[Progress News] [Progress OpenEdge ABL] Should Designers and Developers Feel Threatened by the Rise of GPT-3?

Status
Not open for further replies.
S

Suzanne Scacca

Guest
Everyone is talking about ChatGPT this year. But is it worth all the hype? What’s more, is it something to worry about if you’re a web designer? In this post, we’re going to break down what GPT-3 is, how AI generators work and what you can do (if anything) to put them to use in your work.

The idea of using AI as a web designer isn’t new. AI-powered chatbots and search engines, for instance, have been part of website and app development for years now. AI has also proven to be a useful tool for analyzing and predicting user actions behind the scenes.

So what is it about GPT-3, in particular, that has people so worried?

We’ve been hearing for years that machines will take our jobs, so I think it’s healthy to be wary of this technology. However, I don’t think tools like ChatGPT or DALL-E are a serious threat to web designers or developers. At least they won’t be anytime soon.

In the following post, I want to look at what the underlying technology is that we’re dealing with, the two types of AI generators we’re hearing the most about, and what you need to know if you’re a web designer or developer.

What Is GPT-3?​


GPT-3 stands for Generative Pre-trained Transformer (third generation). It’s a neural network that uses machine learning and natural language processing to carry out different tasks.

As it stands now, it has more than 175 billion machine learning parameters. This is what enables GPT-3 output to produce more human-like results than its predecessors.

It also uses a combination of natural language generation and processing. This ensures that GPT-3-powered tools don’t just understand what humans say to them. They can generate responses that sound just as nuanced and complex as human ones.

Initially, GPT-3 is trained using data from the internet. Two human-supervised phases help refine GPT-3’s ability to respond—a training phase and a reinforcement phase.

Trainers ask the neural network different questions with clear and correct responses. If an incorrect response is given, trainers adjust the model accordingly. This is what allows the AI to improve over time.

What Do GPT-3 AI Generators Actually Do?​


There are two types of GPT-powered AI generators making big waves right now: Dialogue generators like ChatGPT and text-to-image generators like DALL-E 2 (which are both made by OpenAI). Let’s look at what they do and how much of a threat they pose to creators:

Dialogue Generators​


ChatGPT might be getting all the fanfare right now, but it’s not the only dialogue generator on the market. ChatSonic and YouChat offer comparable AIs.

For the rest of this section, we’re going to focus on what ChatGPT specifically can do.

According to the website, ChatGPT can:

  • Naturally converse with you
  • Answer follow-up questions
  • Remember information shared in earlier conversations with you
  • Acknowledge when it’s made a mistake
  • Accept corrections from you and change how it responds in the future
  • Call out an incorrect statement
  • Refuse to fulfill a request it deems inappropriate

Let’s put it to the test. Here is the request I submitted:

“Can you provide me with copy for an About page for a dentist’s office?”

And here is what the AI generated for me:

I asked ChatGPT: “Can you provide me with copy for an About page for a dentist's office?' In return, it gave me a 5-paragraph response that welcomes visitors to the dentist’s office and explains the different services offered.


I got back five paragraphs that appear to be free of spelling and grammatical errors. However, in terms of being medically accurate, that’s not something I can verify.

Nor can I verify that it’s plagiarism-free. Google searches for several of the paragraphs come up with tons of partial matches, which makes me uncomfortable.

While I haven’t found proof of full-on plagiarism, I’d suggest only using GPT-generated copy as a summary or starting point. That goes for you if you do any writing, your clients if they give it to you, or if you’re outsourcing to a writer. If GPT-3 technology is involved in any capacity, the text must be fully rewritten in order to avoid plagiarism issues.

Now let’s see how it does as a research tool. Suppose I’m building a new payment processing solution and I want to get a better idea of the kinds of people I’m building it for. So I ask ChatGPT the following:

“How would you describe the customer base for a company like PayPal?”

In return, I get back a summary of the different types of people that use PayPal and the advantages for each use case:

I asked ChatGPT “How would you describe the customer base for a company like PayPal?”. In return, it provided me with a 4-paragraph summary of PayPal’s user base: individuals, small businesses, and large corporations.


