As mentioned in the synopsis, Amie, Android is written with OpenAI’s GPT-3 neural network. More specifically, I am using the model accessible through the text-based game AI Dungeon 2. The game offers two versions of the AI, called “Griffin” and “Dragon”. Griffin is a simplified version of GPT-3 available to all players, while Dragon is more robust but requires a paying subscription. For the moment, this story is being written with the former, as I like the idea that anyone reading Amie, Android can play the game and interact with the exact same AI used to write the story. There is also the consideration that writing and editing chapters is a time-consuming affair. Writing this as a hobby and posting it for free is one thing; regularly paying to do so is quite another. If we can build up a readership willing to chip in for the ten dollar monthly subscription and there is demand for it, then I will consider testing Dragon and possibly switching to it for the remainder of the story.

Regardless of whether one uses Griffin or Dragon, however, the game’s interface is much the same. Here’s an example of what the story looks like during the writing process:

https://images2.imgbox.com/a0/46/AvCEOt5Z_o.png

On the AI Dungeon 2 server there are no chapters; the story exists as a single mass of text which I have copied over, split into chapter form, and periodically update on my PC. Tragically, the existent formatting doesn’t carry over, giving us this abominable result:

https://images2.imgbox.com/0a/0f/zXTApKry_o.png

The editing process alone therefore eats up a significant chunk of time which could otherwise be spent writing…or, as is often the case, redoing. AI Dungeon 2 has an undo function to compensate for the inconsistent quality of generated text, and I have recourse to it whenever the generated text is truly hopeless. Some chapters are very smoothly written, and require few retries; others demand well over a hundred. Writing a coherent narrative with AI Dungeon 2, and Griffin in particular, is not for the faint of heart. The AI has several negative tendencies to contend with, including but not limited to:

1. Deciding to wrap up a chapter prematurely by means of a fake author’s note:

https://images2.imgbox.com/26/2e/89INfQHA_o.png

2. In the same vein, detecting the poignancy of a statement and deciding to end the entire story right then and there:

https://images2.imgbox.com/d6/12/Di0OpVqD_o.png

(While it’s impressive that the AI can identify profound moments, I think the reader will agree that the above conclusion leaves something to be desired!)

3. Materializing random characters out of thin air:

https://images2.imgbox.com/37/35/eUb8jJiH_o.png

4. A penchant for crass materialism; thirty retries were needed to overcome this fit of epicureanism:

https://images2.imgbox.com/1f/03/aJMjoFym_o.png

5. Conversely, there is sometimes an overly religious bent in need of mitigation:

https://images2.imgbox.com/e8/65/ab21mU04_o.png

https://images2.imgbox.com/a7/1b/na7oZUiB_o.png

6. Other times the text is coherent, but the tone is simply off:

https://images2.imgbox.com/e2/33/6PZuxAzy_o.png

7. Also, the AI has a bit of a dark streak:

https://images2.imgbox.com/82/31/634zMVZM_o.png

https://images2.imgbox.com/a4/c9/Nof6iCDG_o.png

8. And very rarely, off-color humor makes an appearance…

https://images2.imgbox.com/2d/f4/oIHrEuZN_o.png

9. …while on a more light-hearted note, there is no shortage of unintentional comedy:

https://images2.imgbox.com/37/65/ZMXWJo1d_o.png

https://images2.imgbox.com/12/a3/8OixoucB_o.png

https://images2.imgbox.com/d3/d3/Lat8GQlP_o.png

10. And, finally, we have sporadic bouts of sentimentalism. Since Ely and Amie are ostensibly husband and wife, the AI can be tempted to devolve the story into a sappy romance; dialogue like this I delete with a vengeance:

https://images2.imgbox.com/3d/d0/MItA62gs_o.png

I caution the reader against concluding from these unfortunate samples, however, that the AI requires constant handholding to put together useful material. It manages to produce meaningful statements on its first try surprisingly often. Here is an example of one such instance:

"But..." Amie says, stumbling over her words, "We have to give something, so that the gift can continue. We can't just refuse a giver. And we can't be afraid. We have to love."

This outcome was reached in three stages. First, I prompted the AI write on its own, which rendered the following:

1. “But…” Amie says,

At this point I could have pressed Enter to generate more dialogue, or provided my own contribution. I chose to add, ‘stumbling over her words’, giving the AI more context to work with:

2. "But..." Amie says, stumbling over her words,

I then prompted the AI again, and it finished the sentence in the following manner:

3. "But..." Amie says, stumbling over her words, "We have to give something, so that the gift can continue. We can't just refuse a giver. And we can't be afraid. We have to love."

No second attempts were needed here; the original generation sufficed admirably for the story.

The AI is also capable of writing multiple intelligible sentences all at once. On one occasion I wrote the following:

1. After a brief pause, you ask, "Can you identify why, specifically?"

Amie looks at you for a long moment, as if to discern all the hidden meanings in your question.

I then prompted the AI to continue, resulting in the bolded text:

2. After a brief pause, you ask, "Can you identify why, specifically?"

Amie looks at you for a long moment, as if to discern all the hidden meanings in your question. Then, she blushes again and looks away. "I-I'm not sure. I think it's because you...you make me feel cared about. You made me happy when I wasn't happy before, and you're treating me like... like a lady." She lowers her eyes to the table once more. "And I... I really like that."

