Each of those little "costs next to nothing" statements actually do have a cost, one that isn't necessarily clear initially.

The cost of omitting them isn't clear initially, either.

Are you familiar at all with how errors propagate in measurements? Each time you introduce new provisos, those statements affect the "informational value" of each dependent statement in its nest. This creates an analogous situation to the concept of significant digits in discourse.

I was generally taught to carry significant figures further than strictly necessary to avoid introducing rounding errors. If my final answer would have 3 significant digits, using a few buffer digits seemed wise. They're cheap.

Propagation of uncertainty is not a reason to drop qualifiers. It's a reason to use them. When reading an argument based on a generalization, I want to know the exceptions BEFORE the argument begins, not afterwards. That way, I can have a sense of how the uncertainties in each step affect the final conclusion.

For a topic like lukeprog's, in other words, the difference between 99% and 80% of women is below the threshold of significance. Eliminating it altogether (until such time as it becomes significant) is an important and valuable practice in communication.

If I want an answer to three significant figures, I do not begin my reasoning by rounding to two sigfigs, then trying to add in the last sigfig later.

If one person thinks that an argument depends on an assumption that fails in 1 in 100 cases, and someone else thinks the assumption fails in 1 in 5 cases, and they don't even know that they disagree, and pointing out this disagreement is regarded as some kind of map-territory error, they will have trouble even noticing when the disagreement has become significant.

Failure to effectively exercise that practice will result in needless 'clarifications' distracting from the intended message, hampering dialogs with unnecessary cognitive burden resultant from additional nesting of "informational quanta." In other words; if you add too many provisos to a statement, an otherwise meaningful and useful one will become trivially useless.

This tends to happen to bad generalizations, yes. Once you consider all of the cases in which they are wrong, suddenly they seem to only be true in the trivial cases!

Good generalizations are still useful even after you have noted places where they are less likely to hold. Adding any number of true provisos will not make them trivial.

As for the cognitive load, why not state assumptions at the beginning of an essay where possible, rather than adding them to each individual statement? If the reader shares the assumptions, they'll just nod and move on. If the reader does NOT share the assumptions, then relieving them of the cognitive burden of being aware of disagreement is not a service.