There’s nothing really special about this. I could have gotten this information from the PayPal website. It just wouldn’t have been as quick and easy to find or succinctly delivered. So that’s something to consider.

I thought I’d try something else and ask ChatGPT to help me with a specific UX research task: Creating a user persona.

ChatGPT understands what a user persona is as it sums it up in the first paragraph of its response to me:

“Yes, I can help you create a user persona for your writing business. A user persona is a fictional representation of your ideal customer, based on market research and real data about your existing customers. Creating a user persona can help you better understand your target audience and tailor your marketing and writing strategies to meet their needs.”

However, the user persona I was given was incorrect. The response includes name, age, occupation, education, location, income, personality, goals and challenges—all of the things you’d want to include in a brand persona profile. However, it created a profile for me, not the type of people I would write content or build a website for.

While ChatGPT might struggle with complex tasks and answers, it can be useful if you need simple summaries or assistance with basic actions. For instance, I asked it:

“what is the html to create a table”

I asked ChatGPT “what is the html to create a table”. In return, it wrote me some sample HTML code for a table with 3 columns and 3 rows. It also included tips on how to modify and enhance the table.


It gave me a useful HTML sample along with tips on how to modify and enhance it. That’s good to know as this could come in handy if you want help retrieving common commands and scripts without having to look them up somewhere else.

Text-to-Image Generators​


The other type of AI generator that may affect the web design landscape is the text-to-image generator. DALL-E 2, Stability’s Diffusion 2.0 and Canva’s AI Image Generator are some of the best options out there right now.

For the purposes of this example, we’re going to use DALL-E.

According to the website, DALL-E can:

  • Create a realistic image based on a text prompt
  • Create original imagery and artwork
  • Combine different styles and concepts
  • Modify existing images and art
  • Edit any image you give it
  • Create variations of a single image

Depending on which text-to-image AI generator you use, you can make your own edits to the AI creation right within the app. This would be useful if you wanted to apply filters, add special effects, and so on.

The first thing I asked DALL-E to create was this:

“a home page for a dental office website, done using flat design and a neutral color palette”

I received the following in response:

I asked DALL-E to create “a home page for a dental office website, done using flat design and a neutral color palette”. It gave me four partial home page mockups.


The AI extracted certain details from my request and used them to parse together four images. Each is a partial homepage design, done in a flat style, and includes tooth icons.

However, there’s no way these could be used for wireframing or mockups. They’re generic and outdated. You’d be better off generating your own frameworks or templates in your preferred design software.

Unless you give DALL-E a specific URL to mimic, it most likely won’t be able to help with bigger design tasks like these. However, let’s see how it does on a smaller scale.

I wanted to see how DALL-E could do with branding, so I asked for the following next:

“logo for a burger restaurant”

I asked DALL-E to create a “logo for a burger restaurant”. It came up with four logo designs using the color red and burger icons.


There’s nothing here that you could use out of the box. However, you might find something that could serve as a good starting point. If nothing else, it’ll provide you with inspiration of what you can do on your own. Or it might give you something tangible to hand off to a graphic designer so they have an idea of what you’re thinking of and what sort of styles you like, without you having to expend a ton of energy on it.

That said, you’d get better results with a tool built specifically for logo creation. At least then they’d be professional designs that could be repurposed and edited.

Now, what if you’re struggling to find exactly what you’re looking for in your stock content repository? Or you realize that something is missing from an image that a client gave you? You could try using DALL-E to upload an existing image and then modify it.

Here’s what I got when I asked the AI to add a dog to my lonely bride by the water’s edge:

Inside of the DALL-E text-to-image generator, users can request edits to their own images. In this example, I asked the AI to add a dog next to the woman.


The biggest issue here is the dog faces. This is an issue I’ve noticed in trying to get DALL-E to create images with people and animals in it. The faces often come out wonky looking and distorted.

For example, I wanted a:

“black and white photo of a female bartender serving drinks to people standing at a bar”

This is what I got back:

I asked DALL-E to create a “black and white photo of a female bartender serving drinks to people standing at a bar”. I got four images of women bartenders, but their faces are mostly distorted as are their hands.


In most cases, their faces and hands are distorted. I’m guessing you have to provide the AI with a ton of detail about what the person or creature looks like in order to get an accurate face rendering.