As we can see, it’s not only dialogue that the AI handles; Amie’s blush, her looking away, lowering her eyes to the table, etc., were all written by it. In this manner the AI heavily shapes her characterization and the story’s tone, as can also be observed here:

"Yes. And I did not wish to drink this wine until it meant to you as much as it does to me." You give her with a smile—for once, a genuinely gentle smile.

Amie looks down, perhaps a bit bashfully. "Thank you, Mr. Brennan," she says in a quiet voice. Your eyes meet. "I... look forward to the rest of this journey."

You nod. "You're very welcome, Amie. And now, what do you say we have our first sip?"

She smiles. "Yes, sir."

You both lift your glasses. Staring into each other's eyes, you bring together your glasses in a nearly inaudible 'clink', then raise them after your voiceless toast and drink. This moment, this precursor to all the other moments you share with her, is a good one. For now, at least, Amie has her back to the cold, endless void she fears so much. The path you are on is a curious one, and you are joined by an unusual companion, but you feel a burgeoning hope inside you that the rest of the journey will be as fruitful as today.

While only one sentence in the paragraph immediately above is AI generated, there are times when I allow it to write at much greater length. Readers who are interested in knowing whether specific passages are human or AI written are free to inquire to that effect in the comments.

There is another challenge to the writing process, and it is the most consequential of all. Ely’s personal philosophy has essentially been created from scratch, causing the AI to struggle mightily. Though GPT-3 has been fed hundreds of billions of words, since Ely’s discourses have no parallels in its database, that prodigious amount of data is of little help in helping the AI rebound off his statements. More often than not, our poor Amie is stuck. This is the limit of language generating AIs, and I don’t expect it to be broken ever, much less within our lifetimes.

And yet, Amie is our second leading character, and she speaks quite a bit. So what’s the solution? Sometimes, I’ll admit to using opportunities served on a silver platter, as can be seen in the following extract:

You mightily refrain from rolling your eyes. "I'm glad to hear it. What I would also like to hear, however, is a drawing out of the principle we have previously elucidated: to wit, the usage of the material things around us as intermediaries with which to strength our immaterial bond. How is this principle applicable to our dinner?"

Amie looks at you expectantly, and you continue: "In sharing a meal we are speaking a language of sorts. Earlier we went over the various benefits of wine—pleasure, increased sociability, exuberance, even health... similarly, this food offers its own set of benefits, and generally a hearty steak dinner will be well regarded and reserved for a special occasion. By selecting this meal and drink I have therefore made a series of implicit statements. Does that help you get on the right track?"

As a sentence becomes more complex and long-winded, the AI’s ability to respond to it coherently decreases proportionally. The question presented by Ely above far outstrips the AI’s ability. When the AI is well and truly stuck, it will simply reply “I don’t understand”, “That sounds complicated,” “Please, go on,”;—short queries or statements, in other words, which serve as deflections. To compensate for this, I will sometimes begin sentences for Amie, nudging the AI in the right track so it can actually play off my statements; it’s no fun if I do all the writing in response to befuddled “I don’t understand’s”, after all. Other times I simply refuse to let the AI off the hook, as can be seen in this exchange:

You nod, and consider the android's words. "Neither do I. Our very conversation is proof of that. What was my first thought earlier, when you brought up having wine?"

"You thought I wanted to celebrate my humanity." Her eyes widen slightly in understanding. "Oh! That is another reason. Yes, wine is often served at celebrations."

"Which ones specifically?" you prod.

"At any major coming of age or transition in life, like a birthday or marriage. It's...um...it's sort of a sign of affection."

You lean forward a little, trying to keep the intrigue from showing on your face. "Why is it a sign of affection?"

She shifts uncertainly in her seat. "Well, affection is something humans value. To not offer it at a major life transition would be like showing someone you don't care about them."

"So, offering a bottle of wine is like pouring your heart into the gift?"

She nods. "Basically."

"Making wine an instrument, a conduit of affection, of love, between two hearts."

She ventures to nod a few times. "Yes. Yes, exactly."

"What if a man purchases several bottles of wine and drinks alone until he passes out in his apartment?" you suggest. "Is there love there?"

She frowns, unsure what to say. "Well... no?"

"What is the wine a conduit of, in that instance?"

"Of loneliness," she answers thoughtfully.

The reader will notice that Ely’s line, "So, offering a bottle of wine is like pouring your heart into the gift?" is bolded above. This isn’t a mistake. As many of you have undoubtedly deduced while reading, the artificial dialogue in the story is not exclusive to Amie. Though to a much lesser degree, a certain portion of Ely’s lines are artificial as well, making his and Amie’s interactions a blend of organic and artificial writing. While it usually isn’t difficult to tell when Ely has “my” voice, there are a few artificial lines that mesh surprisingly well with my writing style, and even my family members have had trouble distinguishing between my writing and AI generated text at times.

Finally, I’ll take this opportunity to mention that I have a non-AI fiction ongoing on Royal Road, which can be found here: https://www.royalroad.com/fiction/36157/selenas-reign-the-golden-gryphon

Readers may be interested in comparing my “real” writing with that of this story. Of course, I am very grateful for the support both titles have received and appreciate your feedback on either one of them.