ChatGPT and artificial intelligence in general is taking the world by storm. People from every industry, including art, business, and webdev are now using some form of AI in their every day workflows. One of the most popular releases in the AI landscape over the past year is ChatGPT. With it, came a easy to use interface that meant that it became easy to interact with a large language model (LLM).
This allowed virtually anybody to try out artificial intelligence for themselves. And, not surprisingly, it took the WordPress community by storm.
ChatGPT was trained on a massive repository of information from the Internet. Some of that information contained WordPress content, code, and more, which means that when you ask the chatbot a question about code, popular, plugins, and more, it gives an answer that is typically correct.
In this article, I want to explore the intersection of the every day usage of AI and WordPress workflows. We will be taking a look at all angles, including security, ethics, and the underlying theory behind ChatGPT.
I am a WordPress developer, turned AI developer and researcher. While I still run the Isotropic studio, I’m a lot more focused on a newer venture called CodeWP – a platform that uses AI to generate code snippets for WordPress creators. On the platform, there is a chat bot very similar to ChatGPT; essentially, this article speaks from the perspective of somebody who creates artificial intelligence, and is also the end-user as a WordPress developer/agency.
ChatGPT Use Cases In WordPress
Due to the flexibility of the chatbot-method of interacting with artificial intelligence, we’ve seen a lot of use cases pop up in the WordPress community.
If you’re unsure what a specific function in WordPress does, ChatGPT is great at creating explanations about this. However, it’s limited to some of the more popular WordPress products. That’s because it doesn’t have specific training on all plugins/themes out there.
Creating general code snippets is another use case.
Typically, it will also give an explanation of what the code actually does, which can be pretty helpful when it comes to learning. At the same time, this can be annoying if you’re looking to generate many snippets at once. Something else worth noting is that in the above example, we asked for a shortcode, and it only gave us “loose” PHP code.
This is an issue because we can simply wrap it in a function, and then create a shortcode from that, but if I wasn’t a WordPress developer, this wouldn’t be immediately clear.
With that said, this shows benefit of having this type “chat based” of interface. All I need to do is ask again for it to create a shortcode, and it will.
Many are using ChatGPT to create content, both for placeholders, and for production. I discuss this in more detail in a later section of this article.
Everyday Admin Tasks
ChatGPT excels it every day administrative tasks, that every agency, owner, freelancer, and developer need to do. For example, creating email replies is incredibly easy. All you need to do is paste in the email, and you’ll get a great response for it. Categorization, ideation, and more, is also pretty easy here.
Making Plugins (not really)
When ChatGPT first went viral in the WordPress community, one of the biggest show tricks was creating an entire plugin using the bot. I suppose that in a basic term, you could create a WordPress plugin, but this is essentially a code snippet in itself. There’s no real way to build out an entirely functional plugin with multiple files, object, oriented programming, dependencies, and so forth. Instead, you’ll find yourself creating a boilerplate file which you can then zip and install into a WordPress website. However, this is definitely much more of a show trick then something that’s actually useful to professionals.
How Does ChatGPT Actually Work?
To understand the best use cases for this powerful free platform, it’s important to understand how it works. In this section, we’re going to try and break down how the platform actually goes from receiving your prompt to outputting code, explanations, or content.
How it was trained
Training the underlying AI model that powers ChatGPT, also known as InstructGPT, follows a three-step process. First, they took many prompts and created the desired output using human labelers. This was then used to fine-tune GPT 3.5. From there, they take one single prompt and multiple outputs. A human then ranks them from best to worst. The best ones are used to train the model, and what the expected behavior is. The final step further tailors the generations to be accurate, and also valuable.
How it works
There’s no official paper or information by OpenAI on how exactly ChatGPT works, but most people believe this:
it uses InstructGPT with minor variation and training
the user submits a prompt
initially, a response is generated based on the training that it received (mentioned above)
it will refer to previous generations and outputs using embedding for additional context
Where Does ChatGPT Excel?
For anything related to WordPress on a surface level, ChatGPT works really well. For example, if you have an Elementor question about how to do a specific thing in WordPress, you’ll typically get a great step-by-step answer.