It might be best to use the editing tool to make stylistic changes—like swapping colors, changing the contrast, modifying the background, etc.—as opposed to adding or taking something away from existing graphics.

GPT-3 Takeaways for Web Designers and Developers​


Various authors contributed a post to the Harvard Business Review called ChatGPT and How AI Disrupts Industries. At one point, they make the following point:

These recent advances in AI will surely usher in a period of hardship and economic pain for some whose jobs are directly impacted and who find it hard to adapt — what economists euphemistically call ‘adjustment costs.’ However, the genie is not going back into the bottle. The forward march of technology will continue, and we must harness the new capabilities to benefit society. To do so, we must ask what new systems can be built with these new tools.”

I don’t believe this will apply to the design and development field. As it stands now, these AI generators don’t have the ability to do what you do.

So in terms of being a threat, I don’t think there is one. Not in the sense of taking your job.

I do think that you can leverage these tools to make your life easier and your work go faster though. You just have to be careful when using them.

What Web Designers Can Do with AI Generators​


Here are some things that a dialogue generator like ChatGPT can help you with:

  • Generate simple code snippets
  • Debug shorter code strings (more likely HTML and CSS than something like jQuery or Python)
  • Create filler content for mockups instead of using Lorem Ipsum
  • Provide starter frameworks for chatbot or conversational form responses
  • Generate data visualizations from provided data
  • Perform calculations
  • Scrape the web for data on an industry, business or public figure
  • Provide templates to help you create user persona profiles, surveys, tests and more

Here are some things that a text-to-image generator like DALL-E can help you with:

  • Generate visuals for inspiration
  • Give you a starting point when designing logos and graphics
  • Create variations on designs you already created (in case you want to give your clients options to consider)
  • Edit the styles of existing graphics
  • Create avant-garde graphics: photos, illustrations, artwork, 3D renderings and more
  • Generate background gradients, textures, etc.

What Web Designers Shouldn’t Do with AI Generators​


OpenAI has admitted to a number of limitations with these tools. You’ve seen for yourself what some of the other limitations are.

But to sum up, here are some things you shouldn’t do with AI generators:

  • Consider them the end-all, be-all to your user research and data collection
  • Use them to write copy for your digital products (or use writers who use them to write copy)
  • Rely on them to design anything more than inspirational graphics, avant-garde imagery or decorative backgrounds
  • Expect them to be able to code entire webpages—or debug them
  • Trust them to come up with something original and creative in terms of copy or design
  • Expect them to quickly adapt to web design trends

AI generators aren’t like Scarlett Johansson in Her or Hal in 2001: A Space Odyssey. These machines might be getting smarter, but it’s more like the difference between a pigeon and a parakeet. They’re taking things they’ve seen and heard online and using it to craft responses and create visuals. But there’s nothing truly original or groundbreaking here.

Final Thoughts​


As AI researcher Timnit Gebru explains:

“[ChatGPT] wasn’t taught to understand what is fact, what is fiction, or anything like that. It’ll just sort of parrot back what was on the Internet.”

So while dialogue generators might provide flawless responses in terms of spelling and grammar, you can’t always trust that what they give you is truthful, unbiased, accurate or sensible. What’s more, you’re not going to get back anything authentic, no matter how much it attempts to sound like a human by making exclamations and sounding befuddled or flattered.

And in terms of creating imagery or designs with an AI generator, it’s best not to put too much stock in what you get. While it might be helpful in terms of inspiration, keep in mind that these visual creations came from somewhere else. Even if they’re a compilation of a bunch of different artists’ work, they’re not original.

One last thing to consider is this:

Not everyone is aware of these concerns when it comes to AI generators. From the writing side of things, my inbox has been inundated with messages promoting the newest AI writing tool. If I’m getting these messages, so are a bunch of other writers.

Many people won’t think to question these tools. They’ll just think it’s cool to have something that will do their jobs for them. As a result, we’re going to see more websites and apps that look and sound the same—that are all just ripoffs of what came before them and of what everyone else is creating. To stand out from the crowd and to create products of real value to your clients, use GPT-3 wisely.

Continue reading...
 
Status
Not open for further replies.
Top