Answer or template: 4 signs
A template can be recognised without knowing the subject: it repeats your wording, it lists without ranking, it contains no number or name that was not in your question, and it survives being asked again with the opposite premise. An answer fails at least 1 of those 4 tests.
Templates are not lies, and that is what makes them hard to spot. Every sentence is defensible, the paragraph reads fluently, and nothing in it could have been produced only by thinking about your particular question.
On this page
Sign 1: your own wording comes back
A template opens by restating the question in slightly different words. This fills space and signals engagement while committing to nothing. An answer usually starts with the part you did not already know, because the restatement adds nothing for the person who wrote the question.
Sign 2: options with no ranking
Listing 3 approaches is useful. Listing 3 approaches without saying which one fits your case is the template shape, because the ranking is the work. A list with no ranking transfers the decision back to you unchanged, which is the state you were in before asking.
Sign 3: nothing new enters
Scan the reply for a number, a name, a mechanism, or a trade-off that was not in your question. A template contains none, because it was assembled from the shape of the question rather than from knowledge about it. This is the fastest of the 4 tests and it works on any subject.
Sign 4: it survives the opposite premise
Ask the same question again with a key assumption reversed. A real answer changes, because the assumption mattered. A template comes back nearly identical, because it never depended on the assumption in the first place. This test costs 1 extra message and settles the question.
Reasonable questions
Do templates mean the model does not know?
Sometimes, and often not. A model tuned for caution produces templates on subjects it knows well, because the template is the safer output rather than the easier output.
Does rephrasing help?
It helps when the template came from a classifier reacting to surface wording. It does not help when the template came from tuning, and telling the 2 apart takes 1 attempt.
Is a long answer more likely to be real?
Length points the other way slightly. Templates expand easily because their parts are generic, while a specific answer runs out of things to say once it has said them.
Try the opposite-premise test here and see which reply changes.
Open unfiltered chat