LANDMARK · 6 MIN · ORIENT
Capabilities and limits
After this landmark, you can tell a model’s reliable zone from its unreliable one, and keep your important tasks inside the first.
Generative AI has a reliable zone and a shaky one, and they’re easy to confuse because it sounds equally confident in both. It’s genuinely good at working with material you put in front of it: drafting, rewriting, changing tone, brainstorming options, summarising a document that’s right there in the conversation. It’s unreliable when it has to reach into memory: specific facts, exact numbers, recent events, citations, anything that must be exactly right. The trap is that a capable tool feels trustworthy everywhere — so people carry it out of the zone where it’s reliable and get burned on the one fact that mattered.
This is the canonical concept. It stays the same across learner lenses so personalization never changes the underlying facts.
◌ EVERYDAY LENS
What this looks like for you
Ask it to reword your message or talk through an idea, and it shines. Ask it for a date, a statistic, or “which pill interacts with which”, and verify elsewhere — that’s exactly where a confident-sounding answer can be wrong.
MAKE A DECISION
You need three exact figures for a board slide and it’s due in ten minutes. How do you use the model well?
CARRY THISSplit your next AI task into “reshape what I gave it” and “recall a fact for me”, and only trust the first without checking.