Almost one year after ChatGPT was released, OpenAI continue to surprise us with new features and capabilities of its platform.
The last release note from OpenAI provides indeed a new cool feature to create your own customized version of ChatGPT, called GPTs.
GPTs are a way to increase even more your productivity. If you have been a user of ChatGPT, you might have already identified a long list of use cases that help you in your day to day activities : it can be optimizing your tweets, help brainstorm article ideas, generate variations of a sales pitch, review your code or help you learn Esperanto. GPTs come to play to help you make a "specialized" version of ChatGPT, personalized for your needs, and very good at one specific thing, instead of being very generalist.
This opens a new range of possibilities. In this article, we will create, step by step, a new customized GPT than discuss some limitations of this new feature.
Pre-requisites
The pre-requisites are quite simple, you need :
- A ChatGPT Plus license or Enterprise account (this means, basically, you can’t create GPTs using a free account).
- And an idea of a new GPT (think of your main use cases, or ask ChatGPT to help you brainstorm). Here some use cases shared by OpenAI to give some examples:

The use case
For this article, I wanted to work on a real use case, coming from a pain point I faced in the past.
In a previous professional experience, I had to manage customer facing GDPR topics. The GDPR, or General Data Protection Regulation, is a law in the European Union (EU) about data protection and privacy. It aims at giving people more control over their personal information and how it’s used. Yet, its implementation can be tricky. The text law is 150+ pages, and is quite technical. As a technical guy, it might be challenging to explore legal and regulatory topics. At that point of time, ChatGPT didn’t exist, so I had to upskill and get the help of an external legal expert to support me in the journey (which also increased costs), till I had more autonomy. But even then I always faced new topics like : addressing data cross-border transfers, managing security incidents, updating the privacy policy or making a data protection impact assessment.
So I wanted to build, what I didn’t had at that point of time, a GDPR smart companion, available 24/7 that can help me with this topic.
Of course, this is just a use case, but it is an interesting one in my opinion, as it addresses a clear need and have an obvious business case.
When you use ChatGPT, and ask about GDPR, ChatGPT has consumed "all the internet" (not exactly, but good part of it), so it does have some accurate information, but also some less accurate, and less precise. GPTs in this case will allow me to constrain the knowledge using specific options on GPTs that we will see in the guide, as well as tailor the language tone and some other tweaking.
You can apply the same logic afterwards and adapt the use case to your own needs, typically, if you want to build a Chatbot based on a knowledge database, a series of article, on a set of PDF, or similar use cases, you can follow the same steps.
Let’s build it
To create your GPT, you will need to use the editor provided by OpenAI.
The editor is comprised of 2 blocks : The left one, with 2 tabs, for Creation and Configuration, and the right one for the Preview and the test.
You can use the following link to open it (check pre-requisite, you need ChatGPT Plus or Enterprise licnese) : https://chat.openai.com/gpts/editor
The editor looks like this:

We will now follow the steps as suggested in the editor guide. I start with the following prompt "I want to create a new GPT that will help answer GDPR related questions. It can be general questions related to GDPR (like the main principles), or specific to user situations (like, for a given security incident, is a notification required)."
When giving those instructions, you should be as specific as possible. My prompt however might not be perfect, but we will have the possibility to make it better as we progress.
Now, following the description I provided, I am being offered a name suggestion :

I do not like the name "GDPR Guide", so I will suggest another one : "My GDPR companion". The new GPT will get that name, and a logo will be suggested:

I find the logo not that great, as GDPR is a European law, I think the logo should reflect that. However it is interesting to see that I got a logo representing security and privacy which are major pillars of GDPR. We can also see on the Preview part the logo (generated), a short description (generated), and examples of questions, called conversation starters, (generated as well). This helps you get a first sense of what the GPT will look like, and what questions it is expected to answer.
So, as I want to adjust the logo, I will give additional instructions requesting a revision "GDPR is a European law. I think the logo should reflect that."

Now that logo is better in my opinion. The editor took my comments into consideration and provided a Logo with more reference to European flag (the twelve yellow stars and the blue background).
Then, the editor is requesting more information to refine its behavior.

