By the name, noninformative prior, the prior distribution doesn't contain any information about the parameter. Then why we use this thing to estimate the parameter by the Bayesian approach?
The purpose of using a non informative prior is to utilize a more objective approach. Non-informative priors allow the data to have a greater influence on the posterior. Though, if you have enough data, your choice of prior is not really going to matter.
That said, the term non-informative can be misleading. A uniform distribution does input information into a model since it provides equal probability weight to extreme values (imagine SAT scores, a non-informative prior would say that a perfect score is just as likely as an average score on the scale).
In my own research, I have bought into McElreath's line of reasoning and use semi-informative priors. These are priors that work more to regularize results rather than input any information into a model. I find them more useful than a non-informative prior when modeling a parameter that I do not have strong prior information about.