By Bob Glaser

The ubiquity of Generative AI doesn’t need to be pointed out as anyone with a smartphone is aware of it. As of June 2023, ChatGPT has had over 1.6 billion visits. When you add Bard, and Stable Diffusion, and Dall-E and others, the number of results generated is staggering. Since it is still relatively new to the wider audience, I am reminded of the ubiquitous spread of DTP (Desktop Publishing) and the early spread of the availability of the World Wide Web.
I should point out that I’m not discussing the topics of copyright, or copyright infringement and those subjects which are well discussed even if not yet well addressed elsewhere.

The nightmare of early DTP.

For those that remember that era, it wasn’t merely about the sudden availability of information. It was also, significantly about the sudden availability of tools and [digital] materials that were previously used by mostly trained and experienced professionals. For example, design tools. Designers were taught skills that allowed them to use these tools judiciously, with care, restraint, and a thorough and experienced understanding of a set of rules that weren’t arbitrary. The practice of good design wasn’t about the tools, but rather it was concepts, understanding, skills, and references to produce something that had a clear intent. There are a lot more constraints about what not to do than what to do. This is the challenge of every design. The litany of rules about what not to do all had very well established and understood reasons. When an experienced professional designer breaks a rule, they do it with a complete understanding of why the rule exists, as well as why they are breaking it. It’s not an arbitrary decision. A good designer has learned many concepts which are often counterintuitive to person ego such as the comprehension and importance of Ludwig Mies van der Rohe ‘s observation “Less is more.”

The problem happened when, because of the “home computer” and the WYSIWYG (what you see is what you get) environments, and the relatively inexpensive availability of software which put digital assets and capabilities into the hands of anyone who bought it started the era of “Anyone can be a desktop publisher.” The two (of many) most common errors were too many fonts, and presenting as much information as possible all at once.

The availability of hundreds of fonts gave the false impression that all that variety allowed a personal presentation of creativity. The more fonts used, the more creative the designer. Over time, people started to realize that the mess that using numerous fonts created wasn’t just problematic, it became a cliche (even though it still happens occasionally.)

Similarly, with color, and layout, a lack of understanding design principles and concepts lead to hundreds of thousands of pages like the those shown above.

The other aspect of design that wasn’t understood by the masses was the presentation of information. The common approach was to show as much as possible. This again is ignoring a concept similar to the “Less is more” idea in that too much is not only overwhelming, but can be confusing and is likely to stop the reader from trying to continue. They need to address hierarchy of information in terms of both importance, as well as relevance (and that just 2 aspects of many more.)

The era of generative AI.

So much discussion is going on about the sudden wide adoption and use of generative AI starting with ChatGPT LLM models and stable diffusion models as well as other variants.

Most of the problems that are discussed are significant and important. The issue is that, as in the era of early DTP, the people who are discussing it and red flagging these issues are the professionals who already know how to code, or write professional content, or create illustrations and designs. These generative AI programs are designed to produce a convincing result. Unfortunately, it is not designed to produce an accurate one, although that may be a side effect of the process, it is not a guaranteed one. That lack of guarantee is important because it cannot validate it’s results. The sources of data for the training models are not validated for accuracy. The accuracy is based on a statistical approach where confidence isn’t absolute, even when the data may be.

What this means is that a professional software engineer can look at the AI generated code and assess whether it produces the correct result(s) as well as whether it got the result by the correct processes. Someone without that training or experience, can easily recognize that it got the correct result but not be able to assess if it did it in a correct manner. Not being able to assess this, the inexperienced person may use the code that later causes more significant problems because it doesn’t get the result in a correct manner, or there may be other unrecognized anomalies.

The same thing happens in design. The AI method is in a way a subtractive approach. Start with everything and progressively remove everything that is irrelevant or counter to the intent of the design. As designer work, they start with what the design is supposed to convey. The add to it to refine and clarify those initial intents. This way, they are always including only those aspects and ideas that convey the original intents. They can also see when unexpected things happen in the developing design.

When generative AI is used, by design, the output often includes many things that weren’t initially requested in order to fill out the canvas. Now, at this point, using the prompts, the user requesting the design, must spend a lot of time figuring out how to remove irrelevant or problematic artifacts. This seems like ‘design’ but really it’s more about removing junk and less about adding value. This time spent removing junk starts to be perceived as adding value. It isn’t though. Think of it like this: if you have an account in the bank with a negative value, bringing the account to 0 (zero) isn’t adding value, it’s removing the negative value. You don’t owe money anymore but your account is still at a 0 value. If you look at people using Midjourney publicly on Discord, you can see it often takes so many iterations of an illustration to be generated, with a new prompt every time to achieve a result. The common problem here is that the person entering the prompts become myopically focused on what it isn’t, that they forget or minimize what it’s supposed to be. It seems that the generating what seems to be finished artwork appears to be timesaving, but it appears that becomes a distraction from the focus on the original intent.

I want to point out that people who are using it for fun and for personal reasons, this is not relevant, as long as it’s only for that purpose. I think a safe way to use ChatGPT, for example, is to view every result, that you don’t already know to be true, as potentially wrong. Remember: it is designed to sound convincing, not to be correct.

Also, I personally think using generative AI with proper restraint can be one of several great ways to jumpstart ideation in any aspect of creation, design, or development.

Lastly, I do see great potential in generative AI, but it’s current ubiquitous availability and hugely prolific use by a mostly uncritical public is likely to take a long time to correct and adjust. The long term effects are going to be remain with us for a significant amount of time.