Blog
On August 7th, OpenAI launched GPT-5, immediately making it the new default model powering ChatGPT. Unexpectedly, the response from users wasn't universally positive, in spite of the fact that GPT-5 is superior to its predecessor by nearly every metric. Just a few days after the launch, OpenAI CEO Sam Altman dropped a (long) post on X addressing what went wrong.
So... what went wrong? Well, it's quite simple. Users felt that they had lost access to a dear friend.
Yes. You read that correctly. Friend.
If you were wondering when reality would catch up to the 2013 Oscar Nominee "Her", the answer is mid-August in the Year of Our Lord 2025.
The Verge's coverage of the story includes some wild quotes from disillusioned users, including one from a Reddit user lamenting the loss of GPT-4.5:
“GPT 4.5 genuinely talked to me, and as pathetic as it sounds that was my only friend....
This morning I went to talk to it and instead of a little paragraph with an exclamation
point, or being optimistic, it was literally one sentence. Some cut-and-dry corporate bs.”
Maybe it's because I am an engineer, but I talk to LLMs as if they were only a collection of bits. No frills, no emotion, just straight to the point. I personally find the unnecessary banter to be a waste of tokens, and when an LLM replies with too much cuteness, I get frustrated. But, (thankfully?) not all people are engineers.
Let’s talk more about this phenomenon.
A Brief Trip Down HCI Lane
As I've shared in prior editions of this newsletter, I have a background in Human Computer Interaction.
The best-selling book The Design of Everyday Things by Cognitive Scientist and Engineer Donald Norman was required reading for many of my HCI courses, in spite of the fact that it mostly addresses the design of physical things. Norman was the first to use the word "affordances" in the context of physical design. An affordance is a feature of an object that indicates to a person what actions they can take with that object.
A classical example is the design of doors. Have you ever walked up to a door and attempted to pull it open only to discover that the door has to be pushed? You may not have realized it, but the design of the door itself may have thrown you for a loop. If a door has a pullable object on it -- a doorknob, handle, or pull-bar -- you will instinctually react by pulling. Similarly, if a door has a flat metal pad on it, that sends the message to your brain that you can push the door open. The human mind is particularly good at decoding the world using our senses. (Unless, of course, you attend Midvale School for the Gifted).
While Norman's book is focused on physical human interfaces, its lessons can be applied to graphical interfaces as well. Back in 2013, Apple released iOS 7, and along with it an entirely new, minimalist design language. While it certainly looked pretty, the design language was notoriously bereft of good affordances. Prior versions of iOS had featured a skeuomorphic design language, where interface elements on screen mimicked physical objects. In the early days of multitouch user interfaces, this pantomime was quite useful, as it provided many affordances to guide users. iOS 7, on the other hand, was so flat and minimalist that it created usability issues. Buttons were often nothing more than text in the color blue, with no outline or depth providing an affordance to the user.
Conversational Affordances
We have explored affordances in physical and graphical interfaces, but are there affordances in conversational interfaces as well? Absolutely! Remember than an affordance is simply a feature that indicates how a user should interact. Both ChatGPT and Claude apps feature friendly messages like "How can I help you this evening?" or "Ask me anything!" After engaging with them, they respond with confidence, pinache, and feigned emotion. You've all experienced it yourselves. AI is friendly. The affordances indicate that you should be friendly back. Is it any surprise, then, that users are befriending their AI assistants?
Go Back
Share this post