If you’re looking for basic code, snippets, examples, and explanations, the platform will typically perform pretty well.
If you want to create any type of content, including placeholder content, based off of themes, this is a great platform to do that as well. For example, you can prompt for a call to action based on the theme of a fencing company, and it will create a great placeholder for your project.
If you need to do any brainstorming, leveraging a LLM is always helpful. For example, I can ask, “give me five advanced ideas on how to improve the performance of my WordPress website – I have 500,000 individual posts”. From the initial ideas, I can then go into my own research to get to my goal.
I can also use this to create outlines, post titles, and other administrative tasks like responding to emails.
We use this pretty consistently to ideate on research topics. For example, if we have a big problem, but don’t know where to start in Google or stack overflow, we can explain the issue to ChatGPT, and ask for suggestions on what to look into. It’s good at this because it’s trained on so much information in the Internet, it will always come up with something that we’ve never thought about before, but is typically related to our issue.
You’ll notice that we don’t really use this to do any heavy lifting when it comes to offering complete solutions, snippets and more. Let’s talk about why it’s not smart to use this platform as the end all be all.
Now, this is the part of the article where we get into the limitations and negatives of using ChatGPT.
The first major thing to cover is something called hallucinations. This is a general AI concept that is very present in some responses created by ChatGPT.
It seems that this phenomenon occurs more often when the users asking a specific question about a specific topic that is more technical, and the model doesn’t have a ton of training data to reference. At that point, the preferred behavior would be a response like:
However, sometimes it will make up information that seems real, but definitely is not. And if you’re not completely familiar, in the topic that you are discussing, it can lead to some pretty major issues.
For example, say you asked ChatGPT to create a code snippet that does a specific feature for your WordPress website. It hallucinates, and gives you a good looking snippet. However, completely wrong. If you’re a professional developer, chances are you’ll be able to catch this before testing. However, if you’re an every day WordPress user, and go and install it on a production website (confident due to the fact that this is AI, and should ‘always’ be right), you can brick it, causing the white screen of death and no clear path forward.
ChatGPT’s OpenAI and many AI ventures are very focused in mitigating this on ChatGPT, and AI models in general. But, this is a big issue, with no clear way to mitigate, unless you’re able fine tune for a specific use case, incorporate ML, or otherwise have a poke into the internals of the tool (impossible with ChatGPT).
Sometimes, It’s Wrong
AI won’t always be right – but many users don’t understand this, leading to some tricky situations. The wrong facts being displayed on a webpage because somebody didn’t check the content generated, errors on WordPress websites, because code was installed without being tested, the list goes on…
With the hallucinations, this issue can be exacerbated by the fact that it can be very convincing Lee wrong.
Here’s an example, using the code snippet that I had generated for the “Writing Code” use case above. I asked, “how do you install this on the website”. I expected it to tell me that I needed to first create a shortcode, and then install the code in the site.
At first glance, this looks correct, but if you install the code directly on the website, it will cause issues (echoing directly into the website).
We’ve also noticed that there can be security issues with snippet generations as well. This can be mitigated by prompting accurately, but if you don’t know what you’re looking for in the first place, sometimes it can slip by you.
It’s worth noting one of AI’s golden rules applicable when using ChatGPT – low quality prompts will result in low quality generation, and vice versa, high-quality prompts, will result in high-quality generation.
Suggested reading: https://docs.cohere.ai/docs/prompt-engineering
The UI is lacking
If you’re a WordPress professional, and truly want to integrate this tool into your workflow, you’ll realize that there’s a lot of user interface elements either missing or poorly designed.
The platform is missing a search feature, as well as other categorization utilities that would allow you to reference previous generations easily. The overall design and usage is good for small sessions but if you have a large or multiple chat logs, you can quickly get unmanageable.
Realistically, this is probably by design, as this interface is being used by OpenAI to train future models, and isn’t really intended to be used in the professional manner that some expect.
