This is very similar to anchoring. It is also a form of conditioning.
A large language model tends to stay inside the same semantic frame as your initial prompt, so it often gives better phrasing of the same idea instead of a real shift.
If you want novel ideas, force perspective changes on purpose. Ask the model to solve from another industry, another stakeholder, or another constraint. A practical technique is to request 50 to 100 ideas and require each new idea to be orthogonal to the previous one, by changing domain, industry, stakeholder, constraint, or time horizon.
That is usually what unlocks creativity.