How I Designed a Choose-Your-Own-Adventure Vacation Itinerary with ChatGPT

By David Griffin Brown

My partner and I planned a trip to Portland for our ten-year anniversary. I had recently read about how AI can be a helpful tool for creating an itinerary. At the same time, I was brainstorming tenth anniversary gift ideas, so I decided to combine the two concepts.

The grand plan: to create an elaborate choose-your-own-adventure itinerary, complete with restaurant options, and to get it all printed up as a fancy booklet. I would also make a few set bookings for some surprise events along the way.

I could have done this all with a search engine, but I thought it would be fun to use Chatbot, since the newest version can now search the web and even cross-reference different websites.

First things first: planning our arrival

We live in Victoria, BC, so our trip to Portland involved a ferry to Port Angeles, Washington, and then about a four-hour drive. That meant we’d be arriving at our Airbnb just before dinnertime. So the first thing I asked Chatbot to do was to take our culinary preferences into consideration along with the Airbnb address to recommend some nice restaurants within walking distance. Bam. Next I asked for two locations within a short drive where we could go for an evening walk. Bam.

This was too easy, I thought! Yes, too easy, indeed. We would soon discover that Chatbot’s tendency to “hallucinate” can get in the way of even its direct search results—but more on that in a minute.

The rest of the itinerary

I gave Chatbot a list of our interests and asked it to assemble a master list of fun things we could do in Portland. Then I asked it to divide the city into quadrants and group our possible activities by location. Next, I told it to break these into morning and afternoon activities, along with well-rated lunch and dinner locales that we’d find in these areas. It produced a wonderful list with a summary of each location and how long we could expect to spend. For the restaurants, it provided a star rating, an average cost per plate, and an example item from the menu.

Matching the optional itinerary to the set bookings

On Thursday evening, I booked an escape room. On Friday, I got us tickets to see a play (Clyde’s at Portland Center Stage). And on Sunday morning, we would visit Tattoo 34 (Chuck Palahniuk’s favorite studio) for some impromptu ink. I asked Chatbot to take these three locations into account and then plan out our options for the rest of our visit.

Here are some of the very cool places Chatbot sent us:

• Mount Tabor Park
• The Bird Alliance of Oregon
• NW 23^rd Avenue
• Portland Art Museum
• Lan Su Chinese Garden
• Goodwill (we love thrifting!)
• Tom McCall Waterfront Park (and the Saturday market)
• Oaks Bottom Wildlife Refuge
• Voicebox Karaoke

Turning an itinerary into a mystery game

Once I had the choose-your-own-adventure itinerary hammered out, I had a new idea. We both love mysteries, so I wanted to incorporate a game of some kind into our five-day vacation. So I asked Chatbot: can you create a mystery for us to solve in Portland, where we tell you each place we visit, and you give us a clue, leading us to guess a culprit and a hideout location?

Yes, Chatbot said, it could definitely do that.

I could have decided on a culprit and location and made the clues myself, but I wanted to involve Chatbot so that we would have to work together to solve the mystery.

Next I had to come up with a theme. We are huge fans (superfans?) of RuPaul’s Drag Race, so I gave Chatbot the following scenario: a drag queen has stolen the crown of the reigning champion (Nymphia Wind), and she’s hidden it at one of our itinerary locations. When we provide you with a location we’ve visited, give us a clue about either the drag-queen thief’s identity or the hideout location. And when you give us these clues, make sure you do it with a lot of drag queen slang.

Beta testing: Going Bananas in Portland

Generative AI can “hallucinate,” as they say. This is what AI programmers call the strange responses that Chatbot can come up with when its dataset doesn’t have a clear answer and it over-extrapolates. For example, I once gave Chatbot a short summary of my fiction manuscript and asked for comparable titles, and not one of the titles it produced was a real book. It turns out, this can also happen when it isn’t quite sure how to process instructions.

In one of my first run-throughs, it gave the following clue: “A parrot flies by, squawking a name, ‘Shangela! Shangela!’ Clue: The culprit is Shangela.”

No, no, no, Chatbot. That’s not a clue. That’s the answer to the mystery.

I moved my mystery-game prompt into Word and tested the game over and over. It was very difficult to get Chatbot to come up with clues that weren’t super obvious (the culprit’s initials) or  ridiculously vague (“the drag queen was wearing a feather boa”).

My prompt grew longer and longer. I included a breakdown of our itinerary along with a list of drag queens with associated trivia and famous quotes. I included examples of what the clues could and could not include.

Finally, I did one last run through and everything seemed to be working. Or so I thought.

Enter: The Hallucination

It turns out I should have fact-checked Chatbot a lot more than I did, just as I should have run another few beta tests of the game.

When we arrived in Portland, we encountered our first hallucination. The restaurants recommended for our first night were in line with our culinary preferences, but they were nowhere near our AirBnB (both about a 30-minute drive away). We also soon learned that the star ratings for the various restaurants did not always match the actual ratings on Google.

As for our location-based itinerary breakdown? It was surprisingly coordinated, although most of the places it had us visiting were centered around downtown. Still, it made many great recommendations that we may not have visited otherwise. The itinerary kept us busy from morning to night—so busy, in fact, that we didn’t attend any of the late-night options since we were worn out by dinner time.

As for the game, Chatbot went off the rails. The culprit of the mystery was Jynkx Monsoon, but we had to make several guesses since some of the clues came from the trivia I’d provided about other drag queens. For example, it told us the culprit likes to wear 1940s fashion—check! And then it told us the culprit was famous for wearing a pineapple-inspired outfit—nope, that was Manila Luzon.

The hideout clues worked, but they ended up too vague for us to draw any real conclusions: it’s outside, there are public washrooms nearby, there are flowers nearby, etc.

Still, it was a fun experiment, and also a worthwhile exploration of the current capabilities of generative AI. The best part was how creative Chatbot got in delivering the clues, which came from dogs, birds, hidden notes, and engraved lockets, not to mention how great it was at using drag queen slang.

David Griffin Brown is an award-winning short fiction writer and co-author of Immersion and Emotion: The Two Pillars of Storytelling. He holds a BA in anthropology from UVic and an MFA in creative writing from UBC, and his writing has been published in literary magazines such as the Malahat Review and Grain. In 2022, he was the recipient of a New Artist grant from the Canada Council for the Arts. David founded Darling Axe Editing in 2018, and as part of his Book Broker interview series, he has compiled querying advice from over 100 literary agents. He lives in Victoria, Canada, on the traditional territory of the Songhees and Esquimalt Nations.