Summary: For its first two years, the web was a text-only medium with a command-line user interface similar to the UI for current generative AI tools like ChatGPT. Only after GUI browsers launched in 1993 did the web explode. AI needs a similar GUI revolution in usability.
The World Wide Web was invented in 1991. My initial encounter with it at a hypertext convention left me unimpressed, particularly in comparison with the sumptuous HyperCard designs that were the fashion. Apple, ever in favor of proprietary systems (Quelle surprise?), let HyperCard drift so that it is now lost to the mists of history.
Line-mode web browsing worked like this: extensive passages of text [1] would scroll, interspersed with numerically [2] tagged hypertext links [3]. You'd enter the designated number to select a link at the prompt, initiating more scrolling text. The interaction style was like ChatGPT in 2023, except that the “modern” UI doesn’t even number the different items on the screen for easy reference.
For the first two years after 1991, the web's complexity was a playground for geeks alone, so it went nowhere.
1993: Enter Mosaic and the GUI Web
Come 1993; computer history changed: Marc Andreessen’s Mosaic was released as the first widely available graphical user interface to browse the web. (I say “widely available” because earlier GUI prototype software existed in various research labs but didn’t seem much use because it required fancy equipment .) In early 1993, Mosaic was only available on high-end Unix workstations, but in the fall of 1993, Mosaic was released for Windows and Macintosh, finally catering to regular folks. 30 years ago, the world really did change, which doesn’t happen as often as technology visionaries claim to want.
Web navigation started for real with Mosaic because users were finally liberated to click around the web. (Indeed, the successor to Mosaic, also developed by Marc Andreessen, was named the Netscape Navigator.) As shown by the chart later in this article, the GUI browser was the (inter)face that launched a million ships (websites). Ship by Leonardo.
Rather than abstract textual commands typed at a prompt, Mosaic presented web pages adorned with buttons, illustrations, and blue, underlined hypertext links. Most importantly, to follow a link, you simply clicked it, making the interaction easier and more suited for normal people. Free navigation became the dominant interaction style and has remained so to this day.
Early web pages looked like this in Mosaic:
A web page from 1994, in the early days of GUI access to the web, rendered by the NCSA Mosaic browser. The weird “S”-like logo in the upper right corner of the browser was indeed their logo. This was before any marketing professionals or skilled designers were involved with Internet software.
OK, pages were gray in 1993 and 1994, but having an actual page with a layout, as opposed to scrolling screenfuls of text, was a revelation. Not HyperCard (which offered nicer layouts, with animation when you clicked buttons), but close.
1995: The Web Exploded
The web experienced decent growth even during the first two years of line-mode access, growing from that first single website at CERN in 1991 to 26 websites in the entire world at the end of 1992.
But the advent of GUI browsing made the web take off. While two years of line-mode web yielded 26 websites, 11,550 new websites were created during the first two years of GUI web browsing. The only currently-famous website among the bunch was Yahoo!, which joined the web in February 1994.
The web grew by 44,423% during 1993 and 1994, for an annualized growth rate of 2,010%.
User interfaces do matter.
How the web reached a million sites in 6 years and then kept going, albeit at a slower rate: charting the growth of the world-wide web from 1991 to 2012, measured by the number of websites in the world. The launch dates for a few famous websites are marked. The number of websites is shown on a logarithmic scale to accommodate the exponential growth during the early period.
AI Must Use Hybrid UI
The user interfaces to current generative AI bear an uncanny resemblance to the nascent days of the web in 1992, especially in ChatGPT. Déjà vu all over again, with the endless scrolling and the command line as the only way of specifying what you want.
History shows that the web didn’t take off and started evolving useful services until normal users could use it, which required a graphical user interface. GUI doesn’t guarantee usability, but it sure helps.
AI is the first new user-interface paradigm in 60 years, moving from the command-based interaction design that characterized both line-mode and GUI applications to the new form of intent-based outcome specification. Users now tell the computer what outcome they want (say, rewrite this article in a more compelling style) rather than commanding the computer for each detailed step it must perform. Potential usability progress, except that users must tell the computer something. And doing so in the form of prose prompts is a pain in the behind and beyond the writing capabilities of most regular users, leading to an articulation barrier which is my fancy way of saying that conversational user interfaces have huge usability problems.
Solution: reuse the proven GUI techniques for the new outcome specifications. A hybrid user interface that controls AI interactions through traditional graphical UI techniques. GUI has superior UX qualities that AI will need for broad market success:
Visualizing available actions, guiding the user at each step according to the usability heuristic of recognition rather than recall.
Point-and-click interactions, rather than spelling and assembling words, which most people find challenging.
Direct manipulation: Allowing value adjustments through intuitive means, such as sliders.
Error reduction due to these direct manipulation and click-based interactions. I can’t tell you how many times I have made errors in using Midjourney due to typos or forgetting the correct number of dashes or colons in a prompt specification. (Remember that our UX ideology considers user errors the designer’s fault for creating an error-prone user interface.)
Prettier. In fact, much higher aesthetic appeal than the ugly UI sported by current AI tools. (I bet you would never have expected me to highlight good visual design since I was famous for distaining big graphics during the dial-up era of the web, but that was because of response time limits, not because I don’t realize that good-looking designs perform better with users.)
Consistency with the interaction styles of non-AI tools, such as web browsing and application use.