Lessons from AI
By incorporating artificial intelligence into puzzle building, we’ve learned some valuable lessons that are guaranteed to blow your mind. From the complexities of algorithmic decisions to the unpredictable nature of machine learning (ML), building puzzles using artificial intelligence has been a rollercoaster ride of highs and lows. But fear not, fellow puzzlers – in this blog post we will share our top four lessons learned from this puzzling journey.
Overview of using artificial intelligence in puzzle-building
Escape rooms have been creating puzzles using novel applications of technology for ages that inspire and delight. Reason’s escape rooms in San Francisco have successfully used simple AI chatbots and automation in its puzzles for years. Chatbots especially are a novel way to give players additional interactivity as well as a unique hint mechanism. However, chatbots only have a limited range of responses.
The new large language models (LLMs) and generative AI models provide new opportunities to develop more natural interactions between humans and computers. We expect many escape rooms to at least attempt implementing AI solutions, but we also expect that many puzzles will continue to use more analogue solutions until the technology has evolved a bit more. The following lessons are primarily based on our work with Claude as well as OpenAI’s ChatGPT 3.0 and 4.0 models.
Lesson 1, Functionality: AI might not be sentient, but it has a mind of its own
While testing our puzzle, we hard-wired our answer into our AI model. We wanted to give the AI enough flexibility to respond in ways that would surprise and delight. For the most part, this was great! However, we also realized that despite the hard-wiring, our model would sometimes make up something up or be convinced that a different answer was correct. It also was quite happy to occasionally go off-script and offer puzzle hints that were either completely incorrect or otherwise hilariously off the mark.
What we learned: You need to walk a fine line between giving a model room to have a personality and tying it down hard enough to stay on script.
Lesson 2, Creativity: There are more novel ideas for AI than applications available
One of the problems with being on the cutting edge of technology is avoiding the bleeding edge. We discarded quite a few puzzle ideas after we learned that the capability we were looking for has simply not been created yet. AI can be used to solve problems with static variables, but still stumbles quickly when it comes to novel applications. As we are not ready to jump in and build our own machine learning model, we will need to put some of our ideas on the back-burner.
What we learned: Current models are great at automating away repetitive tasks, but when it comes to more creative uses, we need a little more time.
Lesson 3, Accessibility: Adoption and experimentation are easy
Despite running into issues with turning our ideas into realities, the AI tools themselves were easy to setup, learn, and integrate into our current games. Though we expect this to change as competition thins out, the cost to use and test products was very low, mostly consisting of our time.
We also needed very little coding knowledge to integrate and call our models. Instead, most of our issues were with the nuances of real-world human interaction instead of working with AI specifically, which we cover in Lesson 4.
What we learned: With a little initiative, accessing and testing artificial intelligence models is easy!
Lesson 4, Communication: It’s hard to get the tone right
Communication between humans and computers has come a long way, but it still has a long way to go. In order to cut down on screen reading time, we used an AI voice generator to translate text responses to speech, and were amazed how rude most voices came off. Despite using neutral tones, the flat, rapid responses to players’ questions always came off as a little irritated. Don’t get us wrong, we would love to have a sassy robotic companion, but we want players to feel like they are having a conversation and it felt more like we were talking to an insolent teen who just wanted us to leave it alone. We went through at least 10 different text to voice options before we found one that did not instantly sound harsh or off-putting when answering players.
We also found that the output in general varied wildly from way too long to way to short, which swung from too curt to boringly long. In the end, we found that adding in a list of pre-written responses to go along with the AI’s generated responses allowed us to strike the perfect balance.
What we learned: Though new models provide much more natural language responses, they also offer more opportunities for misinterpretation.
Final Thoughts on the Future of AI in Puzzle-Making
The world of artificial intelligence is vast and growing. We look forward to continued advancements in deep learning, computer vision, and neural networks. As for today’s models, we recommend taking a much more constrained approach. AI can assist in automating a ton of simple tasks that enhance individuals’ productivity, but we’re still limited to the minds of its creators and their biases. It’s important to note that, right now, if you want to use artificial intelligence to create something without well designed rules and processes, you need to feel comfortable with any number of unexpected outcomes. You also need to accept that AI is not a replacement for human interaction, i.e your AI’s responses to your players may not instill the right emotions in your players.
While the process was challenging, the end results were worth the effort and we will continue to experiment and implement new puzzles. We hope that our 4 lessons help ease you into your own introduction to AI and we would love to hear about your own explorations in to your own puzzle-building endeavors!