The Fortuna, by Jason Gauci
For the past year or so, the IF community has seen numerous conversations about the ethics, efficacy, and prospects for using AI tools like Large Language Models or image-generation software to create IF. Various arguments have been advanced over epically long threads – often throwing off as much heat as light, it must be admitted – but perhaps our time would have been better spent waiting quietly, because sometimes an ounce of practice is worth a pound of theory. The Fortuna is perhaps a maximalist take on what using AI in IF can look like: every element of it, from the graphics illustrating each location and character, to the descriptions that flesh out its cruise-ship milieu, to the freeform conversation system that’s central to progression, is built around AI. And every single one is awful.
Looks, cards on the table: I come to this debate pretty skeptical of the AI pitch. One of the major reasons I engage with art is because it offers an opportunity to connect with other human minds, to expand my understanding and my perspectives, to experience something idiosyncratic and specific to the person who made it – LLMs and other AI approaches, with their views from nowhere, get in the way of that. I also find most procedural-generation pretty boring; I get that in theory it can open up new intellectual possibilities and reveal surprising connections, but in practice I tend to experience it as one bowl of oatmeal after another. So I may have some biases (I also am not really big on cruise ships, now that I think about it – never been on one, so my major associations with cruises are the destruction of Venice’s lagoon, David Foster Wallace’s A Supposedly Fun Thing I’ll Never Do Again, and the time my brother-in-law went on one and came back with swine flu). But even if I’d gone into The Fortuna a dyed-in-the-wool techno-optimist, what I found there would give me second thoughts.
Start with the art. Due to the use of AI, the game is lavishly illustrated by the standards of amateur IF. But instead of providing a nice aid to the imagination, allowing the player to more fully immerse themself in the world, instead we get, well, this:
Allegedly this is a set of greeters for the cruise ship, not a band of Cenobites on holiday from hell. Allegedly. But it’s not an outlier; all of the images are like this to a greater or lesser extent.
It’s hard to discern exactly what’s gone wrong here – obviously the image-generation tool got confused somehow. But aren’t there other tools that work better that the author could have used? And shouldn’t he have exercised a minimum of quality control and noticed how disturbing this picture is, how out of step it is with the happy vacation vibe the prose is working to convey, and killed it?
Speaking of the writing, it suffers from the same uncanny valley issues as the visuals; the well-known tendency of LLMs to hallucinate is on full display, and the prose manages to be vague and repetitive. Like, here’s the description of a uniform in a gym locker:
The uniform in the gym locker room is a set of clothing worn by individuals for a specific purpose or profession. It is likely made of durable and breathable material to withstand physical activities. The uniform may include a shirt, shorts or pants, and possibly a jacket or other accessories depending on the specific sport or activity taking place in the gym. It may also have the logo or name of the cruise ship or gym embroidered on it. The uniform is typically neatly folded or hung in one of the lockers, ready to be worn by individuals engaging in physical activities in the gym.
This is a hundred words that don’t communicate anything beyond “a uniform in a gym locker”, and of course there are an awful lot of “may”s here – in fact, the uniform may not actually exist, given that I couldn’t try to take it.
The writing issues combine with the game’s conversation-focused design to create still further problems. In several places, progression requires you to build a rapport with various of the characters you meet, which is largely accomplished by asking them about themselves. This is accomplished via a freeform chat system, which is impressively open-ended and does allow for a wide variety of input to be understood (one character writes for the New York Times, and when I asked her about Jamelle Bouie, she correctly identified him as a journalist). Still, this element too runs into some familiar problems. The left hand may not know what the left is up to – as with one early character, who’s described as toying with their wedding ring in a bit of environmental detail, but who steadfastly insists they were never married when you ask them about it. And when you do learn more about them, well:
> ask officer “how are you”
I’m doing well, thank you for asking. I believe in hard work and dedication, and I strive to make the world a safer place as a port security officer. Ensuring that everyone can feel safe is important to me, and I hope to be a role model for young people. Growing up in a rough neighborhood and seeing the violence and crime firsthand motivated me to make a difference. After high school, I joined the military and served in Iraq for four years. I then pursued a degree in criminal justice and landed a job in port security. I want to show others, especially young people, that they can achieve anything they set their minds to.
This is generic and robotic, of course, but it’s maybe worth noting that this character is Black, and I dare say there’s something pretty stereotypical about the backstory being presented. The one other Black character has a similarly tormented backstory involving struggles with addiction. Meanwhile, when I asked the Times reporter about her job, after saying it’s a dream come true, she continued with “[a]s a female reporter, I am proud to work for a newspaper that values diversity and promotes women’s rights.”
To just state what I’m getting at: the characters who aren’t white men are largely stereotypes, defined by their race and gender. This is something a human author can unwittingly wind up doing, of course – but it’s also a real danger of LLMs, trained as they are on the products of a racist, misogynistic society. At least there’s a note of comedy from how the model extrapolates without any understanding of the world: there’s a devout Italian woman who’s super into C.S. Lewis, because the AI’s mashed up different stereotypes without knowing the difference between Catholics and Protestants.
Now, it’s possible that these characters get deeper over the course of the game, and that it improves in some of its other problematic aspects too, because I have to confess that I didn’t get very far into the plot. After boarding the ship, entering the VIP section, and being introduced to the cast of characters, there was an inciting incident that pointed to a mystery to be solved that presumably kicks off the narrative proper – but one of the first steps required solving a riddle to open a safe. The riddle seemed like one of those old chestnuts everyone knows, albeit with awkward, AI-y syntax (”I am owned by the poor and the rich don’t need me” – it’s nothing, right?) except the obvious answer didn’t work, nor did a bunch of non-obvious ones I tried. Is this because I was typing things in wrong, or being especially thick, or did the AI generation flub the riddle? I don’t know, but regardless, there my journey came to an end, and I must admit to feeling some relief.
I am being uncharacteristically harsh on The Fortuna, and I should say it’s not because I think the game is coming from a bad place by any means. In fact, the best piece of writing in the game is the introduction, where the author speaks movingly about being inspired by IF as a child, and wanting to use the new possibilities afforded by AI tools to create a game that would similarly spark joy in its players. That’s a laudable goal, and one that pretty much all first-time authors fall short of (I know I certainly did!)
But I do think there’s something very badly broken about the approach here. While there may be a case for the IF community to embrace AI, it turns out that using it for every aspect of a project and dialing things up to 11 does not make for a convincing demonstration. The Fortuna’s various failures all, I think, have one root cause – which is that time after time, when an AI tool threw up something that didn’t work or didn’t make sense, the author didn’t take action to cut, modify, or customize in service of the larger vision. As the amateur IF scene is currently constituted, it’s an auteurist medium; one can certainly levy critiques about that situation, but as it turns out, taking the author out of the equation is a very bad idea.