Unless you have been comatose for the past eight months, you have heard about ChatGPT. In fact, no technology product has ever seen such rapid mass adoption. ChatGPT reached 100 million users about two months after its launch late last year. The next closest product – TikTok – took almost five times longer to reach that milestone.
ChatGPT and its cousin, Google Bard, are the most public faces of a form of Artificial Intelligence (AI) called Generative AI. I have been watching the space since late last year when ChatGPT was launched by OpenAI, but I intentionally avoided writing about it until now. It seemed like almost every technology product in the industry was suddenly claiming to be on the AI bandwagon. But many of the claims were dubious, often implying the use of Generative AI by products built on much earlier generations of AI (or no AI at all).
Integration of new toolsets like ChatGPT into real applications takes time. I have watched over the past eight months as various companies have tried out the new platforms and explored potential use cases. Before writing about it, though, I wanted to see a few products that had demonstrably and successfully integrated Generative AI into their core platforms and shown the ability to deliver business value.
That is just now starting to happen. I am not going to cite specific products this week, as I have only seen a few in much detail, and I know there are more. The number of commercially available products in hospitality that have truly incorporated Generative AI is still small, but it is growing. I plan to survey the field more thoroughly over the next couple of months. If your company has a Generative AI product, and we have not yet talked about it, or even if you just know of one that I might not be aware of, please reach out to me so I can include it in my review.
There is a lot of reason to be paying close attention to Generative AI, with more and more use cases coming to the attention of hoteliers, so this week I will review the fundamentals of Generative AI (and what distinguishes it from other forms of AI) and outline some of the use cases that have been identified, some of which are being actively worked on. I will close with some comments on the challenges associated with Generative AI.
Fundamentals of AI and Generative AI
The broad field of artificial intelligence is typically defined as computers that can simulate human intelligence. There are many levels of AI, from simply being able to answer a preprogrammed set of questions, to understanding and translating languages, to examining massive data sets looking for interesting patterns too obscure for a human to ever find. You can find an excellent overview of the various types of AI in this recent article by David Chestler, but I am going to focus specifically on Generative AI.
Generative AI refers to the creation of new content using models based on the way humans create content. Most often (so far), this has taken the form of the written word – but it can also be photos, videos, or audio. It can also take the form of data, for example analyzing historical data and predicting the next number in sequence (such as occupied rooms on a future date).
These are all things humans can do with different levels of competence, and things computers can now do too – also with varying levels of competence, although often much faster. The competence level matters: if a typical human can do something 99% accurately, and a computer can do at least as well and possibly faster, then the human (if they are performing no other role) becomes unnecessary. If the computer is less accurate, it becomes less useful. But less useful does not necessarily mean useless: when no human is available (such as when half your work force leaves the industry because of a pandemic), lower accuracy will often be a better option than leaving business needs completely unmet.
Generative AI is built on what are called foundational models. Such models can, for example, read billions of English-language documents and then describe the language based on the usage patterns it finds, distinguishing common ones from less common ones (which may represent errors, regional variations, effects of educational levels, life experiences, or other factors). It does not know “right” usage from “wrong” usage, but rather recognizes more typical usage vs. less typical. It can then reconstruct the data on which it is trained to generate outputs closely resembling that data. A degree of randomness can be added to create variations in the outputs, just as a human answering the same question multiple times will likely use different verbiage, even if the meaning might be essentially the same. This modeling process is handled by Variational Autoencoders (VAEs).
The models are then trained using an adversarial process. One model (the “generator”) creates outputs, while another one (the “distinguisher”) tries to determine which responses were model-generated vs. human. Millions or billions of tests are conducted, and the generator learns which of its outputs pass as “human;” it then adjusts future outputs to suppress more of the ones that do not and promote more of the ones that do. The distinguisher learns as well, by comparing its predictions (human vs. machine-generated) with the true source. Over time, both models improve, and the generator acts more and more “human.” These are referred to as Generative Adversarial Networks (GANs).
The models can be improved by training the agent within a specific environment (such as a hotel chatbot) or constraints (such as avoiding sensitive topics). Reward functions provide feedback to the generative model that reflects the quality of the output, so that the model has a “metric” of how it is doing and can improve. This allows for a balance between generating diverse and novel outputs, and optimizing outputs based on rewards. This is called Reinforcement Learning.
