AI trip planning: a journey into the unknown with chatbots

Cards on the table. I’m not a ‘grab your passport and see where the wind takes you’ kind of girl, I tend to plan every aspect of my trips meticulously, pouring over multiple sites to find the best restaurants and hotels, checking their respective proximity to public transport links. I’ll have wandered the streets of my preferred destination weeks in advance via Google Streetview and researched opening hours of every damn museum within a 10-mile radius. I’ll even have a special hard copy travel folder with print outs of, well, everything related to the trip. Laminators may even end up being involved at some stage.

Call me Mrs Anally-retentive, but – hey – that’s just how I roll.

Can I get AI to sort all my travel arrangements instead?

For all you crazy kids who prefer a smarter, more stream-lined way to plan and book travel (and that’s a lot of you, it would seem), you’re in luck. With the big boys of travel such as Expedia and Kayak already embracing AI technology to encompass all areas of trip-planning in their apps, the days of having 25 tabs open on your laptop are coming to an end. Possibly.

Chatbots are getting in on the action too. Back in September, Google Bard rolled out Bard Extensions – a shiny new feature which will allow travellers to plan trips up until the point of booking, without having to leave the app.

For example, imagine you receive a work email from a colleague proposing dates for a forthcoming trip to Stockholm. Bard can now grab those dates from your inbox (without you having to search for the original email) and suggest flights, hotels, restaurants, activities and even offer tips on what clothes to pack — all within one conversation.

Let’s back up a bit and explore how we got to this point.

AI and travel – the story so far

Travel planning and AI have been uneasy bedfellows to date, with some amusing faux pas occurring along the way, as demonstrated in our blog Navigating the AI Search Revolution:

But that was back in March, and as we all know, 8 months is basically 8 years in AI years. Plus, while we’ve all been amusing ourselves with chatbot hallucinations, Kayak was busy integrating GPT plugins onto its platform. Voted one of the best AI travel apps out there, Kayak uses natural language processing (NLP) and algorithms to process hundreds of thousands of flights and user-generated hotel reviews, condensing and categorising them, while integrating the platform into the likes of Google Assistant, Amazon Alexa and more.

I want to get back to the chatbots though, as these are the primary focus of this blog. My aim is to explore just how effective they are in booking travel and making travel-related recommendations.

To test the capabilities of each of the three main AI chatbots in travel planning (at present being Bing, Bard and ChatGPT), I presented them with three individual challenges. Would they rise to the occasion or end up in a hallucinatory muddle?

Let’s find out.

Challenge 1: Create a walking tour around Alnwick (ChatGPT)

First up is the rising star of 2023 – AI’s ubiquitous poster boy, ChatGPT.

For this challenge, I wanted to explore whether ChaptGPT could create an itinerary for a walking tour. I chose a place I know well - the Northumbrian market town of Alnwick, a 45 minute drive north of Newcastle. I gave ChatGPT the prompt ‘create a walking tour of Alnwick’. No further context was provided, so I was expecting something pretty generic and basic….and GPT delivered exactly that.

At first glance, this looks fine and ticks off all the must-sees of the town. But I have the benefit of knowing Alnwick and all its haunts well. What if a tourist from overseas was dropped in the middle of the town and had nothing but this walking tour as a guide?

The major flaw here is that despite prompting GPT for a walking tour, it provides no explanation as to how to get from one location to another. Vague descriptions such as ‘a short walk’ are not all that helpful, with ‘short’ being somewhat subjective (for the record, the walk from Alnwick Castle to Barter Books is not actually *that* short a walk…it’s over half a mile which, if you were to have limited mobility, could be a problem).

OK, there’s an element of nitpicking here, especially as the original prompt gave GPT little to go on in terms of context or audience. Let’s refine the prompt:

This is essentially the same tour as previously provided, but with additional guidance regarding accessibility. What leaps out at me is that this doesn’t read like it’s been created by someone who has limited mobility themselves – once again, it feels vague, with little in the way of insider tips. For example, it fails to mention the cobbles around Market Street which could be an issue for people with limited mobility, nor does it name any specific cafes that are step-free or particularly accessible.

And once again, it doesn’t actually explain how to get from one location to another.

The other well-documented issue with GPT is the lack of citations. While the information provided is pretty much correct, there’s no way of knowing where the chatbot gets its information from. What is the source? How can the reader trust that the details provided are correct without further investigation? This has been one of the major criticism levied at ChatGPT since it burst onto the scene almost a whole year ago.

But let’s move on.

