An abstract is written to compress: no examples, no digressions, no first person, every noun nominalised. Those constraints are also the features that make text look generated, so a perfectly correct abstract can score higher than a messy body full of specifics.
The fix is not to make the abstract chatty. It is to put one concrete anchor in it: the population, the number, the intervention, the site. A sentence like "we interviewed 24 part-time students at two regional campuses" is compressed and specific at the same time, and specificity is what both readers and detectors reward.
Then check the body separately. If the abstract is clean and the discussion is flagged, the discussion probably has the uniform-paragraph problem, not the other way around.
This piece on AI detector accuracy explains why abstracts are a known weak spot for the models, and this academic-focused page covers the tone balance.