Common myths, and where each one is answered

The beliefs that come up most, and the page that deals with each. Nothing is argued here: every line is a link to an explanation that carries its own sources. The reason for that restraint is the first section.

Published 30 August 2026 · Last verified 30 August 2026 · A router. Every claim it points at is dated on the page that makes it

Why this page argues nothing#

Every line below is a belief people arrive with. None of them is argued against here — each one points at the page that explains the mechanism underneath it, and the explanation lives there.

That is a deliberate constraint rather than a stylistic one. This site's page on whether what you're reading is out of date makes the case that a confident claim about what these systems “can't do” has the shortest shelf life of any sentence in the field. A page built out of such sentences would therefore be the fastest rotting page here, and it would rot invisibly: a myth-busting list reads as authoritative long after the thing it busted stopped being true.

The rule this page runs on

A line only goes on this list if a page here already explains the mechanism, and the line links to it. Nothing is asserted on this page that is not argued somewhere else with its sources attached. If a belief is worth correcting and no page explains why, the work is to write that page, not to add a sentence here. That is what stops a router turning into a place where claims quietly accumulate.

About what it knows#

“It lied to me.”
Answered under It isn't lying, and it isn't malfunctioning.
“It knows when it doesn't know.”
Answered under It doesn't know when it doesn't know.
“Ask it what its cutoff is.”
Answered under The model doesn't know its own cutoff.
“It can tell you how it works.”
Answered under It is unreliable about itself.
“Its knowledge stops on one date.”
Answered under There are actually two dates, and the gap matters.

About what it does#

“An agent is an AI that goes off and does things.”
Answered under Two very different things share the name.
“It went and looked that up.”
Answered under The trick: text that gets treated as a request.
“It has read my documents.”
Answered under What actually happens.
“Fine-tuning will teach it our facts.”
Answered under Why it fails at teaching facts.
“You can read its reasoning to check the answer.”
Answered under Why reading the reasoning doesn't verify the answer.
“It only does what it is asked.”
Answered under The finding that deserves the alarm.

About getting more out of it#

“Being polite to it gets better answers.”
Answered under What doesn't work, or isn't proven.
“A bigger context window is better.”
Answered under Why bigger is not automatically better.
“Just paste the whole thing in.”
Answered under Can't I just paste everything in now?.
“A word is about a token.”
Answered under The rule of thumb everyone repeats is attached to a tokeniser.
“It's bad at writing, maths, and anything creative.”
Answered under What it is not bad at, despite the reputation. The reverse myth, and the one this site is most wary of: it is the shape of claim that goes stale fastest.

Next → Is what you're reading out of date? · All → Guides