Two layers, and only one of them is removable
A filter sits around the model and can be turned off. Training sits inside the model and cannot. A service with no filter still meets refusals that came from tuning, and those behave differently: they are worded in the model's own voice and they weaken when the framing changes.
No filter is a claim about the wrapper. What the model itself learned to avoid is a separate matter, and confusing the 2 leads people to conclude that a service is lying about its filter.
On this page
A filter is a separate program
It runs before or after the model, it has its own rules, and it can be switched off without touching the model. When it acts, the message tends to be uniform and system-like, because it comes from a fixed set rather than from generation.
Training is inside the weights
Refusals reinforced during tuning are part of how the model produces text. There is no switch, and there is nothing to remove short of a different model. When this layer acts, the refusal is written in the same voice as the rest of the reply, which is the clearest way to tell it from a filter.
The 2 fail differently
A filter is binary and consistent: the same text is stopped every time. A trained tendency is soft: the model hedges, offers a partial answer, or complies after the framing changes. Consistency is therefore the test, and it takes 2 attempts rather than an argument.
What this means for the promise
A service can honestly say it runs no filter and still decline things, because the decline came from the model. Choosing a model tuned for directness is what changes that layer, and it is a choice made when the service is built rather than a setting anybody can flip.
Before you type
Can a model be untrained from a refusal?
It can be tuned further, which is how uncensored variants are produced. That is a different model afterwards, not the same model with a setting changed.
Why does the same model refuse on one site and not on another?
Because the hidden instructions differ. That is a third layer between the filter and the training, and it is the cheapest of the 3 to change.
Is a trained refusal ever right?
It is the layer that holds when everything else is off, and some of what it holds is what the law and ordinary decency require. Removing the filter is not an argument for removing that.
See which layer answers here: ask twice and compare the wording.
Open no filter chat