This is a real addition to classic ChatGPT, because it actually, proactively, asks you questions and guide you to ensure it provides the right tone, and is well adapted to your needs. I provided the answers to those questions, and got confirmation that we are all set. It will also rephrase your inputs in a synthetic manner to show that it "understands" the input you shared.

As you can see, I also requested to add a kind of disclaimer. In some fields like legal advice or financial advice, such disclaimers are needed to ensure that the user clearly understands that such tool is only here to help, but not to take the accountability.
Now, we have our GPT setup, before testing, we will need to add some configurations in the configuration tab:

Multiple fields are not exactly as I would like them to be. So I will adjust some of them, including
- The Description: I want a shorter one,
- The instructions: refine some instructions,
- The Conversation Starters: to add one conversation started on how to start with GDPR,
- and I will add the knowledge base as well. For the knowledge base I recommend adding a (plain) Text file. I tried with a PDF, but the results were not very accurate (check Limitations). For your reference, the raw source I used is located here. (You can upload multiple files).
- There are also some advanced configuration features allowing interactions with API. These are also interesting to explore, but on this article, I am limiting this guide to a "No coding" approach.

Now it is time to test!
Starting with one simple one, requesting information from a specific article in the law. Note that I can just give quick prompts. No need to say it is about GDPR or to get the data from the Knowledge set:

And here is another example as well, this one for a specific case. The answer is very well explained with the exact references to the right law articles (despite it being very long and complex, and despite my prompt not using the exact correct technical terms):

You can continue testing your GPT. If some answers does not suit you, you can go back to the Create tab and add some comments and feedback so it can be made better. You can also refine the configuration by tweaking the instructions section or updating the Knowledge base and other parameters.
Once you reach the right configuration, you can now save and publish your GPT, you use the Save button in the upper right part of the editor, and you will have some options. In my case I will chose to make it public:

Note that for Enterprise users, an option to publish the GPT at the company level is also available.
Once saved, I am redirected to the ChatGPT screen, with my GPT selected,

I can now share the link of by GPT, here is the link of the GPT I created for this test.
OpenAI mentioned that monetization will be available for some GPTs. It is not yet clear how this will work. It should be announced in the upcoming weeks.
Limitations
GPTs come also with limitations. It is still a new tool and should improve over time, but for now, I observed the following limitations:
- Hallucinations : GPTs are powered by GPT-4. They inherit a good part from GPT-4, including hallucinations. This does not change with GPTs.
- Accuracy when using PDF is Knowledge base: Regarding the feature of using the "knowledge" feature, allowing to upload various documents, I tried initially with a PDF, but the results showed some limitations, probably related to scraping issues. I retried with a plan text format, and this seems to work better.
- Accuracy for long or complex instructions: When given long instructions, GPT might take into account some, and not the others, or might not apply all the instructions to all answers. Typically, in my tests, it didn’t mention the disclaimer in all answers.
- Response times : The response times are much longer than GPT-4. For some prompts, it took more than 1 minute. This of course depends on many factors and can be related to the recent release. But it looks like, the bigger the Knowledge base is, the longer the response times. Indeed, each time a prompt is sent, the GPT seems to access the Knowledge.
On a related note, and even if this is not a limitation per se, this new feature will de facto make some plugins and third party tools obsolete, typically, some plugins that manage PDFs (like AskYourPDF), or some applications that generate Chatbots based on Knowledge databases.
Closing words
For now, GPTs looks like a cool new feature, but it is more "incremental" to ChatGPT than really disruptive.
If you manage to prompt well in ChatGPT or in the Playground (using the System field), meaning, you provide clear context, instructions and tone, you probably do not need this new feature, unless, you want to couple it with some more advanced feature.
What is really new, is the capability to share your Gpt and to monetize it. These cannot be tested right now, and we should expect it to be available in the next few weeks.
If you created some GPTs, feel free to share your feedback as well as the limitations you faced!
If you created some GPTs, feel free to share your thoughts and the limitations you faced!
If you want to learn about AI, join my course!