Why does a human review sometimes become a rubber stamp?
Human review turns into a rubber stamp when the reviewer gets a huge queue, no supporting evidence and an immediate deadline. Time one review task, make rejection usable, and ask reviewers what they need to decide. Then adjust volume and timing to match the responsibility.
Calling a step human review doesn't show what the person can really do. A large queue, no evidence and an immediate deadline can produce approval with very little examination behind it.
Choose a review task and time it for real, counting the time to open the sources and check the important claims, plus a moment to ask a question. Then compare that with the volume the workflow sends this person.
Rejection needs somewhere to go. Decide where the work lands, who fixes it and whether the reviewer can pause a release. A disagreement shouldn't vanish into a comment nobody owns.
Ask reviewers what they'd need to make the decision. A comparison might do it, or a source excerpt, or a short account of what changed. Skip anything that makes the packet longer without making the choice clearer.
A role needs workable conditions. Accountability with little time, information or authority leaves the important part of the review undefined.