Capacity and Throttling
This is a massive problem, that’s difficult to fix for the company behind the product. As with anything AI, this requires a bunch of processing power to operate. In fact, ‘According to OpenAI co-founder and CEO Sam Altman, ChatGPT’s operating expenses are “eye-watering”‘ (source).
Because it’s a completely free platform that everybody and their mother is using, there are major capacity issues simply because there isn’t enough power for everybody to use it all at once. During the business day, you’ll typically notice outages and bugs because of this.
An example of a capacity error
If you’re lucky enough to load the platform and enter a prompt, you’ll also notice that the answers “render” (stream is the proper term) very slow. That’s because OpenAI has been forced to limit and throttle the response rate to allow more users to access the platform at once.
We were lucky enough to have access to the first iteration of the platform, before it went main stream. Answers would render in seconds, there would be no limitations, and the models were much more raw at that point. Now, there is a massive difference, and it’s only getting slower as the popularity increases. A potential fix to this is a pro plan (which is in the works), but I wouldn’t expect the free plan to get any better anytime soon. Instead, you’ll likely need to end up paying for the pro plan to get performance like the early days.
Response Character Limits
This one isn’t the fault of ChatGPT, because it’s present in pretty much every single OpenAI product: the dreaded token or character limit.
With this offering in particular, there’s a limitation to how much content it can reference in the past, as well as how much content it can actually generate. Therefore, you’ll typically see a drop off if you’re trying to refer to something that occurred a while ago in your conversation. Additionally, if you’re generating large code snippets, you may need to chunk it into smaller generations, or risk having your creation cut off at a “random” point.
Something to note is that you can instruct it to continue the previous generation, but if you’re looking to create a massive amount of code, you’ll only be able to reference back to the previous 4000 tokens generated, meaning that after you get away from that limit, the generation may not be able to reference previous variables or functions that you’ve created, essentially creating code that won’t work. (this is the major reason I said, “not really”, after “creating plugins” in the use case section)
Good at everything, great at nothing
LLM General Purpose Transformers (GPT) are typically good at pretty much everything you throw at them at a surface level. As mentioned before, these models were trained on massive repository as a data script from the Internet, and know a lot of different things. But, in many cases, specifically development, it’s good at everything but great at nothing.
That’s because to create a model that generates accurate code, and truly understands specific plugins and products, you’ll need to fine-tune it based on different examples, as well as the entire code base of the tool (context of generating code).
Fine-tuning, embedding, and other methods are common place in scenario specific products (like CodeWP), but it’s simply not feasible with ChatGPT, because it needs to be good at everything. In the future, if there were an update to introduce, fine-tuning and personalization, based on your specific generation cases and rankings, then it would be a completely different story. But I don’t really see that happening, especially given that if you’re looking for a specific used case, there’s probably already a tool out there that does it.
It’s Not Up To Date
A limitation of all large language models, and GPT is that it doesn’t have access to up-to-date information. Instead, it’s drawing on the knowledge it received from its training data. Training data has “cutoffs” as models are trained in one go.
ChatGPT has a cut off date of somewhere in 2021-2022. Therefore, if you’re asking it to create code, explain a function, or do some other thing for WordPress, or a WordPress specific product that has, since been updated, it will very likely be wrong, hallucinate, or refuse to answer your question.
There’s another issue where ChatGPT is also biased towards information that it knows the most about. For example, there’s a lot more code on the Internet discussing Chart.js v3 or Chart.js v2, as opposed to the most recent version of 4.0+.
Therefore, if you ask to generate code, using this library, you will most likely get a response that uses depreciated methods. When you go to install it, and use the most recent version of the library, it simply won’t work because it’s outdated.
This issue is present for pretty much every piece of technology, library, platform, or anything else that has undergone a major version upgrade with breaking changes.
If you were to create your own model oriented specifically towards this, you would simply train it on the newest version, and then ensure that it doesn’t use code from version two or three (this is relatively easy when using positive negative pairings). But, this is definitely outside of the scope of this offering.
WordPress Specific – Affiliate Favoritism
Because of the underlying model that ChatGPT uses is based off of data scraped from the Internet, there’s a surprising amount of affiliate favoritism here. This is more of a qualitative then quantitative observation, but it’s definitely true, and can be a big issue for WordPress beginners, who don’t have a lot of experience in the industry.
