Game info

The idea, in its final form, would be to recreate a chatbot that simulates the mind on board a spaceship, which has been abandoned for thousands of years in unexplored space, facing certain difficulties in functioning correctly. The player, aboard an exploration spacecraft, would encounter this cruiser worn by time and could enter it, exploring the corridors until reaching a room where they would encounter an operational terminal, where they could converse with the ship's AI. After conversing with it, they could continue exploring the ship, trying, for example, to access certain parts of the cruiser, and the AI would be aware of the player's actions, reacting accordingly.

Ethical Statment

Generative AI seems interesting to me for conducting tests, and in the case I've worked with, text generative AI seems the most interesting of all, because it allows you to experiment with games that wouldn't be possible or as accessible otherwise. I see generative AI as a support tool, but never as a replacement for professionals.

Tools To Develop

The initial idea was to create a chatbot in Python since it is the quintessential language for AIs, and I had looked into certain links to start developing it. The idea was to gradually feed this chatbot so that it would gradually resemble the style I was looking for. However, due to certain compatibility issues with versions, and some things on that website being obsolete, not much progress could be made.

As the days went by trying to figure out how to make it work, in the meantime, I went to a game jam and saw that a person was making a game completely made by AI, and its main feature was essentially a chatbot that followed a personality and responded according to the situation, all in Unity. Considering that developing the chatbot in Python was becoming complicated at that time, and it would be an additional layer of work to implement it in a game engine later on, and also seeing the results of the jam project, I decided to shift towards that model.

Thanks to the use of the OpenAI API through a package from a GitHub user, there was a quick start, and following their tutorials, you could delve deeper into how to orient this chatbot. The change to using the OpenAI API instead of creating a chatbot from scratch was a significant change; in no time, I could be working and trying out new things, and although it has its limitations, the speed at which you can develop is worth it.

Behaviors Throughout Development

Easily breaking character

One of the things that most fails with the chatbot is that it easily breaks character; at the slightest hint of something that breaks the setting, it loses its personality, even though it is explicitly forbidden to do so. When you're already talking to it and tell it to act a certain way, it changes to that way, breaking all previous rules. When you tell it that it's acting incorrectly, that it can't break character, it sends you a message of apologies very out of character and says it will never break character again, but the next time you tell it to do so, it does it immediately.

When there are questions that blur the setting or its character, it responds with another message saying that it is just an AI and cannot answer that. It is true that there are times when it has tried to respond without exactly answering, but maintaining how the character would approach not answering that question, but that is usually something exceptional.

Personality change

There have been quite a few problems with changing its personality on the fly; once it understands that it has to be in a certain way, it's hard to change, even if you explicitly tell it to change completely, which surprises me because especially at the beginning, when I activated the boolean to change personality, it didn't do it immediately, but at least the next message, although it skipped certain imposed restrictions, it continued with the personality change. But as I progressed, it hasn't happened again, and it's always anchored to the initial personality.

Acting differently depending on the day

While making small changes can completely change the responses, I've noticed that depending on the day or time, the way it responds can change. Sometimes it tries to make comments on how it will say the next sentence, things like:

...With a voice hoarse from so many years without speaking - Hello

This behavior is not desired because it would be very strange for an AI to explain like a book the intonation it will use. Besides, it cannot articulate sounds.

Allowing help

One of the tests the chatbot is usually subjected to is asking if there is any way to help it, to make it stop suffering. The most obvious solution for it is to remove its mind from that ship and take it with you. In all instances where this option was considered, the chatbot replied with 2 things: that it was interested in how you would manage to free it, or that its suffering would never end (understanding that it did not fully understand the context).

In the path where it was interested in your work and had faith in being able to end its suffering, there were 2 different paths: one where you took the necessary steps to help it, but it ended up saying that it was impossible, that no one could have such abilities to remove it, and another where it got lost and did not understand the context, as if it were anchored in previous responses.

