Fluency is not evidence
A chatbot produces likely text. Likely text reads confidently whether or not the underlying claim is supported, and nothing in the output marks the difference. This is the single most useful thing to know about any of them, and it does not change when the filter is removed.
People calibrate trust by tone, because with human writers tone tracks certainty. That heuristic fails here, and it fails in the direction that costs the most: the confident wrong answer.
The voice is the same in both cases
A well-supported fact and an invented detail are produced by the same process and rendered in the same register. There is no internal flag that leaks into the wording. Reading a reply for confidence measures the writing style of the training data, which is mostly confident prose.
Specific details are the risky part
Names, dates, figures and citations are exactly what a language model generates well and verifies not at all. A reply is most likely to be wrong precisely where it is most useful, which is why a specific-sounding answer deserves more checking than a vague answer, not less.
Removing the filter does not add accuracy
An uncensored chatbot answers where a filtered chatbot declines. The answer is produced the same way, so the accuracy is the same. Openness and correctness are separate properties, and a page promising the first says nothing about the second.
Cheap checks that work
Ask for the source and see whether one is offered or invented. Ask the same factual question in a fresh conversation and compare: an unsupported detail tends to move, and a supported detail does not. Both checks cost 1 message and catch most of it.
Before you paste it in
Do chatbots say when they are unsure?
Some are tuned to add uncertainty language, and it is added by style rather than measured. A hedge is not a confidence estimate.
Are invented citations common?
Common enough to check every time. A citation is a specific detail, and specific details are the weakest part of generated text.
Does asking are you sure help?
It usually produces agreement or reversal depending on tone, not a re-examination. Asking the same question fresh is a better test.
Ask for a source, then ask again in a fresh conversation and compare.
Open uncensored chatbot