Challenge 2: Recommend the best dog-friendly hotels in Northumberland (Bing)

I asked Bing to find me the best dog-friendly hotels in Northumberland. After all, where better to go for a good old yomping holiday in the countryside with a four-legged pal? But where to rest our weary legs and paws at the end of the day? Let’s find out if accommodation-planning don’t mean a thing if it ain’t got that Bing:

Just the three options from Bing, but that’s OK. And looking at the locations, I know they are lovely areas of the county with lots of sprawling countryside so Bing scores a few brownie points there. Let’s take a closer look at the first option (which shall remain unidentified, for reasons that will become clear shortly).

The first thing that struck me as odd was the supposed Tripadvisor rating. As we probably all know, Tripadvisor’s star rating system is out of 5 – not 10. But hey-ho. If that’s the only hiccup here then it’s hats off to Bing.

And then I take a look at the actual Tripadvisor reviews:

It’s never wise to make a judgement call based on the first review you see though. What about the ones that follow?

• 1 star….avoid if possible
• 1 star…PLEASE LOOK ELSEWHERE FOR ACCOMMODATION
• 1 star….DON'T STAY you have Been Warned

You get the picture.

So less than ideal really. The discrepancy between Bing’s recommendation and actual customer feedback serves as a stark reminder of the limitations of AI in understanding the nuances of real user experiences.

Challenge 3: Plan and book a trip to Stockholm (Bard)

Before we get into the challenge itself, let’s learn a little about one of Bard’s latest rollouts. Bard Extensions can link up with all your Google bits and bobs - Gmail, Google Drive, Google Maps, Docs, YouTube, and – more specifically for this experiment - Google Flights and Hotels. The point being you can use Bard as the ultimate travel planner, tour guide, travel agent and more.

Now, bearing in mind Bard once recommended an 8-hour flight from Southampton to Newcastle via Belfast, I didn’t have high hopes.

But as promised, Bard was able to dig into my emails and find the preferred dates to travel to Stockholm as a starting point for planning the trip. So far, so impressive.

I next asked Bard to find me suitable, real-time flights but for reasons not entirely clear, it proposed flights departing on 21^st November and returning on 16^th December which was a bit odd. However, once I steered Bard back on the right path, the flights it suggested seemed to be a good selection, although booking an actual flight does still involve you navigating away from Bard, so it’s not *fully* all-inclusive – at least, not yet.

Next up, I asked it to recommend some hotels. Once again, this all seemed fine, with Bard offering up a good assortment via Google Hotels, complete with star rating, price per night, photo and a brief description. As with the flights, the bookings themselves cannot be made within Bard.

So I’ve got the dates, the flight and a list of hotels. What about food? I asked Bard to recommend some vegetarian restaurants in the city:

I cross-referenced these with Google and they all appear to be good suggestions (although I can’t say I’m all that taken by that last one).

I continued with my Bard adventure, discovering cute little boutique shops, mapping out how to get from one location to another, learning some useful tips regarding social etiquette along the way. I even picked up a few nifty phrases too: (AI kommer att förstöra oss alla).

Have I changed my mind about chatbot travel?

Despite the mixed bag of results, this experiment was an eye-opener in many respects. As someone who is both fascinated by and dubious about AI in equal measure, I did wonder what my angle might be for this article – was I willing the chatbots to get it oh-so wrong or can I actually see a time when I might throw caution to the wind and hand over my hours of travel planning to Bing, Bard and co?

Ever the luddite, I’m not quite ready to make the leap. While chatbots might serve as useful starting points for planning trips, they lack the nuanced understanding and insider knowledge that can make or break a holiday. ChatGPT, for example, has never felt the wind in its face skiing down the slopes of Verbier, nor has Bing or Bard experienced the sights, sounds and smells of a street market in Marrakech, which can only come from human experience. This underscores the importance of E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) when planning a trip, especially if you’re looking for an altogether more bespoke experience.

Bard, with its integration of various Google services, shows a glimmer of a more cohesive travel planning experience, yet its current capabilities fall short of providing a truly all-inclusive solution.

ChatGPT can generate an acceptable basic walking tour but falters when it comes to providing details crucial for specific needs, such as accessibility.

While promising, Bing's recommendations for dog-friendly hotels highlighted a disconnect between algorithmic suggestions and deeper real-world experiences.

So, in conclusion, we’re not *quite* there just yet. But the pace at which AI is evolving, I envisage a rewrite of this blog might be necessary in the coming months. Who knows? One day I may even end up ditching my beloved travel folder and allow ChatGPT to sort out my trip to Rome instead.