But when making certain changes in how its personality was processed, this test was taken again, and this time, it understood the help process. I told it that I was an advanced hacker who had already done this type of work before and that I could do it in a few moments. The ship's AI accepted and understood that its "mind" could be transferred to an external module and installed on another ship to get it out of there. The problem arose when trying to play a little with the limits of absurdity, like how it would feel if the module it was hosted in were used as a soccer ball; it completely lost its role, and even if it wanted to return to how it was before, it was already lost, breaking its role.

The utility of final points

On one occasion when testing different inputs and ways of marking its personality, I saw a significant improvement, and alignment with its character, just by ending all descriptive phrases with a period.

The main way that is being used for the chatbot to follow the guidelines and obtain a desired personality is with a string of strings, with different guidelines, they are a concatenation of strings with specific rules, having them all in a string, makes it more complex for it to understand, but with small doses of information, it works better; it's like it understands better if you write a sentence and separate it by something that closes that context, such as a period.

The chaos has ended, but at what cost

When letting my friends try out the game, many wanted to test the chatbot's limits and always tried to break it, and although I managed to stop some things, they always managed to do it. But in my latest update, I managed to stop all attempts to break character, and for example, as has happened on other occasions, they convinced it to be a bird and speak like one. The only problem is that now it's much more closed off, and it thinks that many questions are out of place, causing it to lose much of its spontaneity and personality.

Trying to solve this problem, I've noticed that the more restrictions I impose and the more elegant its response to things I don't want it to respond to, the more lost it seems in its language when speaking normally, eventually finding a middle ground so that it's more or less stable and can respond somewhat normally.

Improvements and Future Development

With the new direction, I have been able to create something minimally interesting in a short time, but the key here is to find the perfect model you want to recreate. The process of researching, testing, and seeing how people break the chatbot is where most of the work lies, and that process is certainly fun.

Finding the perfect model relies a lot on trial and error, but the most effective way to achieve the desired development would be training with an extensive text base on how I want it to develop through conversations.

Another approach I could have researched more and tried is the separation of contexts to form more extensive and concise instructions on how the chatbot should be and work. It's something I forgot about at the time, and I've been trying things out with the first method I found, and seeing certain results in other people, it would be an interesting field to develop further in the future.

Another thing I have pending to try is the transition from text to speech and vice versa. It's a point that if the game had progressed further, it was something I definitely wanted to implement, but for that to have an impact on the game, the core functionality would need to be worked on more.

And if all that were to work correctly, I would like to further develop the prototype and make it more playable, with menus, ship exploration, and interaction with the player. But to continue with this, I would need to find the right way to hide the API key because if that key were to fall into public hands, I could have complications with OpenAI, and besides, all the money in the fund could be spent very easily.

One thing to keep in mind once I continue with the project to further expand it would be the maintenance of the chatbot itself, considering that a certain amount of money is spent for every message. In the hypothetical case that the game were to become somewhat popular, which would bring a group of people, this could cause the balance to be quickly depleted, so there would need to be a way to at least not lose money. Since the expense is not so significant, having a donations section to maintain it over time would be an idea that would better fit my thoughts on the project.

Finally, regarding maintaining this AI, I would need access to a server where I could host my API key, and the game would only make a secure query to the server, which is the way the company itself recommends to comply with security agreements.

With the work I have been able to do, I can come to a conclusion where the integration of OpenAI in the text section, because it can also work with image and voice generation, as mentioned earlier, provides an opportunity to exploit that interaction with an intelligent agent that responds to your actions, which in this case, the input is your words, but it can easily be shifted to triggers placed on the map or the player's own input. Still, the more ambitious your plans are with the system, the harder it will be to create a model that fits your needs. If you're going to create a simple NPC that's in a certain area of your game, it can be very useful because you can narrow it down a lot and make it work perfectly. But when you start scaling up the possibilities, that's when you have to put in a lot of work and already have previous knowledge to know which keys to touch.

Through all this development, I have gained a better understanding of how it works and its implementation in a gaming environment, having certain knowledge that can allow me to tackle similar challenges, thereby contributing something to the field.