AI Product Feature Amnesia
It's naïve, irresponsible, and dangerous to expect perfection from general-purpose AI chatbots. Claude, ChatGPT, and Gemini are certainly impressive, but only a fool believes they bat 1.000. Hallucinations are inevitable in trillion-parameter large language models trying to achieve artificial general intelligence. Always treat their claims and answers as dubious. The fleas come with the dog.
At the other end of the spectrum is weak or narrow AI. A tool that falls under this umbrella performs a discrete task better than humans can. Hawk-Eye is my favorite example of a system that does one thing exceptionally well: determine whether a tennis ball landed in or out. At a high level, we ask weak AI to do comparatively less while demanding far greater accuracy from it.
Somewhere in between these two extremes are AI companions that emphasize depth over breadth. These in-app chatbots don't purport to answer every conceivable question. Rather, companies design them to:
- Tell customers exactly what their core products can and can't do.
- Let them interact with software via natural language. (Case in point: Salesforce's Einstein.)
- Help them save time, make better business decisions, and the like.
At least that's the theory.
In today's post, I'll demonstrate yet another example of how AI companions are less accurate than they should be. One in particular exhibited an astonishing case of what I call AI product feature amnesia.
A Simple Product Prompt
I was curious about Notion's native ability to handle large surveys. I knew that it wouldn't be the best tool for the job but figured that I'd go straight to the source. Here's my prompt from earlier this month:
It'll only take a moment.
