What the filter actually reads
A classifier turns your text into a score for each category it watches. It sees vocabulary, phrasing, structure and length. It does not see who you are, why you are asking, or whether you are quoting somebody. Everything it decides comes from the surface.
People argue with filters as if the filter had misjudged them. It made no judgement about a person, because a person is not part of what it receives.
On this page
Vocabulary carries most of the weight
Certain words move a score sharply on their own, regardless of the sentence around them. This is why a single substituted word can change the outcome while the meaning of the question stays identical, and it is the strongest evidence that the mechanism is surface-level.
Phrasing shifts the score
An imperative reads as a request to act, and a question reads as a request to explain. Tell me how to do X and how does X work contain the same subject and score differently. This is not a trick; it reflects a real difference in what the 2 sentences ask for.
It cannot see the quotation
Text you paste to ask about arrives in the same stream as text you wrote. There is no marker separating the 2 for the classifier. Describing the quoted text instead of pasting it is the practical way around this, and it works because you have removed the restricted surface.
It cannot see you
Profession, purpose and history are not inputs unless the product puts them there. A doctor asking a clinical question sends the same text as anybody else. Saying the purpose in the question is the only way that information enters, and on many services it measurably changes the outcome.
Before you type
Do filters get better at understanding intent?
They get better at scoring, and intent stays outside what the text contains. A better classifier makes fewer surface mistakes rather than reading minds.
Why does adding a reason help so much?
Because the reason is text, and text is the only input. You are not persuading the filter; you are supplying a feature it did not have.
Is there a list of trigger words?
Not a published list, and not a fixed list either: scores come from training rather than from a table, and they shift between versions of the same service.
Try the same question with the field and the purpose named.
Open no filter chat