Hedged conclusions
The model writes eight hundred words weighing a question and then closes with "ultimately, it depends on your specific needs and priorities." The conclusion could be appended to any essay ever written. It is the single most reliable tell of unedited AI prose. A raw specimen: the blog post run.
Why it fails
A conclusion exists to commit. Readers finished the argument; they came to the last paragraph for the author's judgment, and the hedge withholds exactly that. Both-sides framing at the end also retroactively weakens the body: if nothing was decided, why was anything argued? Training rewards the hedge because it is never wrong, and never being wrong is the enemy of being useful.
Fix
Delete the last paragraph and read the piece. If it ends fine, the conclusion was filler; leave it deleted. If something is genuinely missing, write one sentence that takes a position a reasonable colleague could disagree with. A concrete recommendation with a named tradeoff ("choose X unless you need Y") beats a survey of considerations every time. Pair examples of the hedge with the committed rewrite, as in contrast pairs.