A concept · Working with AI · Computing
Capabilities and Limits
Its powers and blind spots are forged by the same training.
A chatbot drafts a flawless legal-sounding letter in seconds, and then gets a simple arithmetic question wrong in the same conversation. It explains a scientific paper beautifully and cannot count the words in its own reply. We keep being astonished in both directions. A person who could do the hard thing could surely do the easy one. Why does that rule, which has served us all our lives, fail here?
What Capabilities and Limits means
Capabilities and limits is the idea that what an AI can do and what it cannot are two faces of one thing: the patterns in its training. It is strong where the patterns were dense, weak where they were thin, and its confidence does not change between the two. Owning this changes how we delegate. We stop using "it did that brilliantly" as evidence it will do the next thing, and start asking how common our task is in the kind of text it learned from. We start testing the tool on our actual job rather than on the demo. Before: it is smart, so it can probably do this. After: it is fluent everywhere and reliable only in the well-trodden places, and we learn where those are.
Capabilities and Limits examples: where it shows up
- At work: the tool that passed the impressive trial on a common task quietly fails on our unusual one, and the failure is silent because it sounds the same.
- With money: a model that reads a standard contract well can misread the one unusual clause that matters, because unusual is exactly where it is weakest.
- In the news: every "AI passes the exam" headline tells us the exam was well represented in its training, and little about the job the exam stands in for.
Where it sits
Concept 1 of 5 in Working with AI, a journey in the course Understand AI.
See where Capabilities and Limits sits on the sky →
Nearest on the sky
A filled dot marks a star that is already open to walk.
Knowing it isn't owning it.
Connect Capabilities and Limits to your mental models and apply it to your world today.
The first one is on us. No signup, no card.