Combined, these and other Generative AI techniques can enable agents to generate human-like responses in many situations. Like humans, the information may not always be correct. Unlike humans, AI has not reached the point where it can reason, such as by detecting obvious nonsense. It may generate what are called hallucinations – in writing these may manifest as false or even impossible statements. The better the training, the less this should happen. Unfortunately, there are as yet no common metrics for measuring this.
In the past couple of months, I have seen a few impressive implementations of Generative AI in hotel chatbots, where the desired behavior, tone, constraints, and specialized knowledge are entered in plain text, and the chatbot uses those instructions to correctly respond to guest requests. It is much like training a human, except the chatbot does not forget its instructions as often as humans do. The specialized knowledge can include lots of important details about a hotel, such as restaurant operating hours, equipment options in the gym, check-out time, or how to get to nearby sights.
What I have not seen (but only because I have not played with them enough) is what kind of hallucinations these chatbots come up with, how often, and what effect they might have on an interaction. It is one thing to provide slightly out-of-date directions to a nearby restaurant because of a road detour, but quite something else to start hurling racial or religious epithets or to provide inaccurate life-safety information. And it matters how often such issues arise. If it is one time in a million, it will beat the accuracy of almost any human response. But if it is one time in 100 and causes major offense to a guest, then maybe not.
Chatbots are among the more challenging applications because they are designed to ultimately run unsupervised. If I were running an economy or midmarket hotel, I would probably try one out, but I would still review the prompts and answers on a regular basis to assess its performance. In a luxury hotel, I might let the chatbot generate responses, but then only send them after human review and approval (maybe with exceptions for simple questions like the Wi-Fi password). Based on today’s technology, though, in neither case would I be willing to turn it loose without some monitoring. That day is probably coming, but it is not here yet.
Hotel Applications for Generative AI
There are many potential applications of Generative AI in hotels. Some of them are relatively obvious, others less so. A few may require other forms of AI as well, but still depend on Generative AI for the final output. I have listed some useful applications below, starting with lower-risk, easy-to-implement ones and moving up toward the more complex or higher-risk ones.
- Generating marketing copy for things like room types, restaurants, or recreational facilities. Many hotel staff get writers’ block trying to draft this type of material from scratch, but it is important if you want to sell, and is something Generative AI does well. A staff member can easily review the result and make any necessary tweaks. Generative AI can (at least in concept) also handle different languages, meaning not just translating text literally, but writing a description in a marketing style that is appropriate for each language.
- Generating photos, graphics, or music for the hotel website, app, or digital signage. Why schedule a professional photo shoot or pay royalties for stock photos if all you need is a generic photo of a hotel front desk that a tool like DALL-E can conjure up to your specifications?
- Writing job descriptions. Many hiring managers and even some human resources professionals have a really difficult time doing this, but Generative AI can put together an excellent first draft.
- Creating interview questions tailored to a job and specific candidate, based on their resume and the job requirements.
- Generating draft (or even final) responses to online reviews. Not only can this save time, but it can avoid the overuse of templates, which may end up posting so many identical responses that the reader will immediately recognize them as impersonal boilerplate. The randomness inherent in Generative AI can vary the wording enough to make responses look more human.
- Creating regular content for posting on social media. Depending on the training database, this could for example automatically highlight local events, destination information, or feedback from happy guests.
- Generating marketing copy and FAQ answers for a hotel based on guest reviews. The collection of all reviews ever written about a hotel includes a wealth of information that matters to real-life guests: things they like or dislike about the hotel (often by type of guest), reactions to specific facilities or services, and so forth. If the reviews are fed into a Generative AI model, you can refine your marketing descriptions even further, making sure they highlight the things that guests like, set expectations properly to avoid disappointments, and address questions or concerns that past reviewers have highlighted. With the right website functionality, you could even personalize things like room type or restaurant descriptions, for example emphasizing the business features of a room for a business traveler.
- Responding to social media posts. I would not yet let Generative AI actually post responses, but it could very useful in preparing responses for a human to review and post.
