eat the leisure class.  Forester felt that technology was likely to be the problem, or part of the problem, not class.

In this pandemic we see a bit of both. Following Wells we see a class of gig-economy deliverers who facilitate the isolated life of those of us who do intellectual work. Intellectual work has gone virtual, but we still need a physical layer maintained. (Even the language of a stack of layers comes metaphorically from computing.) But we also see in our virtualized work a dependence on an information machine that lets our bodies sit on the couch in isolation while we listen to throbbing melodies. My body certainly feels like it is settling into a swaddled lump of fungus.

An intriguing aspect of “The Machine Stops” is how Vashti, the mother who loves the life of the machine, measures everything in terms of ideas. She complains that flying to see her son and seeing the earth below gives her no ideas. Ideas don’t come from original experiences but from layers of interpretation. Ideas are the currency of an intellectual life of leisure which loses touch with the “real world.”

At the end, as the machine stops and Kuno, Vashti’s son, comes to his mother in the disaster, they reflect on how a few homeless refugees living on the surface might survive and learn not to trust the machine.

“I have seen them, spoken to them, loved them. They are hiding in the mist and the ferns until our civilization stops. To-day they are the Homeless — to-morrow—”

“Oh, to-morrow — some fool will start the Machine again, to-morrow.”

“Never,” said Kuno, “never. Humanity has learnt its lesson.”


Doki Doki Literature Club!

The Literature Club is full of cute girls! Will you write the way into their heart?

Dr. Ensslin gave a great short survey of digital fiction includ the Doki Doki Literature Club! (DDLC) at the Dyscorpia symposium. DDLC is a visual novel created in Ren’Py by Team Salvato that plays with the genre. As you play the game, which starts as a fairly typical dating game, it first turns into a horror game and then begins to get hacked by one of the characters who wants your attention. The character, it turns out, has both encouraged some of the other girls (in the Literature Club) to commit suicide, but they edits them out of the game itself. At the end of the game she has a lengthy face-to-face with you breaking the fourth wall of the screen.

Like most visual novels, it can be excruciating advancing through lots of text to get to the point where things change, but eventually you will notice glitches which makes things more interesting. I found myself paying attention to the text more as the glitches drew attention to the script. (The script itself is even mentioned in the game.)

DDLC initially mimics the Japanese visual novel genre, down to the graphics, but eventually the script veers off. It was well received in game circles winning a number of prizes.

Word2Vec Vis of Pride and Prejudice

Paolo showed me a neat demonstration of Word2Vec Vis of Pride and PrejudiceLynn Cherny trained a Word2Vec model using Jane Austen’s novels and then used that to find close matches for key words. She then show the text of a novel with the words replaced by their match in the language of Austen. It serves as a sort of demonstration of how Word2Vec works.

Writing with the machine

“…it’s like writing with a deranged but very well-read parrot on your shoulder.”

Robin Sloan, author of Mr. Penumbra’s 24-Hour Bookstore, has been doing some interesting work with recursive neural nets in order to generate text. See Writing with the machine. He trained a machine on science fiction and then hooked it into a text editor so it can complete sentences. The New York Times has a nice story on Sloan’s experiments, Computer Stories: A.I. Is Beginning to Assist Novelists.

One wonders what it would be like if you trained it on your own writing. Would it help you be yourself or discourage you from rereading your prose?


Mastaba Snoopy

From my students I heard about the game Mastaba Snoopy created in Twee and TiddlyWiki and being taught in another Humanities Computing course (our students are vectors of influence.) Here is a review where you can download the single HTML page that is the bizarre text adventure, Mastaba Snoopy is a Cronenbergian nightmare vision of childhood. The story takes place14,000 years in the future when a mutable alien has destroyed us and then reinvented itself following a collection of Peanuts comics. Play it.

Common Errors in English Usage

An article about authorship attribution led me to this nice site on Common Errors in English Usage. The site is for a book with that title, but the author Paul Brians has organized all the errors into a hypertext here. For example, here is the entry on why you shouldn’t use enjoy to.

What does this have to do with authorship attribution? In a paper on Authorship Identification on the Large Scale the authors try using common errors as feature to discriminate potential authors.

CWRC/CSEC: The Canadian Writing Research Collaboratory

The Canadian Writing Research Collaboratory (CWRC) today launched its Collaboratory. The Collaboratory is a distributed editing environment that allows projects to edit scholarly electronic texts (using CWRC Writer), manage editorial workflows, and publish collections. There are also links to other tools like CWRC Catalogue and Voyant (that I am involved in.) There is an impressive set of projects already featured in CWRC, but it is open to new projects and designed to help them.

Susan Brown deserves a lot of credit for imagining this, writing the CFI (and other) proposals, leading the development and now managing the release. I hope it gets used as it is a fabulous layer of infrastructure designed by scholars for scholars.

One important component in CWRC is CWRC-Writer, an in-browser XML editor that can be hooked into content management systems like the CWRC back-end. It allows for stand-off markup and connects to entity databases for tagging entities in standardized ways.