An experiment with ChatGPT-Powered Customer Support
By Jim Kogler (**edited by ChatGPT, Morgan, and Dan)
I have been approached by multiple customers suggesting the idea of developing a support system by training an OpenAI/Chat-GPT system on our MAK documentation set. They believed it could potentially provide some level of autonomous help desk for our customers. Intrigued by the possibility of automating support, I decided to give it a try. However, I must say it hasn't gone well, and as a result, we won't be releasing it just yet.
Learn how AI can sometimes be misleading...
ChatGPT, a large language model developed by OpenAI, has been trained on diverse text sources, granting it a broad knowledge base and the ability to generate coherent responses. OpenAI offers an API that allows us to process specific data, such as our MAK ONE User Manuals, and integrate that information into the model. (There's more to it, so for a comprehensive understanding of how ChatGPT works, you might want to explore other resources.)
After spending hours researching the process, I found several websites and sought advice from ChatGPT itself. Armed with this newfound knowledge, I managed to write code to efficiently parse through our extensive documentation. Although the processing took a significant amount of time, I eventually created a functional web page where I could enter prompts and expect satisfactory outcomes. For this article, I'll refer to this web interface as the Bot.
Excited about the results, I eagerly fed the Bot an email from a customer asking about how to configure a weapon system in VR-Forces. To my surprise, the response I received was detailed and well-written but included some incorrect elements. It turned out that about 20% of the information provided was fictional, and there were steps that didn't actually exist! The most concerning part was that I didn't realize it was incorrect. This discrepancy became apparent only when I delved deeper into it, and needless to say, I didn’t send that response to the customer.
Maybe I expected too much by trying to copy and paste the support email as a prompt. Prompt engineering plays a crucial role in obtaining desired responses from ChatGPT. It involves constructing the initial instructions or queries in a way that encourages the model to generate the desired information or perform a specific task.
To better understand how this all works, I spent several additional hours asking the Bot questions, each phrased to hopefully get a useful answer. The Bot provided partially correct answers but also interjected incorrect information.
When I asked, “Does VR-Forces Support Behavior Trees?” The simple answer is “yes,” but the Bot gave a verbose answer filled with details, many of which were related to capabilities of VR-Forces but it added in details that are definitely not part of VR-Forces.
When I asked how to task a plane to take off and fly to an airport “in VR-Forces”, I received a correct answer about the preflight checkout process in ‘the real world’, not an answer about the ‘simulation conditions’ that must be met before takeoff.
When I asked about the meaning of "MAK" in MAK Technologies, and it provided a plausible but false explanation, potentially influenced by my earlier support questions related to maritime modeling.
As you can see, releasing an AI-powered customer support system on our customers at this point would be counterproductive. Our engineering team is safe — no AI is going to hold them back from customer support, this year at least.
Before closing, I would like to offer some higher-level thoughts about the use of AI and LLMs like ChatGPT.
The fundamental problem with this AI isn't just that it makes things up. The fundamental problem is that it makes things up in such a convincing way that even experts on the topic sometimes have a hard time fully understanding what went wrong. Humans often equate 'well-written' with 'knowledgeable.' That is the problem. Providing a ChatGPT version of MAK support would not only lead customers astray but would also do so in ways that are exceedingly frustrating and time-consuming. As ChatGPT says, “My purpose is to provide information and assist with generating human-like text based on the input I receive and the data I've been trained on.” Its goal is not to provide knowledgeable answers based on the data it has been trained on.
**Jim's original blog post was three pages long, so we used ChatGPT to shorten it to its main points. The prompts for ChatGPT were: “I am going to copy/paste an article. Please condense it into half the length.” It provided a summary of the original blog but in the third person. The next prompt was: “Ok, now rewrite that in first person.” Then we argued a bit about how good of a job ChatGPT did and we spent a fair amount of time editing to arrive at the blog post that you see above. AI has its place, but be careful.