- Generating personalized training curriculum and materials for staff members, based on their job requirements, prior experience, and prior education or training.
- Generating personalized virtual tours for guests or meeting planners that will allow them to see the relevant parts of the hotel and to get immersive audio-visual responses to questions like “What do your beach cabanas look like?” or “Can you show me the Grand Ballroom?”
- Generating personalized marketing messages for each guest, to send via text or email or other means, suggesting activities, tours, events, restaurants, or retail stores that are selected based on the guest history and profile.
- Generating personalized responses to inquiries on a job board, answering questions about the facility, roles, key staff, policies, and the like.
- Responding via AI chatbot to text prompts (or even voice prompts). Chatbots can provide booking assistance (“which hotel is closest to The Louvre?”, “do you have any rooms for a family of five?”), information before or during the stay (“what are the pool hours?”, “how much is a taxi from the airport?”), or fulfill service requests (“please send me more towels”). The better chatbots that have already incorporated Generative AI are already approaching and sometimes even exceeding the capabilities of human staff to respond – and they are still in their early days; they will improve.
- Personalizing instructions for delivery robots or mingling robots to interact more engagingly with specific guests.
Using Generative AI in the ways described above requires technical skills that hotels will rarely have in-house, so they will be dependent on commercial products and/or customization by vendors. Today, however, all the products are new, and there are few metrics for comparing their relative quality. Extensive testing is needed, especially when AI results will be communicated to guests or the public without human review. This testing needs to be done with multiple people, preferably from the target user group, to ensure that responses are universally understandable, relevant, and accurate. It needs to be done over multiple types of hotels. For now, a key part of the evaluation process should be an in-depth discussion with staff that are maintaining and monitoring the product in a comparable hotel.
Training a Generative AI model requires data that encompasses the entire experience that the model is intended to address. But in a hotel environment this means getting data about everything in the hotel, and the siloed nature of most hotels’ information systems can make this very difficult. It can be much easier to deploy chatbots and similar tools if the hotel first creates a centralized source of truth that covers everything that might be asked. A chatbot will not provide correct operating hours or menu options for a restaurant if they are only stored in the head of the food and beverage manager or the point-of-sale system.
While some of the above applications are essentially standalone, others may require integration with hotel systems, or may be able to do more if they are. A chatbot that has access to the history and preferences of a specific customer can generate personalized responses that are not possible without it. A chatbot that is connected to guest-room systems can answer requests to change the TV channel or temperature, dim the lights, or set the do-not-disturb signal. Such integrations exist for most systems, but hotels may not already have them and they may be expensive to acquire.
The biggest issue with any form of AI in hotels, of course, is the potential loss of the human touch. This is a more critical issue for luxury hotels and for guests who value that human touch (which not all do, even at luxury hotels). This is not a black or white issue but has shades of gray. Almost every guest will find at least some AI applications as satisfying (or more satisfying) than a human interaction, but for some guests it may be just a handful while for others it may be most. In luxury hotels the right answer is to meet the guests wherever they want to meet you – at the front desk, on the phone, on a chatbot via text, on a chatbot via voice response. But if you offer an AI option, it needs to be as good as if not better than staff – meaning that just as with staff, it needs to know when to escalate a question to someone who has more information.
At an economy or select service hotel, this is much less of an issue. With understaffing and high turnover, it is hard to get a person, even via chat, who knows the answers to more than the most basic questions. An AI chatbot may not be perfect, but it essentially adds a 24x7 staff member whose sole job is to answer customer questions, and who can do it reasonably well. I stay at many select service hotels, particularly outside major cities, but I am tired of asking the front desk clerk to recommend a nice restaurant and being directed to a McDonald’s or Denny’s. These are perhaps considered “nice restaurants” by the typical 19-year-old working the front desk, but to me, while I am not knocking them, they fall a few notches below “nice.” I suspect most guests would agree. I like the human touch, but in this case, I will take the AI hands-down.
Even today’s first-generation Generative AI tools can address issues like this. And while I would deploy them with caution in luxury hotels, I see few downsides for select service and below, particularly if they suffer from staff shortages or high turnover. I plan to return with more specifics, and some vendors to consider, in a future installment. Stay tuned!