Essentially, the larger a company is, and the bigger their affiliate program, the more likely they are to show up in a response by any LLM, including ChatGPT. That’s because big affiliate programs result in more content published praising that company and pushing the product. All of this information biases the models reasoning as to the best product. There’s simply no way around this, and it’s difficult to train against this.
Let’s ask some questions to illustrate this point:
“What’s the best page builder (bulleted list of top 5 names, no description)”
This is basically a list of the top five page builders with the largest affiliate presence and program. Are they the best on the market in 2023? In my opinion, absolutely not.
“What’s the best hosting company for WordPress (bulleted list of top 5 names, no description)”
Another list of the top hosting companies, essentially based on the affiliate program size and digital presence.
And, of course, this makes perfect sense. If there are hundreds of articles raving about a specific service because they’re making 50% when somebody purchases, and the LLM is unable to distinguish this from genuine praise, it will always rank these services the highest.
Copywriting, Content and SEO
A lot of people, WordPress and Non-WordPress creators alike, are thrilled at the possibility of replacing copywriters, SEO specialists, and content writers with artificial intelligence.
It should be noted that this isn’t a new concept, and a AI content generators have been around for a very very long time at this point. However ChatGPT is one of the first that allows you to create content using a chat, bot like interface, and it’s also freely available for anybody to use.
Because of this, it brings up the topic of using artificial intelligence to replace copywriters, and create content again, primarily in the lens of WordPress creators and agencies.
For me, the general rule of thumb for our agency is that we will use artificially generated content to replace placeholder text. But when it comes to actual copywriting and content creation, this is done using professional writers and copywriters. AI will never understand every aspect of a business like a professional human being can.
When it comes to producing content that will rank online, Google has made it clear that it discourages this. AI content is not undetectable. The last thing you want is for your Google search console (even worse, your client’s) to look like this:
Instead of replacing content and copywriting, with an artificial counterpart, ChatGPT it’s great to augment workflows here as well. As mentioned above, we love delivering mock ups and ideation to clients with “real world” content attached to it (lispum replacement).
Using ChatGPT to ideate on different article structures, and titles is also a great use of it’s inherent skills of understanding, and being trained on Internet content. But I would strongly caution against using this to replace traditional authors and copywriters. Not only is it discouraged by the largest search engine, humans are simply better.
ChatGPT and the Impact On WordPress Developer Jobs
I’ve been getting a lot of questions about how ChatGPT, an artificial intelligence there’s a whole can impact WordPress developer jobs in both the near term and far future. Many of my thoughts are summed up in this article that I published on dev.to:
Here are the overall footnotes.
in the short term, and users have no idea what to do with code, and artificial intelligence isn’t able to install and configure it. Therefore, it’s really a developer specific tool, making workflows more efficient if you know how to use it.
in the midterm, AI will be able to create plug-ins, and developers will be in charge of overseeing the process of creation, testing, and installing
the long term, WordPress developers will likely transition more to a consulting role, thinking through the process of how to use AI to get to an end goal
ChatGPT is an amazing example of how artificial intelligence works, and what can happen when it’s brought to an easy to access user interface.
For WordPress professionals, this free tool can be a major benefit, clearing up administrative workflows like responding to emails, creating one off paragraphs of content, populating, mock ups with real world information, answering questions and explaining WordPress functions, etc…
However, when it comes to more advanced use cases like generating code and coverage of lesser known known products and plugins or for those that are constantly updated (after the training data cut off), things can get pretty dicey.
You’ll notice that a common theme of this article was “sure, you can use it for WordPress, but there are some major limitations that aren’t immediately clear”. The general solution to this issue is fine-tuning and other methods, but this isn’t available to the end users of ChatGPT.
It’s important to know the limitations of this platform, which aren’t immediately clear after signing up. On the other side of things, if you understand how to use ChatGPT, and the best use cases for it, it can seriously help any WordPress creator with their day to day operations.