Where the threshold is set
Every filter produces a score, and every service picks the point where a score becomes a block. That point differs by service, by category and by version. Nothing about the question changed between the 2 services; the number they compare it against did.
Comparing services makes the differences look like differences of principle. Most of them are differences of one number, chosen by weighing 2 kinds of mistake against each other.
On this page
A score, then a cut-off
The classifier does not decide; it scores. Somebody then chooses the point at which a score stops the reply. Moving that point by a small amount changes the behaviour on thousands of borderline questions, and it costs the service nothing to move.
Two mistakes pulling in opposite directions
A low cut-off lets through things the service will regret. A high cut-off stops ordinary questions and drives people away. Every service sits somewhere between the 2 mistakes, and where it sits reflects who it answers to rather than what it believes.
Categories are set separately
A service can be relaxed about strong language and strict about medical questions. This is why a single description such as strict or permissive fits no service accurately, and why your experience depends on which subjects you bring.
Thresholds move without notice
A change after an incident, a new model version, a different region: all move the line, and none of them is announced. A service that stopped answering something it used to answer has usually not changed its policy text, which is why the policy text is weak evidence about current behaviour.
Before you type
Do published policies tell me the threshold?
They describe categories rather than cut-offs. Two services with identical policy text can behave very differently, and only the behaviour is measurable.
Is a permissive service using a worse filter?
It is using a different cut-off, often with the same class of filter. Permissive and careless are not the same thing.
Why does the same service differ by country?
Legal exposure differs, so the cut-off is set per region. This is a business calculation, and it is applied before your question reaches the model.
Bring the question that was stopped elsewhere and compare the result.
Open no filter chat