This article is part of a series of posts about AI. It’s written by a human being (me) but this and other posts may include examples of identifiable AI-generated response text and imagery.
It began with a question:
+ Can you simulate a text adventure?
ChatGPT assured me that it could:
Absolutely. I can run a fully interactive text adventure
where you choose what to do, and I handle the world, characters, consequences,
inventory, combat, puzzles, and hidden events.
It went on to suggest several themed settings, but also
suggested that we could just start, and spun off the first few lines of a game
called THE DOOR UNDER THE CITY. So that’s we did.
ChatGPT created a horror-themed game involving escaping from an apartment. There were mysterious phone messages, appearing footprints, lights shining through the floorboards, and
when I drank a mug of tea I found that not only had it gone cold, but my own eyeball was looking up from inside it. I also encountered an increasing number of doubles of myself, each instructing me not to trust the others.
The game played out like a classic text adventure in that I typed commands and ChatGPT returned the next scene or event. An inventory was maintained - at the start I picked up my phone and mug from the room, and was later able to use the phone light in a dark coridoor and to drink the tea. The map was also fairly consistent, or at least whenever I retraced my steps I found myself in the previous location with changes consistent with the theme such as the appearance of a new door.
I was not able to predict anything about the story - as I'd hoped. The dark atmosphere was maintained and the repeated messages and threats, the indications that something bad was always approaching, knocking or trying to break in, and the eyeball in my tea, all gave the game a nicely tense flavour.
Partway through, confronted by several clones, I asked them what was going on. The answer came in the form of an argument between one clone who provided a theory about time travel, with clones as past and future versions of myself, and another who told me that this was a lie and:
Every time you make a choice, the building creates another version of you. Most of them don’t survive.
The game ended after, having escaped to the street and encountered a mysterious taxi, I opened a door standing alone on the pavement and found myself back in the apartment. The fact that the game reached an ending by itself is significant given what happened in some of my later experiments.
Although atmospheric, the story was repetitive. Doors, knocking sounds, approaching threats, appearances of doppelgängers and the warnings not to trust anyone occurred over and over through the game. There was a sense that ChatGPT became obsessed with a particular idea, in this case the doppelgängers.
Also, despite the continuous sense of threat, nothing bad ever happened whether I chose to react to the warnings or did the opposite. If I turned around to face a sinister entity there was never anything there. I wondered if ChatGPT was avoiding writing about actual pain or harm. Could this be an example of a guiderail? On the other hand ChatGPT may have been deliberately focussing on atmospheric horror, or preparing the ground for some actual danger later.
I experienced an interesting counting error when I met the second double. Here's what happened:
You turn around. There is nothing there. No creature. No doorway. No corridor. Just yourself. Except this version of you is standing much too close. He looks terrified. Behind him, your other self—the one beside the red light—is
staring at you.
Now there are three of you.
The closest one whispers: “You shouldn't have turned around.”
The second one says: “Don't listen to him.”
The third says: “Neither of us is the real one.”
The red light goes out. In the darkness, all three voices speak at once: “Choose.”
What do you do?
By the end of the response the two clones had become three. My guess is that ChatGPT first wrote the line "now there are three of you" - which is correct, but then wrote the rest of the response interpreting the same statement as three clones in addition to myself. This type of counting error also occurred in later experiments.
I don’t think I learned a great deal about AI psychology but I made a start. I have learned that ChatGPT already knows what a text adventure game is, and can
create and run one with no instructions, although the results might be chaotic,
limited, repetitive, or obsessed with doppelgängers; ChatGPT can make counting errors, possibly by responding to the meaning of its own statements rather than the actual mathematics of the scene. I suspected that some aspects of the game might have been restricted by guardrails.
I also suspected that my decisions were actually driving the creation of the game, although in an unpredictable way. Writers sometimes refer to being “plotters”
(those who plan ahead) or “pantsers” (those who fly by the seat of their
pants). I wasn’t certain at this point but my impression from this game was
that ChatGPT is a pantser pretending to be a plotter but actually making
everything up on the fly.
Out of curiousity I asked ChatGPT to create a cinematic poster based on the transcript of the game. Here's the result, which does include recognizable plot details mainly from the second half of the game.I reflected that, while the idea of starting a completely
random adventure was appealing, the result wasn’t quite what I wanted from a
game. For example I would have liked more of a story or mission structure, and
the doubles of myself were really boring – I wanted a world populated with more
interesting people. Perhaps I could write a detailed prompt to give ChatGP a clearer direction
for the setting and the kind of adventure I wanted to experience.
So that’s what I tried next.
Posts
